Note
Go to the end to download the full example code
Off-the-shelf models quantization
Warning
QuantizeML ONNX quantization is an evolving feature. Some models may not be compatible.
The Global Akida workflow and the
PyTorch to Akida workflow guides
describe all the steps required to create, train, quantize and convert a model for Akida,
respectively using TensorFlow/Keras and PyTorch frameworks.
Here we will illustrate off-the-shelf/pretrained CNN models quantization for Akida using
MobileNet V2 from
the Hugging Face Hub.
Note
Off-the-shelf CNN models refer to already trained floating point models.
Their training recipe and framework have no importance as long as they can be exported
to ONNX.
Note however that this pathway offers slightly less flexibility than our default,
TensorFlow-based pathway - specifically, fine tuning of the quantized model is
not possible.
In most cases, that won’t matter, there should be almost no performance drop when
quantizing to 8-bit anyway.
Note
This tutorial leverages the Optimum toolkit,
an external tool, based on PyTorch, that allows models direct
download and export to ONNX.
pip install optimum[exporters]
1. Workflow overview
Off-the-shelf CNN models Akida workflow
As shown in the figure above, the QuantizeML toolkit allows the Post Training Quantization of ONNX models.
2. Data preparation
Given that the reference model was trained on ImageNet dataset
(which is not publicly available), this tutorial used a subset of 10 copyright free images.
A helper function imagenet.preprocessing.get_preprocessed_samples loads
and preprocesses (decodes, crops and extracts a square 224x224x3 patch from an input image)
these images.
import numpy as np
from akida_models.imagenet import get_preprocessed_samples
# Model specification and hyperparameters
NUM_CHANNELS = 3
IMAGE_SIZE = 224
# Load the preprocessed images and their corresponding labels for the test set
x_test_raw, labels_test = get_preprocessed_samples(IMAGE_SIZE, NUM_CHANNELS)
num_images = x_test_raw.shape[0]
# Get labels for the test set by index
# Note: Hugging Face models reserve the first index to null predictions
# (labeled as 'background' id). That is why we increase in '1' the original label id.
labels_test = labels_test + 1
print(f'{num_images} images and their labels are loaded and preprocessed.')
10 images and their labels are loaded and preprocessed.
As illustrated in 1. Workflow overview, the model’s source is at the user’s discretion. Here, we know a priori that MobileNet V2 was trained with images normalized within [-1, 1] interval. Also, ONNX models are usually saved with a channels-first format, input images are expected to be passed with the channels dimension on axis = 1.
# Project images in the range [-1, 1]
x_test = (x_test_raw / 127.5 - 1).astype('float32')
# Transpose the channels to the first axis
x_test = np.transpose(x_test, (0, 3, 1, 2))
3. Download and export
3.1. Download ONNX MobileNet V2
There are many repositories with models saved in ONNX format. In this example the Optimum API is used for downloading and exporting models to ONNX.
from optimum.exporters.onnx import main_export
# Download and convert MobiletNet V2 to ONNX
main_export(model_name_or_path="google/mobilenet_v2_1.0_224",
task="image-classification",
output="./")
Framework not specified. Using pt to export to ONNX.
config.json: 0%| | 0.00/69.8k [00:00<?, ?B/s]
config.json: 100%|██████████| 69.8k/69.8k [00:00<00:00, 501kB/s]
config.json: 100%|██████████| 69.8k/69.8k [00:00<00:00, 497kB/s]
model.safetensors: 0%| | 0.00/14.2M [00:00<?, ?B/s]
model.safetensors: 74%|███████▍ | 10.5M/14.2M [00:01<00:00, 8.85MB/s]
model.safetensors: 100%|██████████| 14.2M/14.2M [00:01<00:00, 11.2MB/s]
model.safetensors: 100%|██████████| 14.2M/14.2M [00:01<00:00, 10.5MB/s]
preprocessor_config.json: 0%| | 0.00/406 [00:00<?, ?B/s]
preprocessor_config.json: 100%|██████████| 406/406 [00:00<00:00, 200kB/s]
/usr/local/lib/python3.8/dist-packages/transformers/models/mobilenet_v2/feature_extraction_mobilenet_v2.py:28: FutureWarning: The class MobileNetV2FeatureExtractor is deprecated and will be removed in version 5 of Transformers. Please use MobileNetV2ImageProcessor instead.
warnings.warn(
Using the export variant default. Available variants are:
- default: The default ONNX variant.
Using framework PyTorch: 2.0.1+cu118
/usr/local/lib/python3.8/dist-packages/transformers/models/mobilenet_v2/modeling_mobilenet_v2.py:258: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
in_height = int(features.shape[-2])
/usr/local/lib/python3.8/dist-packages/transformers/models/mobilenet_v2/modeling_mobilenet_v2.py:259: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
in_width = int(features.shape[-1])
/usr/local/lib/python3.8/dist-packages/torch/onnx/_internal/jit_utils.py:306: UserWarning: Constant folding - Only steps=1 can be constant folded for opset >= 10 onnx::Slice op. Constant folding not applied. (Triggered internally at ../torch/csrc/jit/passes/onnx/constant_fold.cpp:179.)
_C._jit_pass_onnx_node_shape_type_inference(node, params_dict, opset_version)
/usr/local/lib/python3.8/dist-packages/torch/onnx/utils.py:689: UserWarning: Constant folding - Only steps=1 can be constant folded for opset >= 10 onnx::Slice op. Constant folding not applied. (Triggered internally at ../torch/csrc/jit/passes/onnx/constant_fold.cpp:179.)
