Note
Click here to download the full example code
AkidaNet/ImageNet inference
This CNN2SNN tutorial presents how to convert an AkidaNet pre-trained model into Akida.
As ImageNet images are not publicly available, performances are assessed using a set of 10 copyright free images that were found on Google using ImageNet class names.
1. Dataset preparation
Test images all have at least 256 pixels in the smallest dimension. They must
be preprocessed to fit in the model. The imagenet.preprocessing.preprocess_image
function decodes, crops and extracts a square 224x224x3 patch from an input image.
Note
Input size is here set to 224x224x3 as this is what is used by the model presented in the next section.
import akida
import os
import numpy as np
from tensorflow.io import read_file
from tensorflow.image import decode_jpeg
from tensorflow.keras.utils import get_file
from akida_models.imagenet import preprocessing
# Model specification and hyperparameters
NUM_CHANNELS = 3
IMAGE_SIZE = 224
NUM_CLASSES = 1000
num_images = 10
# Retrieve dataset file from Brainchip data server
file_path = get_file(
"imagenet_like.zip",
"http://data.brainchip.com/dataset-mirror/imagenet_like/imagenet_like.zip",
cache_subdir='datasets/imagenet_like',
extract=True)
data_folder = os.path.dirname(file_path)
# Load images for test set
x_test_files = []
x_test = np.zeros((num_images, 224, 224, 3)).astype('uint8')
for id in range(num_images):
test_file = 'image_' + str(id + 1).zfill(2) + '.jpg'
x_test_files.append(test_file)
img_path = os.path.join(data_folder, test_file)
base_image = read_file(img_path)
image = decode_jpeg(base_image, channels=NUM_CHANNELS)
image = preprocessing.preprocess_image(image, IMAGE_SIZE)
x_test[id, :, :, :] = np.expand_dims(image, axis=0)
print(f'{num_images} images loaded and preprocessed.')
Downloading data from http://data.brainchip.com/dataset-mirror/imagenet_like/imagenet_like.zip
8192/20418307 [..............................] - ETA: 0s
204800/20418307 [..............................] - ETA: 4s
499712/20418307 [..............................] - ETA: 4s
974848/20418307 [>.............................] - ETA: 3s
1417216/20418307 [=>............................] - ETA: 2s
1908736/20418307 [=>............................] - ETA: 2s
2383872/20418307 [==>...........................] - ETA: 2s
2711552/20418307 [==>...........................] - ETA: 2s
3129344/20418307 [===>..........................] - ETA: 2s
3481600/20418307 [====>.........................] - ETA: 2s
3850240/20418307 [====>.........................] - ETA: 2s
4227072/20418307 [=====>........................] - ETA: 2s
4603904/20418307 [=====>........................] - ETA: 2s
4988928/20418307 [======>.......................] - ETA: 2s
5390336/20418307 [======>.......................] - ETA: 1s
5808128/20418307 [=======>......................] - ETA: 1s
6234112/20418307 [========>.....................] - ETA: 1s
6660096/20418307 [========>.....................] - ETA: 1s
7094272/20418307 [=========>....................] - ETA: 1s
7544832/20418307 [==========>...................] - ETA: 1s
7995392/20418307 [==========>...................] - ETA: 1s
8454144/20418307 [===========>..................] - ETA: 1s
8937472/20418307 [============>.................] - ETA: 1s
9412608/20418307 [============>.................] - ETA: 1s
9904128/20418307 [=============>................] - ETA: 1s
10395648/20418307 [==============>...............] - ETA: 1s
10895360/20418307 [===============>..............] - ETA: 1s
11419648/20418307 [===============>..............] - ETA: 1s
11919360/20418307 [================>.............] - ETA: 1s
12451840/20418307 [=================>............] - ETA: 0s
12992512/20418307 [==================>...........] - ETA: 0s
13541376/20418307 [==================>...........] - ETA: 0s
14073856/20418307 [===================>..........] - ETA: 0s
14639104/20418307 [====================>.........] - ETA: 0s
15196160/20418307 [=====================>........] - ETA: 0s
15769600/20418307 [======================>.......] - ETA: 0s
16351232/20418307 [=======================>......] - ETA: 0s
16941056/20418307 [=======================>......] - ETA: 0s
17530880/20418307 [========================>.....] - ETA: 0s
18120704/20418307 [=========================>....] - ETA: 0s
18710528/20418307 [==========================>...] - ETA: 0s
19292160/20418307 [===========================>..] - ETA: 0s
19881984/20418307 [============================>.] - ETA: 0s
20418307/20418307 [==============================] - 2s 0us/step
10 images loaded and preprocessed.
