Source code for quantizeml.onnx_support.layers.conv2d

#!/usr/bin/env python
# ******************************************************************************
# Copyright 2023 Brainchip Holdings Ltd.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ******************************************************************************
__all__ = ["QuantizedConv2D", "get_qconv"]

import numpy as np

from onnx import AttributeProto as AP, TensorProto as TP
from onnx.helper import make_node

from .base_layer import OnnxLayer
from .subgraph_ops import cast_tensors_to, get_pool_ops, get_scale_out_ops
from .subgraph_ops.activation import get_activation_ops
from .subgraph_ops.padding import get_padding_ops, transform_pads_into_array
from .compute_shapes import compute_onnx_conv_output
from .node_constraints import check_supported_attributes
from ..graph_tools import (TENSOR_SHAPE, get_field, get_node, get_variable,
                           value_info_to_tensor_shape)
from ..quantization.weights import quantize_weights, quantize_vector, fold_zero_point, align_to
from ..quantization.outputs import downscale


def get_qconv(nodes, graph):
    conv_node = nodes[0]
    assert conv_node.op_type == 'Conv'
    check_supported_attributes(nodes)
    strides = get_field(conv_node, 'strides', (1, 1))

    pool_type = "none"
    pool_size = (2, 2)
    pool_strides = (1, 1)
    pool_node = get_node(nodes, 'MaxPool')
    pool_pads = [0, 0, 0, 0]
    if pool_node:
        pool_type = "max"
        # kernel_shape attribute is mandatory for MaxPool
        pool_size = get_field(pool_node, 'kernel_shape')
        pool_strides = get_field(pool_node, 'strides', pool_strides)
        pool_pads = get_field(pool_node, "pads", pool_pads)
    pool_node = get_node(nodes, 'GlobalAveragePool')
    if pool_node:
        pool_type = "gap"

    act_node = get_node(nodes, 'Relu')
    qconv = QuantizedConv2D(strides=strides,
                            pool_type=pool_type,
                            pool_size=pool_size,
                            pool_strides=pool_strides,
                            pool_pads=pool_pads,
                            activation=bool(act_node),
                            name=conv_node.name)

    # Sets the weights to configure the operation chain
    qconv.set_weight("kernel", get_variable(conv_node.input[1], graph))
    # If third attribute is there and it is not empty, then there is a bias
    if len(conv_node.input) == 3 and conv_node.input[2]:
        qconv.set_weight("bias", get_variable(conv_node.input[2], graph))
    pads = get_field(conv_node, 'pads', False)
    if pads:
        qconv.set_weight("pads", transform_pads_into_array(pads))
    if act_node and len(act_node.input) > 2 and act_node.input[2]:
        qconv.set_weight("max_value", get_variable(act_node.input[2], graph))
    return qconv