_C._jit_pass_onnx_graph_shape_type_inference(
/usr/local/lib/python3.8/dist-packages/torch/onnx/utils.py:1186: UserWarning: Constant folding - Only steps=1 can be constant folded for opset >= 10 onnx::Slice op. Constant folding not applied. (Triggered internally at ../torch/csrc/jit/passes/onnx/constant_fold.cpp:179.)
_C._jit_pass_onnx_graph_shape_type_inference(
============= Diagnostic Run torch.onnx.export version 2.0.1+cu118 =============
verbose: False, log level: Level.ERROR
======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================
Post-processing the exported models...
Weight deduplication check in the ONNX export requires accelerate. Please install accelerate to run it.
Validating ONNX model model.onnx...
-[✓] ONNX model output names match reference model (logits)
- Validating ONNX Model output "logits":
-[✓] (2, 1001) matches (2, 1001)
-[✓] all values close (atol: 0.0001)
The ONNX export succeeded and the exported model was saved at: .
import onnx
# Load the model in memory
model_onnx = onnx.load_model("./model.onnx")
print(onnx.helper.printable_graph(model_onnx.graph))
graph torch_jit (
%pixel_values[FLOAT, batch_sizex3x224x224]
) initializers (
%classifier.weight[FLOAT, 1001x1280]
%classifier.bias[FLOAT, 1001]
%onnx::Conv_1737[FLOAT, 32x3x3x3]
%onnx::Conv_1738[FLOAT, 32]
%onnx::Conv_1740[FLOAT, 32x1x3x3]
%onnx::Conv_1741[FLOAT, 32]
%onnx::Conv_1743[FLOAT, 16x32x1x1]
%onnx::Conv_1744[FLOAT, 16]
%onnx::Conv_1746[FLOAT, 96x16x1x1]
%onnx::Conv_1747[FLOAT, 96]
%onnx::Conv_1749[FLOAT, 96x1x3x3]
%onnx::Conv_1750[FLOAT, 96]
%onnx::Conv_1752[FLOAT, 24x96x1x1]
%onnx::Conv_1753[FLOAT, 24]
%onnx::Conv_1755[FLOAT, 144x24x1x1]
%onnx::Conv_1756[FLOAT, 144]
%onnx::Conv_1758[FLOAT, 144x1x3x3]
%onnx::Conv_1759[FLOAT, 144]
%onnx::Conv_1761[FLOAT, 24x144x1x1]
%onnx::Conv_1762[FLOAT, 24]
%onnx::Conv_1764[FLOAT, 144x24x1x1]
%onnx::Conv_1765[FLOAT, 144]
%onnx::Conv_1767[FLOAT, 144x1x3x3]
%onnx::Conv_1768[FLOAT, 144]
%onnx::Conv_1770[FLOAT, 32x144x1x1]
%onnx::Conv_1771[FLOAT, 32]
%onnx::Conv_1773[FLOAT, 192x32x1x1]
%onnx::Conv_1774[FLOAT, 192]
%onnx::Conv_1776[FLOAT, 192x1x3x3]
%onnx::Conv_1777[FLOAT, 192]
%onnx::Conv_1779[FLOAT, 32x192x1x1]
%onnx::Conv_1780[FLOAT, 32]
%onnx::Conv_1782[FLOAT, 192x32x1x1]
%onnx::Conv_1783[FLOAT, 192]
%onnx::Conv_1785[FLOAT, 192x1x3x3]
%onnx::Conv_1786[FLOAT, 192]
%onnx::Conv_1788[FLOAT, 32x192x1x1]
%onnx::Conv_1789[FLOAT, 32]
%onnx::Conv_1791[FLOAT, 192x32x1x1]
%onnx::Conv_1792[FLOAT, 192]
%onnx::Conv_1794[FLOAT, 192x1x3x3]
%onnx::Conv_1795[FLOAT, 192]
%onnx::Conv_1797[FLOAT, 64x192x1x1]
%onnx::Conv_1798[FLOAT, 64]
%onnx::Conv_1800[FLOAT, 384x64x1x1]
%onnx::Conv_1801[FLOAT, 384]
%onnx::Conv_1803[FLOAT, 384x1x3x3]
%onnx::Conv_1804[FLOAT, 384]
%onnx::Conv_1806[FLOAT, 64x384x1x1]
%onnx::Conv_1807[FLOAT, 64]
%onnx::Conv_1809[FLOAT, 384x64x1x1]
%onnx::Conv_1810[FLOAT, 384]
%onnx::Conv_1812[FLOAT, 384x1x3x3]
%onnx::Conv_1813[FLOAT, 384]
%onnx::Conv_1815[FLOAT, 64x384x1x1]
%onnx::Conv_1816[FLOAT, 64]
%onnx::Conv_1818[FLOAT, 384x64x1x1]
%onnx::Conv_1819[FLOAT, 384]
%onnx::Conv_1821[FLOAT, 384x1x3x3]
%onnx::Conv_1822[FLOAT, 384]
%onnx::Conv_1824[FLOAT, 64x384x1x1]
%onnx::Conv_1825[FLOAT, 64]
%onnx::Conv_1827[FLOAT, 384x64x1x1]
%onnx::Conv_1828[FLOAT, 384]
%onnx::Conv_1830[FLOAT, 384x1x3x3]
%onnx::Conv_1831[FLOAT, 384]
%onnx::Conv_1833[FLOAT, 96x384x1x1]
%onnx::Conv_1834[FLOAT, 96]
%onnx::Conv_1836[FLOAT, 576x96x1x1]
%onnx::Conv_1837[FLOAT, 576]
%onnx::Conv_1839[FLOAT, 576x1x3x3]
%onnx::Conv_1840[FLOAT, 576]
%onnx::Conv_1842[FLOAT, 96x576x1x1]
%onnx::Conv_1843[FLOAT, 96]
%onnx::Conv_1845[FLOAT, 576x96x1x1]
%onnx::Conv_1846[FLOAT, 576]
%onnx::Conv_1848[FLOAT, 576x1x3x3]
%onnx::Conv_1849[FLOAT, 576]
%onnx::Conv_1851[FLOAT, 96x576x1x1]
%onnx::Conv_1852[FLOAT, 96]
%onnx::Conv_1854[FLOAT, 576x96x1x1]
%onnx::Conv_1855[FLOAT, 576]
%onnx::Conv_1857[FLOAT, 576x1x3x3]
%onnx::Conv_1858[FLOAT, 576]
%onnx::Conv_1860[FLOAT, 160x576x1x1]
%onnx::Conv_1861[FLOAT, 160]
%onnx::Conv_1863[FLOAT, 960x160x1x1]
%onnx::Conv_1864[FLOAT, 960]
%onnx::Conv_1866[FLOAT, 960x1x3x3]
%onnx::Conv_1867[FLOAT, 960]
%onnx::Conv_1869[FLOAT, 160x960x1x1]
%onnx::Conv_1870[FLOAT, 160]
%onnx::Conv_1872[FLOAT, 960x160x1x1]
%onnx::Conv_1873[FLOAT, 960]
%onnx::Conv_1875[FLOAT, 960x1x3x3]
%onnx::Conv_1876[FLOAT, 960]
%onnx::Conv_1878[FLOAT, 160x960x1x1]
%onnx::Conv_1879[FLOAT, 160]
%onnx::Conv_1881[FLOAT, 960x160x1x1]
%onnx::Conv_1882[FLOAT, 960]
%onnx::Conv_1884[FLOAT, 960x1x3x3]
%onnx::Conv_1885[FLOAT, 960]
%onnx::Conv_1887[FLOAT, 320x960x1x1]
%onnx::Conv_1888[FLOAT, 320]
%onnx::Conv_1890[FLOAT, 1280x320x1x1]
%onnx::Conv_1891[FLOAT, 1280]
) {
%/mobilenet_v2/conv_stem/first_conv/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/conv_stem/first_conv/Constant_output_0)
%/mobilenet_v2/conv_stem/first_conv/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/conv_stem/first_conv/Constant_1_output_0, %/mobilenet_v2/conv_stem/first_conv/ConstantOfShape_output_0)
%/mobilenet_v2/conv_stem/first_conv/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/Reshape_output_0 = Reshape(%/mobilenet_v2/conv_stem/first_conv/Concat_output_0, %/mobilenet_v2/conv_stem/first_conv/Constant_2_output_0)
%/mobilenet_v2/conv_stem/first_conv/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/Slice_output_0 = Slice(%/mobilenet_v2/conv_stem/first_conv/Reshape_output_0, %/mobilenet_v2/conv_stem/first_conv/Constant_4_output_0, %/mobilenet_v2/conv_stem/first_conv/Constant_5_output_0, %/mobilenet_v2/conv_stem/first_conv/Constant_3_output_0, %/mobilenet_v2/conv_stem/first_conv/Constant_6_output_0)
%/mobilenet_v2/conv_stem/first_conv/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/conv_stem/first_conv/Slice_output_0)
%/mobilenet_v2/conv_stem/first_conv/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/first_conv/Reshape_1_output_0 = Reshape(%/mobilenet_v2/conv_stem/first_conv/Transpose_output_0, %/mobilenet_v2/conv_stem/first_conv/Constant_7_output_0)
%/mobilenet_v2/conv_stem/first_conv/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/conv_stem/first_conv/Reshape_1_output_0)
%/mobilenet_v2/conv_stem/first_conv/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_stem/first_conv/Pad_output_0 = Pad[mode = 'constant'](%pixel_values, %/mobilenet_v2/conv_stem/first_conv/Cast_output_0, %/mobilenet_v2/conv_stem/first_conv/Constant_8_output_0)
%/mobilenet_v2/conv_stem/first_conv/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]](%/mobilenet_v2/conv_stem/first_conv/Pad_output_0, %onnx::Conv_1737, %onnx::Conv_1738)
%/mobilenet_v2/conv_stem/first_conv/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_stem/first_conv/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_stem/first_conv/activation/Clip_output_0 = Clip(%/mobilenet_v2/conv_stem/first_conv/convolution/Conv_output_0, %/mobilenet_v2/conv_stem/first_conv/activation/Constant_output_0, %/mobilenet_v2/conv_stem/first_conv/activation/Constant_1_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/conv_stem/conv_3x3/Constant_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/conv_stem/conv_3x3/Constant_1_output_0, %/mobilenet_v2/conv_stem/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/conv_stem/conv_3x3/Concat_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/conv_stem/conv_3x3/Reshape_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Constant_4_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Constant_5_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Constant_3_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/conv_stem/conv_3x3/Slice_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/conv_stem/conv_3x3/Transpose_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/conv_stem/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_stem/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/conv_stem/first_conv/activation/Clip_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Cast_output_0, %/mobilenet_v2/conv_stem/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/conv_stem/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 32, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/conv_stem/conv_3x3/Pad_output_0, %onnx::Conv_1740, %onnx::Conv_1741)
%/mobilenet_v2/conv_stem/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_stem/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_stem/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/conv_stem/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/conv_stem/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/conv_stem/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/conv_stem/reduce_1x1/Constant_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/conv_stem/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/conv_stem/reduce_1x1/Concat_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/conv_stem/reduce_1x1/Reshape_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/conv_stem/reduce_1x1/Slice_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/conv_stem/reduce_1x1/Transpose_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/conv_stem/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_stem/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/conv_stem/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Cast_output_0, %/mobilenet_v2/conv_stem/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/conv_stem/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/conv_stem/reduce_1x1/Pad_output_0, %onnx::Conv_1743, %onnx::Conv_1744)
%/mobilenet_v2/layer.0/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.0/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.0/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.0/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.0/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.0/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.0/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.0/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.0/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.0/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.0/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.0/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.0/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.0/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.0/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.0/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.0/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.0/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.0/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.0/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.0/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.0/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.0/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/conv_stem/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.0/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.0/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.0/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.0/expand_1x1/Pad_output_0, %onnx::Conv_1746, %onnx::Conv_1747)
%/mobilenet_v2/layer.0/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.0/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.0/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.0/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.0/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.0/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.0/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.0/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.0/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.0/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.0/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.0/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.0/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.0/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.0/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.0/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.0/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.0/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.0/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.0/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.0/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.0/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.0/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.0/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.0/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.0/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 96, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]](%/mobilenet_v2/layer.0/conv_3x3/Pad_output_0, %onnx::Conv_1749, %onnx::Conv_1750)