Labels for test images are stored in the akida_models package. The matching
between names (string) and labels (integer) is given through the
imagenet.preprocessing.index_to_label method.
import csv
# Parse labels file
fname = os.path.join(data_folder, 'labels_validation.txt')
validation_labels = dict()
with open(fname, newline='') as csvfile:
reader = csv.reader(csvfile, delimiter=' ')
for row in reader:
validation_labels[row[0]] = row[1]
# Get labels for the test set by index
labels_test = np.zeros(num_images)
for i in range(num_images):
labels_test[i] = int(validation_labels[x_test_files[i]])
2. Create a Keras AkidaNet model
The AkidaNet architecture is available in the Akida model zoo as akidanet_imagenet. In this tutorial, the alpha = 0.5 version of AkidaNet will be used, where alpha is the parameter controlling the width of the model.
from tensorflow.keras.models import load_model
# Retrieve the float model with pretrained weights and load it
model_file = get_file(
"akidanet_imagenet_224_alpha_50.h5",
"http://data.brainchip.com/models/akidanet/akidanet_imagenet_224_alpha_50.h5",
cache_subdir='models/akidanet_imagenet')
model_keras = load_model(model_file)
model_keras.summary()
Downloading data from http://data.brainchip.com/models/akidanet/akidanet_imagenet_224_alpha_50.h5
8192/5714072 [..............................] - ETA: 0s
212992/5714072 [>.............................] - ETA: 1s
720896/5714072 [==>...........................] - ETA: 0s
1212416/5714072 [=====>........................] - ETA: 0s
1703936/5714072 [=======>......................] - ETA: 0s
2195456/5714072 [==========>...................] - ETA: 0s
2711552/5714072 [=============>................] - ETA: 0s
3039232/5714072 [==============>...............] - ETA: 0s
3440640/5714072 [=================>............] - ETA: 0s
3776512/5714072 [==================>...........] - ETA: 0s
3989504/5714072 [===================>..........] - ETA: 0s
4202496/5714072 [=====================>........] - ETA: 0s
4423680/5714072 [======================>.......] - ETA: 0s
4644864/5714072 [=======================>......] - ETA: 0s
4849664/5714072 [========================>.....] - ETA: 0s
5103616/5714072 [=========================>....] - ETA: 0s
5349376/5714072 [===========================>..] - ETA: 0s
5611520/5714072 [============================>.] - ETA: 0s
5714072/5714072 [==============================] - 1s 0us/step
Model: "akidanet_0.50_224_1000"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 224, 224, 3)] 0
rescaling (Rescaling) (None, 224, 224, 3) 0
conv_0 (Conv2D) (None, 112, 112, 16) 432
conv_0/BN (BatchNormalizati (None, 112, 112, 16) 64
on)
conv_0/relu (ReLU) (None, 112, 112, 16) 0
conv_1 (Conv2D) (None, 112, 112, 32) 4608
conv_1/BN (BatchNormalizati (None, 112, 112, 32) 128
on)
conv_1/relu (ReLU) (None, 112, 112, 32) 0
conv_2 (Conv2D) (None, 56, 56, 64) 18432
conv_2/BN (BatchNormalizati (None, 56, 56, 64) 256
on)
conv_2/relu (ReLU) (None, 56, 56, 64) 0
conv_3 (Conv2D) (None, 56, 56, 64) 36864
conv_3/BN (BatchNormalizati (None, 56, 56, 64) 256
on)
conv_3/relu (ReLU) (None, 56, 56, 64) 0
separable_4 (SeparableConv2 (None, 28, 28, 128) 8768
D)
separable_4/BN (BatchNormal (None, 28, 28, 128) 512
ization)
separable_4/relu (ReLU) (None, 28, 28, 128) 0
separable_5 (SeparableConv2 (None, 28, 28, 128) 17536
D)
separable_5/BN (BatchNormal (None, 28, 28, 128) 512
ization)
separable_5/relu (ReLU) (None, 28, 28, 128) 0
separable_6 (SeparableConv2 (None, 14, 14, 256) 33920
D)
separable_6/BN (BatchNormal (None, 14, 14, 256) 1024
ization)
separable_6/relu (ReLU) (None, 14, 14, 256) 0
separable_7 (SeparableConv2 (None, 14, 14, 256) 67840
D)
separable_7/BN (BatchNormal (None, 14, 14, 256) 1024
ization)
separable_7/relu (ReLU) (None, 14, 14, 256) 0
separable_8 (SeparableConv2 (None, 14, 14, 256) 67840
D)
separable_8/BN (BatchNormal (None, 14, 14, 256) 1024