[docs]class QuantizedConv2D(OnnxLayer): """Intermediate representation of QLinearConv() + MaxPool() + ReLU() as an exportable node. Args: strides (list of int, optional): the convolutional strides. Defaults to [1, 1]. pool_type (str, optional): the pool type, one of {"none", "max", "gap"}. Defaults to "none". pool_size (list of int, optional): the kernel pool shape. Ignore it when pool_type != "max". Defaults to (2, 2). pool_stride (list of int, optional): the kernel strides. Ignore it when pool_type != "max". Defaults to (2, 2). pool_pads (list of int, optional): the size of each padding dimension. Ignore it when pool_type != "max". Defaults to [0, 0, 0, 0]. input_conv (bool, optional): whether it is extended the set of operations of the basic QuantizedConv2D, allowing to modify the padding value per input channel. Defaults to False. activation (bool, optional): whether to apply relu operation. Defaults to False. name (str, optional): the node name. Defaults to ''. """ def __init__(self, strides=[1, 1], pool_type="none", pool_size=(2, 2), pool_strides=(2, 2), pool_pads=[0, 0, 0, 0], activation=False, name=''): assert pool_type in ["none", "max", "gap"] super().__init__("QuantizedConv2D", strides=strides, pool_size=pool_size, pool_strides=pool_strides, pool_pads=pool_pads, name=name) # Save properties need to serialize operation name self.serialize_attr["pool_type"] = pool_type self.serialize_attr["activation"] = activation self.serialize_attr["scale"] = True # Declare weights self._add_weight("kernel") self._add_weight("bias") self._add_weight("max_value") self._add_weight("pads", dtype="int64") def __build__(self, input_ts, downscale=True): assert input_ts.dtype in (np.uint8, np.int8) assert downscale, f"{self.name} ({self.base_name}) does not support 32bit output" assert self.weights["kernel"].ndim == 4 # The chain of operations is modified by the type of input: kernel_shape = self.weights["kernel"].shape if input_ts.dtype == np.uint8: self.base_name = "QuantizedInputConv2D" if self.weights["bias"].size == 0: # Bias is mandatory on this configuration filters = kernel_shape[0] self.set_weight("bias", np.zeros(filters, dtype="float32")) # Initialize weights if self.weights["pads"].size == 0: self.set_weight("pads", np.zeros(len(kernel_shape) * 2, dtype="int64")) # Compute output shape conv_output_shape = compute_onnx_conv_output(self, input_ts.shape) output_ts = TENSOR_SHAPE(conv_output_shape, np.dtype("int8")) return output_ts def __quantize__(self, qinput, out_tensor_range, force_fp=False): i_scale = qinput.weights["scale"] # Perform cross-layer equalization, i.e.: rescale weights with input scale. # To do that first reshape i_scale to put it into axis = 1 and be capable of broadcasting. assert i_scale.ndim <= 1 kernel = self.weights["kernel"] kernel = kernel / align_to(i_scale, kernel.ndim, axis=1) # Quantize and set weights qweights, i_scale = quantize_weights(kernel) # Prepare tensors list with unique names conv_name = self.name prefix = conv_name + "_" weights_dict = {} bias = self.weights["bias"] if "InputConv" in self.op_type: # If calibration was done per tensor, repeat zero point over each channel zero_point = qinput.weights["zero_point"] if zero_point.size == 1: zero_point = np.repeat(zero_point, kernel.shape[1]) weights_dict[prefix + "Xpad"] = zero_point # Fold zero point in bias # Note: Dequantize kernel instead to use the float to reduce quantization error kernel = qweights / align_to(i_scale, qweights.ndim, axis=0) bias = fold_zero_point(bias, kernel, zero_point) weights_dict[prefix + "Wi"] = qweights if "Biased" in self.op_type: qbias = quantize_vector(bias, i_scale) weights_dict[prefix + "B"] = qbias weights_dict[prefix + "pads"] = self.weights["pads"] # Reshape i_scale to match with channel axis i_scale = align_to(i_scale, qweights.ndim) # Quantize max value when there is an activation if "Clipped" in self.op_type: qmax_value = quantize_vector(self.weights["max_value"], i_scale, signed=False) weights_dict[prefix + "max_value"] = qmax_value # Fold spatial dimension when GAP if "GlobalAvgPool" in self.op_type: input_shape = value_info_to_tensor_shape(self.input).shape input_shape = compute_onnx_conv_output(self, input_shape, apply_pool=False) i_scale *= input_shape[-2] * input_shape[-1] # Now consider calibrated output range scale, s_out, ocalib_scale = downscale(out_tensor_range, i_scale, force_fp=force_fp) weights_dict.update({prefix + "M": scale.astype("uint8"), prefix + "S_out": s_out}) # Return quantized weights and ouput scale return weights_dict, ocalib_scale @staticmethod def build_subgraph(op_type): # Cast input, weights (and bias) into float. t_names = ["X", "", "W", ""] if "InputConv" in op_type: t_names[1] = "x_pad_value" if "Biased" in op_type: t_names[-1] = "bias" nodes, t_names = cast_tensors_to(t_names) # Pad + convolution nodes += get_padding_ops(t_names[0], "Xi", t_names[1]) conv_tensor_names = nodes[-1].output[:1] + t_names[2:] nodes.append(make_node("Conv", inputs=conv_tensor_names, outputs=["Yi"])) nodes[-1].attribute.append(AP(name="strides", ref_attr_name="strides", type=AP.INTS)) # Maxpool (optional) if "MaxPool" in op_type: # Replace previous output as maxpool input nodes[-1].output.__setitem__(0, nodes[-1].op_type) nodes += get_pool_ops(nodes[-1].output[0], "Yi", pool_op_type="MaxPool") # Activation (optional) if "ReLU" in op_type: # Replace previous output as relu input nodes[-1].output.__setitem__(0, nodes[-1].op_type) nodes += get_activation_ops(nodes[-1].output[0], "Yi", "ReLUClipped" in op_type) # AvgPool (optional) if "GlobalAvgPool" in op_type: # Replace previous output as maxpool input nodes[-1].output.__setitem__(0, nodes[-1].op_type) nodes += get_pool_ops(nodes[-1].output[0], "Yi", pool_op_type="GlobalAvgPool") # Scale out (with saturation) in float domain shift_nodes, shift_t_names = cast_tensors_to(["Scale", "Shift"]) nodes += shift_nodes nodes += get_scale_out_ops("Yi", "Yscaled", *shift_t_names) # Cast output to expect type nodes.append(make_node("Cast", ["Yscaled"], ["Y"], to=TP.INT8)) return nodes