%/mobilenet_v2/layer.0/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.0/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.0/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.0/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.0/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.0/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.0/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.0/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.0/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.0/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.0/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.0/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.0/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.0/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.0/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.0/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.0/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.0/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.0/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.0/reduce_1x1/Pad_output_0, %onnx::Conv_1752, %onnx::Conv_1753)
%/mobilenet_v2/layer.1/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.1/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.1/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.1/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.1/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.1/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.1/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.1/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.1/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.1/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.1/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.1/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.1/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.1/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.1/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.1/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.1/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.1/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.1/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.1/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.1/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.1/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.1/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.0/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.1/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.1/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.1/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.1/expand_1x1/Pad_output_0, %onnx::Conv_1755, %onnx::Conv_1756)
%/mobilenet_v2/layer.1/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.1/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.1/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.1/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.1/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.1/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.1/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.1/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.1/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.1/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.1/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.1/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.1/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.1/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.1/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.1/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.1/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.1/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.1/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.1/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.1/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.1/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.1/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.1/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.1/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.1/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 144, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.1/conv_3x3/Pad_output_0, %onnx::Conv_1758, %onnx::Conv_1759)
%/mobilenet_v2/layer.1/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.1/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.1/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.1/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.1/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.1/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.1/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.1/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.1/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.1/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.1/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.1/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.1/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.1/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.1/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.1/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.1/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.1/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.1/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.1/reduce_1x1/Pad_output_0, %onnx::Conv_1761, %onnx::Conv_1762)
%/mobilenet_v2/layer.1/Add_output_0 = Add(%/mobilenet_v2/layer.0/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.1/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.2/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.2/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.2/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.2/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.2/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.2/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.2/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.2/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.2/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.2/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.2/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.2/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.2/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.2/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.2/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.2/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.1/Add_output_0, %/mobilenet_v2/layer.2/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.2/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.2/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.2/expand_1x1/Pad_output_0, %onnx::Conv_1764, %onnx::Conv_1765)
%/mobilenet_v2/layer.2/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.2/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.2/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.2/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.2/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.2/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.2/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.2/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.2/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.2/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.2/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.2/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.2/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.2/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.2/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.2/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.2/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.2/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.2/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.2/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.2/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.2/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.2/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.2/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.2/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.2/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 144, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]](%/mobilenet_v2/layer.2/conv_3x3/Pad_output_0, %onnx::Conv_1767, %onnx::Conv_1768)
%/mobilenet_v2/layer.2/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.2/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.2/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.2/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.2/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.2/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.2/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.2/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.2/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.2/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.2/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.2/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.2/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.2/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.2/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.2/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.2/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.2/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.2/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.2/reduce_1x1/Pad_output_0, %onnx::Conv_1770, %onnx::Conv_1771)
%/mobilenet_v2/layer.3/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.3/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.3/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.3/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.3/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.3/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.3/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.3/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.3/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.3/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.3/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.3/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.3/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.3/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.3/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.3/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.3/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.3/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.3/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.3/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.3/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.3/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.3/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.2/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.3/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.3/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.3/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.3/expand_1x1/Pad_output_0, %onnx::Conv_1773, %onnx::Conv_1774)