ization)
separable_8/relu (ReLU) (None, 14, 14, 256) 0
separable_9 (SeparableConv2 (None, 14, 14, 256) 67840
D)
separable_9/BN (BatchNormal (None, 14, 14, 256) 1024
ization)
separable_9/relu (ReLU) (None, 14, 14, 256) 0
separable_10 (SeparableConv (None, 14, 14, 256) 67840
2D)
separable_10/BN (BatchNorma (None, 14, 14, 256) 1024
lization)
separable_10/relu (ReLU) (None, 14, 14, 256) 0
separable_11 (SeparableConv (None, 14, 14, 256) 67840
2D)
separable_11/BN (BatchNorma (None, 14, 14, 256) 1024
lization)
separable_11/relu (ReLU) (None, 14, 14, 256) 0
separable_12 (SeparableConv (None, 7, 7, 512) 133376
2D)
separable_12/BN (BatchNorma (None, 7, 7, 512) 2048
lization)
separable_12/relu (ReLU) (None, 7, 7, 512) 0
separable_13 (SeparableConv (None, 7, 7, 512) 266752
2D)
separable_13/global_avg (Gl (None, 512) 0
obalAveragePooling2D)
separable_13/BN (BatchNorma (None, 512) 2048
lization)
separable_13/relu (ReLU) (None, 512) 0
reshape_1 (Reshape) (None, 1, 1, 512) 0
dropout (Dropout) (None, 1, 1, 512) 0
separable_14 (SeparableConv (None, 1, 1, 1000) 516608
2D)
act_softmax (Activation) (None, 1, 1, 1000) 0
reshape_2 (Reshape) (None, 1000) 0
=================================================================
Total params: 1,388,464
Trainable params: 1,382,480
Non-trainable params: 5,984
_________________________________________________________________
Top-1 accuracy on the actual ImageNet is 64.58%, the perfomance given below uses the 10 images subset.
from timeit import default_timer as timer
# Check model performance
def check_model_performances(model, x_test=x_test, labels_test=labels_test):
num_images = len(x_test)
start = timer()
potentials_keras = model.predict(x_test, batch_size=100)
end = timer()
print(f'Keras inference on {num_images} images took {end-start:.2f} s.\n')
preds_keras = np.squeeze(np.argmax(potentials_keras, 1))
accuracy_keras = np.sum(np.equal(preds_keras, labels_test)) / num_images
print(f"Keras accuracy: {accuracy_keras*100:.2f} %")
check_model_performances(model_keras)
1/1 [==============================] - ETA: 0s
1/1 [==============================] - 0s 392ms/step
Keras inference on 10 images took 0.42 s.
Keras accuracy: 100.00 %
3. Quantized model
Quantizing a model is done using cnn2snn.quantize.
The quantized model satisfies the Akida NSoC requirements:
the first layer has 8-bit weights,
all other convolutional layers have 4-bit weights,
all convolutional layers have 4-bit activations.
However, this model will suffer from a drop in accuracy due to quantization as shown in the table below for ImageNet and in the next cell for the 10 images set.
Float accuracy |
Quantized accuracy |
|---|---|
64.58 % |
1.00 % |
from cnn2snn import quantize
# Quantize the model to 4-bit weights and activations, 8-bit weights for the
# first convolutional layer
model_keras_quantized = quantize(model_keras, 4, 4, 8)
# Check Model performance
check_model_performances(model_keras_quantized)
1/1 [==============================] - ETA: 0s
1/1 [==============================] - 1s 594ms/step
Keras inference on 10 images took 0.62 s.
Keras accuracy: 0.00 %
4. Pretrained quantized model
The Akida models zoo also contains a pretrained quantized helper that was obtained after fine tuning the model for 10 epochs.
Tuning the model, that is training with a lowered learning rate, allows to recover performances up to the initial floating point accuracy.
Performances on the full ImageNet dataset are:
Float accuracy |
Quantized accuracy |
After tuning |
|---|---|---|
64.58 % |
1.00 % |
61.30 % |
from akida_models import akidanet_imagenet_pretrained
# Use a quantized model with pretrained quantized weights
model_keras_quantized_pretrained = akidanet_imagenet_pretrained(0.5)
model_keras_quantized_pretrained.summary()
Downloading data from http://data.brainchip.com/models/akidanet/akidanet_imagenet_224_alpha_50_iq8_wq4_aq4.h5.
0/5612520 [..............................] - ETA: 0s
204800/5612520 [>.............................] - ETA: 1s
630784/5612520 [==>...........................] - ETA: 0s
1073152/5612520 [====>.........................] - ETA: 0s
1622016/5612520 [=======>......................] - ETA: 0s
2211840/5612520 [==========>...................] - ETA: 0s
2801664/5612520 [=============>................] - ETA: 0s
3391488/5612520 [=================>............] - ETA: 0s
3973120/5612520 [====================>.........] - ETA: 0s
4562944/5612520 [=======================>......] - ETA: 0s
5152768/5612520 [==========================>...] - ETA: 0s
5612520/5612520 [==============================] - 1s 0us/step
Model: "akidanet_0.50_224_1000"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 224, 224, 3)] 0
rescaling (Rescaling) (None, 224, 224, 3) 0
conv_0 (QuantizedConv2D) (None, 112, 112, 16) 448
conv_0/relu (ActivationDisc (None, 112, 112, 16) 0
reteRelu)
conv_1 (QuantizedConv2D) (None, 112, 112, 32) 4640
conv_1/relu (ActivationDisc (None, 112, 112, 32) 0
reteRelu)
conv_2 (QuantizedConv2D) (None, 56, 56, 64) 18496
conv_2/relu (ActivationDisc (None, 56, 56, 64) 0
reteRelu)
conv_3 (QuantizedConv2D) (None, 56, 56, 64) 36928
conv_3/relu (ActivationDisc (None, 56, 56, 64) 0
reteRelu)
separable_4 (QuantizedSepar (None, 28, 28, 128) 8896
ableConv2D)
separable_4/relu (Activatio (None, 28, 28, 128) 0
nDiscreteRelu)
separable_5 (QuantizedSepar (None, 28, 28, 128) 17664
ableConv2D)
separable_5/relu (Activatio (None, 28, 28, 128) 0
nDiscreteRelu)
separable_6 (QuantizedSepar (None, 14, 14, 256) 34176
ableConv2D)
separable_6/relu (Activatio (None, 14, 14, 256) 0
nDiscreteRelu)
separable_7 (QuantizedSepar (None, 14, 14, 256) 68096
ableConv2D)
separable_7/relu (Activatio (None, 14, 14, 256) 0
nDiscreteRelu)
separable_8 (QuantizedSepar (None, 14, 14, 256) 68096
ableConv2D)
separable_8/relu (Activatio (None, 14, 14, 256) 0
nDiscreteRelu)
separable_9 (QuantizedSepar (None, 14, 14, 256) 68096
ableConv2D)
separable_9/relu (Activatio (None, 14, 14, 256) 0
nDiscreteRelu)
separable_10 (QuantizedSepa (None, 14, 14, 256) 68096
rableConv2D)
separable_10/relu (Activati (None, 14, 14, 256) 0
onDiscreteRelu)
separable_11 (QuantizedSepa (None, 14, 14, 256) 68096
rableConv2D)
separable_11/relu (Activati (None, 14, 14, 256) 0
onDiscreteRelu)
separable_12 (QuantizedSepa (None, 7, 7, 512) 133888
rableConv2D)
separable_12/relu (Activati (None, 7, 7, 512) 0
onDiscreteRelu)
separable_13 (QuantizedSepa (None, 7, 7, 512) 267264
rableConv2D)
separable_13/global_avg (Gl (None, 512) 0
obalAveragePooling2D)
separable_13/relu (Activati (None, 512) 0
onDiscreteRelu)
reshape_1 (Reshape) (None, 1, 1, 512) 0
dropout (Dropout) (None, 1, 1, 512) 0
separable_14 (QuantizedSepa (None, 1, 1, 1000) 516608
rableConv2D)
act_softmax (Activation) (None, 1, 1, 1000) 0
reshape_2 (Reshape) (None, 1000) 0
=================================================================
Total params: 1,379,488
Trainable params: 1,379,488
Non-trainable params: 0
_________________________________________________________________
# Check model performance
check_model_performances(model_keras_quantized_pretrained)
1/1 [==============================] - ETA: 0s
1/1 [==============================] - 1s 598ms/step
Keras inference on 10 images took 0.62 s.