%/mobilenet_v2/layer.3/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.3/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.3/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.3/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.3/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.3/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.3/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.3/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.3/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.3/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.3/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.3/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.3/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.3/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.3/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.3/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.3/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.3/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.3/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.3/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.3/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.3/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.3/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.3/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.3/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.3/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 192, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.3/conv_3x3/Pad_output_0, %onnx::Conv_1776, %onnx::Conv_1777)
%/mobilenet_v2/layer.3/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.3/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.3/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.3/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.3/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.3/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.3/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.3/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.3/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.3/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.3/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.3/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.3/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.3/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.3/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.3/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.3/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.3/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.3/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.3/reduce_1x1/Pad_output_0, %onnx::Conv_1779, %onnx::Conv_1780)
%/mobilenet_v2/layer.3/Add_output_0 = Add(%/mobilenet_v2/layer.2/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.3/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.4/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.4/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.4/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.4/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.4/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.4/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.4/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.4/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.4/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.4/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.4/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.4/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.4/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.4/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.4/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.4/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.3/Add_output_0, %/mobilenet_v2/layer.4/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.4/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.4/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.4/expand_1x1/Pad_output_0, %onnx::Conv_1782, %onnx::Conv_1783)
%/mobilenet_v2/layer.4/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.4/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.4/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.4/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.4/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.4/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.4/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.4/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.4/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.4/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.4/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.4/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.4/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.4/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.4/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.4/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.4/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.4/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.4/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.4/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.4/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.4/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.4/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.4/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.4/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.4/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 192, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.4/conv_3x3/Pad_output_0, %onnx::Conv_1785, %onnx::Conv_1786)
%/mobilenet_v2/layer.4/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.4/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.4/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.4/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.4/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.4/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.4/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.4/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.4/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.4/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.4/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.4/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.4/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.4/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.4/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.4/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.4/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.4/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.4/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.4/reduce_1x1/Pad_output_0, %onnx::Conv_1788, %onnx::Conv_1789)
%/mobilenet_v2/layer.4/Add_output_0 = Add(%/mobilenet_v2/layer.3/Add_output_0, %/mobilenet_v2/layer.4/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.5/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.5/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.5/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.5/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.5/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.5/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.5/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.5/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.5/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.5/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.5/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.5/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.5/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.5/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.5/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.5/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.4/Add_output_0, %/mobilenet_v2/layer.5/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.5/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.5/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.5/expand_1x1/Pad_output_0, %onnx::Conv_1791, %onnx::Conv_1792)
%/mobilenet_v2/layer.5/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.5/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.5/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.5/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.5/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.5/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.5/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.5/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.5/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.5/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.5/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.5/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.5/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.5/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.5/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.5/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.5/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.5/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.5/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.5/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.5/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.5/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.5/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.5/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.5/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.5/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 192, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]](%/mobilenet_v2/layer.5/conv_3x3/Pad_output_0, %onnx::Conv_1794, %onnx::Conv_1795)
%/mobilenet_v2/layer.5/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.5/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.5/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.5/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.5/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.5/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.5/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.5/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.5/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.5/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.5/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.5/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.5/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.5/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.5/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.5/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.5/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.5/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.5/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.5/reduce_1x1/Pad_output_0, %onnx::Conv_1797, %onnx::Conv_1798)
%/mobilenet_v2/layer.6/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.6/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.6/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.6/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.6/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.6/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.6/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.6/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.6/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.6/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.6/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.6/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.6/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.6/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.6/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.6/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.6/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.6/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.6/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.6/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.6/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.6/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.6/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.5/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.6/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.6/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.6/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.6/expand_1x1/Pad_output_0, %onnx::Conv_1800, %onnx::Conv_1801)
%/mobilenet_v2/layer.6/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.6/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.6/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.6/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.6/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.6/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.6/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.6/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.6/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.6/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.6/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.6/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.6/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.6/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.6/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.6/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.6/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.6/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.6/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.6/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.6/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.6/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.6/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.6/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.6/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.6/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 384, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.6/conv_3x3/Pad_output_0, %onnx::Conv_1803, %onnx::Conv_1804)
%/mobilenet_v2/layer.6/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.6/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.6/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.6/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.6/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.6/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.6/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.6/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.6/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.6/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.6/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.6/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.6/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.6/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.6/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.6/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.6/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.6/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.6/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.6/reduce_1x1/Pad_output_0, %onnx::Conv_1806, %onnx::Conv_1807)
%/mobilenet_v2/layer.6/Add_output_0 = Add(%/mobilenet_v2/layer.5/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.6/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.7/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.7/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.7/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.7/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.7/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.7/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.7/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.7/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.7/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.7/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.7/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.7/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.7/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.7/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.7/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.7/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.6/Add_output_0, %/mobilenet_v2/layer.7/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.7/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.7/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.7/expand_1x1/Pad_output_0, %onnx::Conv_1809, %onnx::Conv_1810)
%/mobilenet_v2/layer.7/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.7/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.7/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.7/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.7/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.7/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.7/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.7/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.7/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.7/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.7/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.7/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.7/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.7/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.7/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.7/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.7/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.7/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.7/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.7/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.7/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.7/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.7/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.7/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.7/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.7/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 384, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.7/conv_3x3/Pad_output_0, %onnx::Conv_1812, %onnx::Conv_1813)
%/mobilenet_v2/layer.7/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.7/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.7/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.7/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.7/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.7/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.7/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.7/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.7/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.7/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.7/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.7/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.7/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.7/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.7/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.7/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.7/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.7/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.7/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.7/reduce_1x1/Pad_output_0, %onnx::Conv_1815, %onnx::Conv_1816)
%/mobilenet_v2/layer.7/Add_output_0 = Add(%/mobilenet_v2/layer.6/Add_output_0, %/mobilenet_v2/layer.7/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.8/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.8/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.8/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.8/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.8/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.8/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.8/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.8/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.8/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.8/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.8/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.8/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.8/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.8/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.8/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.8/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.7/Add_output_0, %/mobilenet_v2/layer.8/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.8/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.8/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.8/expand_1x1/Pad_output_0, %onnx::Conv_1818, %onnx::Conv_1819)
%/mobilenet_v2/layer.8/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.8/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.8/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.8/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.8/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.8/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.8/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.8/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.8/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.8/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.8/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.8/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.8/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.8/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.8/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.8/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.8/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.8/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.8/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.8/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.8/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.8/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.8/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.8/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.8/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.8/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 384, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.8/conv_3x3/Pad_output_0, %onnx::Conv_1821, %onnx::Conv_1822)
%/mobilenet_v2/layer.8/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.8/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.8/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.8/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.8/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.8/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.8/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.8/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.8/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.8/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.8/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.8/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.8/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.8/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.8/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.8/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.8/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.8/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.8/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.8/reduce_1x1/Pad_output_0, %onnx::Conv_1824, %onnx::Conv_1825)
%/mobilenet_v2/layer.8/Add_output_0 = Add(%/mobilenet_v2/layer.7/Add_output_0, %/mobilenet_v2/layer.8/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.9/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.9/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.9/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.9/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.9/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.9/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.9/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.9/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.9/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.9/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.9/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.9/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.9/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.9/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.9/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.9/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.8/Add_output_0, %/mobilenet_v2/layer.9/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.9/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.9/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.9/expand_1x1/Pad_output_0, %onnx::Conv_1827, %onnx::Conv_1828)
%/mobilenet_v2/layer.9/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.9/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.9/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.9/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.9/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.9/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.9/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.9/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.9/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.9/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.9/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.9/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.9/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.9/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.9/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.9/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.9/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.9/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.9/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.9/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.9/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.9/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.9/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.9/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.9/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.9/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 384, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.9/conv_3x3/Pad_output_0, %onnx::Conv_1830, %onnx::Conv_1831)
%/mobilenet_v2/layer.9/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.9/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.9/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.9/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.9/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.9/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.9/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.9/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.9/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.9/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.9/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.9/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.9/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.9/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.9/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.9/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.9/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.9/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.9/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.9/reduce_1x1/Pad_output_0, %onnx::Conv_1833, %onnx::Conv_1834)
%/mobilenet_v2/layer.10/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.10/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.10/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.10/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.10/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.10/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.10/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.10/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.10/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.10/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.10/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.10/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.10/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.10/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.10/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.10/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.10/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.10/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.10/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.10/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.10/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.10/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.10/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.9/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.10/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.10/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.10/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.10/expand_1x1/Pad_output_0, %onnx::Conv_1836, %onnx::Conv_1837)
%/mobilenet_v2/layer.10/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.10/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.10/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.10/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.10/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.10/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.10/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.10/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.10/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.10/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.10/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.10/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.10/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.10/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.10/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.10/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.10/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.10/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.10/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.10/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.10/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.10/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.10/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.10/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.10/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.10/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 576, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.10/conv_3x3/Pad_output_0, %onnx::Conv_1839, %onnx::Conv_1840)
%/mobilenet_v2/layer.10/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.10/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.10/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.10/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.10/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.10/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.10/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.10/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.10/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.10/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.10/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.10/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.10/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.10/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.10/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.10/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.10/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.10/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.10/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.10/reduce_1x1/Pad_output_0, %onnx::Conv_1842, %onnx::Conv_1843)
%/mobilenet_v2/layer.10/Add_output_0 = Add(%/mobilenet_v2/layer.9/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.10/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.11/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.11/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.11/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.11/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.11/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.11/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.11/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.11/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.11/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.11/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.11/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.11/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.11/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.11/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.11/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.11/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.10/Add_output_0, %/mobilenet_v2/layer.11/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.11/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.11/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.11/expand_1x1/Pad_output_0, %onnx::Conv_1845, %onnx::Conv_1846)
%/mobilenet_v2/layer.11/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.11/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.11/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.11/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.11/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.11/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.11/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.11/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.11/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.11/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.11/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.11/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.11/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.11/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.11/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.11/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.11/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.11/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.11/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.11/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.11/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.11/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.11/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.11/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.11/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.11/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 576, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.11/conv_3x3/Pad_output_0, %onnx::Conv_1848, %onnx::Conv_1849)
%/mobilenet_v2/layer.11/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.11/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.11/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.11/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.11/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.11/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.11/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.11/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.11/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.11/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.11/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.11/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.11/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.11/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.11/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.11/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.11/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.11/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.11/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.11/reduce_1x1/Pad_output_0, %onnx::Conv_1851, %onnx::Conv_1852)
%/mobilenet_v2/layer.11/Add_output_0 = Add(%/mobilenet_v2/layer.10/Add_output_0, %/mobilenet_v2/layer.11/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.12/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.12/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.12/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.12/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.12/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.12/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.12/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.12/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.12/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.12/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.12/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.12/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.12/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.12/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.12/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.12/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.11/Add_output_0, %/mobilenet_v2/layer.12/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.12/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.12/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.12/expand_1x1/Pad_output_0, %onnx::Conv_1854, %onnx::Conv_1855)
%/mobilenet_v2/layer.12/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.12/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.12/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.12/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.12/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.12/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.12/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.12/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.12/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.12/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.12/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.12/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.12/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.12/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.12/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.12/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.12/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.12/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.12/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.12/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.12/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.12/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.12/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.12/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.12/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.12/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 576, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]](%/mobilenet_v2/layer.12/conv_3x3/Pad_output_0, %onnx::Conv_1857, %onnx::Conv_1858)
%/mobilenet_v2/layer.12/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.12/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.12/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.12/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.12/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.12/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.12/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.12/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.12/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.12/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.12/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.12/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.12/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.12/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.12/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.12/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.12/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.12/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.12/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.12/reduce_1x1/Pad_output_0, %onnx::Conv_1860, %onnx::Conv_1861)
%/mobilenet_v2/layer.13/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.13/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.13/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.13/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.13/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.13/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.13/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.13/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.13/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.13/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.13/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.13/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.13/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.13/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.13/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.13/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.13/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.13/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.13/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.13/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.13/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.13/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.13/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.12/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.13/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.13/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.13/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.13/expand_1x1/Pad_output_0, %onnx::Conv_1863, %onnx::Conv_1864)
%/mobilenet_v2/layer.13/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.13/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.13/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.13/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.13/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.13/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.13/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.13/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.13/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.13/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.13/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.13/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.13/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.13/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.13/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.13/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.13/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.13/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.13/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.13/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.13/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.13/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.13/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.13/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.13/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.13/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 960, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.13/conv_3x3/Pad_output_0, %onnx::Conv_1866, %onnx::Conv_1867)
%/mobilenet_v2/layer.13/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.13/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.13/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.13/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.13/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.13/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.13/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.13/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.13/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.13/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.13/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.13/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.13/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.13/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.13/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.13/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.13/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.13/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.13/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.13/reduce_1x1/Pad_output_0, %onnx::Conv_1869, %onnx::Conv_1870)
%/mobilenet_v2/layer.13/Add_output_0 = Add(%/mobilenet_v2/layer.12/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.13/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.14/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.14/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.14/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.14/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.14/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.14/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.14/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.14/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.14/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.14/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.14/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.14/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.14/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.14/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.14/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.14/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.13/Add_output_0, %/mobilenet_v2/layer.14/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.14/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.14/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.14/expand_1x1/Pad_output_0, %onnx::Conv_1872, %onnx::Conv_1873)
%/mobilenet_v2/layer.14/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.14/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.14/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.14/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.14/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.14/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.14/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.14/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.14/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.14/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.14/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.14/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.14/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.14/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.14/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.14/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.14/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.14/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.14/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.14/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.14/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.14/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.14/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.14/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.14/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.14/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 960, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.14/conv_3x3/Pad_output_0, %onnx::Conv_1875, %onnx::Conv_1876)
%/mobilenet_v2/layer.14/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.14/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.14/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.14/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.14/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.14/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.14/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.14/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.14/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.14/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.14/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.14/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.14/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.14/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.14/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.14/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.14/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.14/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.14/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.14/reduce_1x1/Pad_output_0, %onnx::Conv_1878, %onnx::Conv_1879)
%/mobilenet_v2/layer.14/Add_output_0 = Add(%/mobilenet_v2/layer.13/Add_output_0, %/mobilenet_v2/layer.14/reduce_1x1/convolution/Conv_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.15/expand_1x1/Constant_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.15/expand_1x1/Constant_1_output_0, %/mobilenet_v2/layer.15/expand_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.15/expand_1x1/Concat_output_0, %/mobilenet_v2/layer.15/expand_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.15/expand_1x1/Reshape_output_0, %/mobilenet_v2/layer.15/expand_1x1/Constant_4_output_0, %/mobilenet_v2/layer.15/expand_1x1/Constant_5_output_0, %/mobilenet_v2/layer.15/expand_1x1/Constant_3_output_0, %/mobilenet_v2/layer.15/expand_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.15/expand_1x1/Slice_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/expand_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.15/expand_1x1/Transpose_output_0, %/mobilenet_v2/layer.15/expand_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.15/expand_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.15/expand_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.15/expand_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.14/Add_output_0, %/mobilenet_v2/layer.15/expand_1x1/Cast_output_0, %/mobilenet_v2/layer.15/expand_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.15/expand_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.15/expand_1x1/Pad_output_0, %onnx::Conv_1881, %onnx::Conv_1882)
%/mobilenet_v2/layer.15/expand_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.15/expand_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.15/expand_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.15/expand_1x1/convolution/Conv_output_0, %/mobilenet_v2/layer.15/expand_1x1/activation/Constant_output_0, %/mobilenet_v2/layer.15/expand_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.15/conv_3x3/Constant_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.15/conv_3x3/Constant_1_output_0, %/mobilenet_v2/layer.15/conv_3x3/ConstantOfShape_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.15/conv_3x3/Concat_output_0, %/mobilenet_v2/layer.15/conv_3x3/Constant_2_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/Slice_output_0 = Slice(%/mobilenet_v2/layer.15/conv_3x3/Reshape_output_0, %/mobilenet_v2/layer.15/conv_3x3/Constant_4_output_0, %/mobilenet_v2/layer.15/conv_3x3/Constant_5_output_0, %/mobilenet_v2/layer.15/conv_3x3/Constant_3_output_0, %/mobilenet_v2/layer.15/conv_3x3/Constant_6_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.15/conv_3x3/Slice_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/conv_3x3/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.15/conv_3x3/Transpose_output_0, %/mobilenet_v2/layer.15/conv_3x3/Constant_7_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.15/conv_3x3/Reshape_1_output_0)
%/mobilenet_v2/layer.15/conv_3x3/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.15/conv_3x3/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.15/expand_1x1/activation/Clip_output_0, %/mobilenet_v2/layer.15/conv_3x3/Cast_output_0, %/mobilenet_v2/layer.15/conv_3x3/Constant_8_output_0)
%/mobilenet_v2/layer.15/conv_3x3/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 960, kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.15/conv_3x3/Pad_output_0, %onnx::Conv_1884, %onnx::Conv_1885)
%/mobilenet_v2/layer.15/conv_3x3/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.15/conv_3x3/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.15/conv_3x3/activation/Clip_output_0 = Clip(%/mobilenet_v2/layer.15/conv_3x3/convolution/Conv_output_0, %/mobilenet_v2/layer.15/conv_3x3/activation/Constant_output_0, %/mobilenet_v2/layer.15/conv_3x3/activation/Constant_1_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/layer.15/reduce_1x1/Constant_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/layer.15/reduce_1x1/Constant_1_output_0, %/mobilenet_v2/layer.15/reduce_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/layer.15/reduce_1x1/Concat_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Constant_2_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/Slice_output_0 = Slice(%/mobilenet_v2/layer.15/reduce_1x1/Reshape_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Constant_4_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Constant_5_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Constant_3_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Constant_6_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/layer.15/reduce_1x1/Slice_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/layer.15/reduce_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/layer.15/reduce_1x1/Transpose_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Constant_7_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/layer.15/reduce_1x1/Reshape_1_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/layer.15/reduce_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.15/conv_3x3/activation/Clip_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Cast_output_0, %/mobilenet_v2/layer.15/reduce_1x1/Constant_8_output_0)
%/mobilenet_v2/layer.15/reduce_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/layer.15/reduce_1x1/Pad_output_0, %onnx::Conv_1887, %onnx::Conv_1888)
%/mobilenet_v2/conv_1x1/Constant_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/Constant_1_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/ConstantOfShape_output_0 = ConstantOfShape[value = <Tensor>](%/mobilenet_v2/conv_1x1/Constant_output_0)
%/mobilenet_v2/conv_1x1/Concat_output_0 = Concat[axis = 0](%/mobilenet_v2/conv_1x1/Constant_1_output_0, %/mobilenet_v2/conv_1x1/ConstantOfShape_output_0)
%/mobilenet_v2/conv_1x1/Constant_2_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/Reshape_output_0 = Reshape(%/mobilenet_v2/conv_1x1/Concat_output_0, %/mobilenet_v2/conv_1x1/Constant_2_output_0)
%/mobilenet_v2/conv_1x1/Constant_3_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/Constant_4_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/Constant_5_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/Constant_6_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/Slice_output_0 = Slice(%/mobilenet_v2/conv_1x1/Reshape_output_0, %/mobilenet_v2/conv_1x1/Constant_4_output_0, %/mobilenet_v2/conv_1x1/Constant_5_output_0, %/mobilenet_v2/conv_1x1/Constant_3_output_0, %/mobilenet_v2/conv_1x1/Constant_6_output_0)
%/mobilenet_v2/conv_1x1/Transpose_output_0 = Transpose[perm = [1, 0]](%/mobilenet_v2/conv_1x1/Slice_output_0)
%/mobilenet_v2/conv_1x1/Constant_7_output_0 = Constant[value = <Tensor>]()
%/mobilenet_v2/conv_1x1/Reshape_1_output_0 = Reshape(%/mobilenet_v2/conv_1x1/Transpose_output_0, %/mobilenet_v2/conv_1x1/Constant_7_output_0)
%/mobilenet_v2/conv_1x1/Cast_output_0 = Cast[to = 7](%/mobilenet_v2/conv_1x1/Reshape_1_output_0)
%/mobilenet_v2/conv_1x1/Constant_8_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_1x1/Pad_output_0 = Pad[mode = 'constant'](%/mobilenet_v2/layer.15/reduce_1x1/convolution/Conv_output_0, %/mobilenet_v2/conv_1x1/Cast_output_0, %/mobilenet_v2/conv_1x1/Constant_8_output_0)
%/mobilenet_v2/conv_1x1/convolution/Conv_output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%/mobilenet_v2/conv_1x1/Pad_output_0, %onnx::Conv_1890, %onnx::Conv_1891)
%/mobilenet_v2/conv_1x1/activation/Constant_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_1x1/activation/Constant_1_output_0 = Constant[value = <Scalar Tensor []>]()
%/mobilenet_v2/conv_1x1/activation/Clip_output_0 = Clip(%/mobilenet_v2/conv_1x1/convolution/Conv_output_0, %/mobilenet_v2/conv_1x1/activation/Constant_output_0, %/mobilenet_v2/conv_1x1/activation/Constant_1_output_0)
%/mobilenet_v2/pooler/GlobalAveragePool_output_0 = GlobalAveragePool(%/mobilenet_v2/conv_1x1/activation/Clip_output_0)
%/mobilenet_v2/Flatten_output_0 = Flatten[axis = 1](%/mobilenet_v2/pooler/GlobalAveragePool_output_0)
%logits = Gemm[alpha = 1, beta = 1, transB = 1](%/mobilenet_v2/Flatten_output_0, %classifier.weight, %classifier.bias)
return %logits
}
3.2. Evaluate model performances
The ONNXRuntime package is a cross-platform accelerator capable of loading and running models described in ONNX format. We take advantage of this framework to run the ONNX models and evaluate their performances.
Note
We only compute accuracy on 10 images.
from onnxruntime import InferenceSession
def evaluate_onnx_model(model):
sess = InferenceSession(model.SerializeToString())
# Calculate outputs by running images through the session
outputs = sess.run(None, {model.graph.input[0].name: x_test})
# The class with the highest score is what we choose as prediction
predicted = np.squeeze(np.argmax(outputs[0], 1))
# Compute the number of valid predictions
return int((predicted == labels_test).sum())
correctly_classified_floating = evaluate_onnx_model(model_onnx)
print(f'Floating point model accuracy: {correctly_classified_floating}/{num_images}.')
Floating point model accuracy: 10/10.
4. Quantize
Akida processes integer activations and weights. Therefore, the floating point model
must be quantized in preparation to run on an Akida accelerator.
QuantizeML quantize()
function recognizes ModelProto objects
and quantizes them for Akida. The result is another
ModelProto, compatible with the
CNN2SNN Toolkit.The table below summarizes the obtained accuracy at the various stages using the full
ImageNet dataset.
Calibration parameters |
Float accuracy |
Quantized accuracy |
Akida accuracy |
|---|---|---|---|
Random samples / per-tensor |
71.790 |
70.550 |
70.588 |
Imagenet samples / per-tensor |
71.790 |
70.472 |
70.628 |
Note
Please refer to the QuantizeML toolkit user guide and the Advanced QuantizeML tutorial for details about quantization parameters.
from quantizeml.models import quantize, QuantizationParams
# Quantize with activations quantized per tensor
qparams = QuantizationParams(per_tensor_activations=True)
model_quantized = quantize(model_onnx, qparams=qparams, num_samples=5)
# Evaluate the quantized model performance
correctly_classified = evaluate_onnx_model(model_quantized)
print(f'Quantized model accuracy: {correctly_classified_floating}/{num_images}.')
Calibrating with 5/5.0 samples
Quantized model accuracy: 10/10.
Note
Once the model is quantized, the convert function must be used to retrieve a model in Akida format ready for inference. Please refer to the PyTorch to Akida workflow for further details.
Total running time of the script: (0 minutes 13.665 seconds)