Keras accuracy: 90.00 %
5. Conversion to Akida
5.1 Convert to Akida model
Here, the Keras quantized model is converted into a suitable version for the Akida NSoC. The cnn2snn.convert function needs as arguments the Keras model and the input scaling parameters.
from cnn2snn import convert
model_akida = convert(model_keras_quantized_pretrained)
The Model.summary method provides a detailed description of the Model layers.
model_akida.summary()
Model Summary
________________________________________________
Input shape Output shape Sequences Layers
================================================
[224, 224, 3] [1, 1, 1000] 1 15
________________________________________________
_____________________________________________________________
Layer (type) Output shape Kernel shape
============= SW/conv_0-separable_14 (Software) =============
conv_0 (InputConv.) [112, 112, 16] (3, 3, 3, 16)
_____________________________________________________________
conv_1 (Conv.) [112, 112, 32] (3, 3, 16, 32)
_____________________________________________________________
conv_2 (Conv.) [56, 56, 64] (3, 3, 32, 64)
_____________________________________________________________
conv_3 (Conv.) [56, 56, 64] (3, 3, 64, 64)
_____________________________________________________________
separable_4 (Sep.Conv.) [28, 28, 128] (3, 3, 64, 1)
_____________________________________________________________
(1, 1, 64, 128)
_____________________________________________________________
separable_5 (Sep.Conv.) [28, 28, 128] (3, 3, 128, 1)
_____________________________________________________________
(1, 1, 128, 128)
_____________________________________________________________
separable_6 (Sep.Conv.) [14, 14, 256] (3, 3, 128, 1)
_____________________________________________________________
(1, 1, 128, 256)
_____________________________________________________________
separable_7 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1)
_____________________________________________________________
(1, 1, 256, 256)
_____________________________________________________________
separable_8 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1)
_____________________________________________________________
(1, 1, 256, 256)
_____________________________________________________________
separable_9 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1)
_____________________________________________________________
(1, 1, 256, 256)
_____________________________________________________________
separable_10 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1)
_____________________________________________________________
(1, 1, 256, 256)
_____________________________________________________________
separable_11 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1)
_____________________________________________________________
(1, 1, 256, 256)
_____________________________________________________________
separable_12 (Sep.Conv.) [7, 7, 512] (3, 3, 256, 1)
_____________________________________________________________
(1, 1, 256, 512)
_____________________________________________________________
separable_13 (Sep.Conv.) [1, 1, 512] (3, 3, 512, 1)
_____________________________________________________________
(1, 1, 512, 512)
_____________________________________________________________
separable_14 (Sep.Conv.) [1, 1, 1000] (3, 3, 512, 1)
_____________________________________________________________
(1, 1, 512, 1000)
_____________________________________________________________
5.2 Check performance
While we compute accuracy for the 10 images set in the next cell, the following table summarizes results obtained on ImageNet.
Keras accuracy |
Akida accuracy |
|---|---|
61.30 % |
61.37 % |
# Check Model performance
start = timer()
accuracy_akida = model_akida.evaluate(x_test, labels_test)
end = timer()
print(f'Inference on {num_images} images took {end-start:.2f} s.\n')
print(f"Accuracy: {accuracy_akida*100:.2f} %")
# For non-regression purpose
assert accuracy_akida >= 0.89
Inference on 10 images took 0.30 s.
Accuracy: 90.00 %
5.3 Show predictions for a random image
import matplotlib.pyplot as plt
import matplotlib.lines as lines
# Functions used to display the top5 results
def get_top5(potentials, true_label):
"""
Returns the top 5 classes from the output potentials
"""
tmp_pots = potentials.copy()
top5 = []
min_val = np.min(tmp_pots)
for ii in range(5):
best = np.argmax(tmp_pots)
top5.append(best)
tmp_pots[best] = min_val
vals = np.zeros((6,))
vals[:5] = potentials[top5]
if true_label not in top5:
vals[5] = potentials[true_label]
else:
vals[5] = 0
vals /= np.max(vals)
class_name = []
for ii in range(5):
class_name.append(preprocessing.index_to_label(top5[ii]).split(',')[0])
if true_label in top5:
class_name.append('')
else:
class_name.append(
preprocessing.index_to_label(true_label).split(',')[0])
return top5, vals, class_name
def adjust_spines(ax, spines):
for loc, spine in ax.spines.items():
if loc in spines:
spine.set_position(('outward', 10)) # outward by 10 points
else:
spine.set_color('none') # don't draw spine
# turn off ticks where there is no spine
if 'left' in spines:
ax.yaxis.set_ticks_position('left')
else:
# no yaxis ticks
ax.yaxis.set_ticks([])
if 'bottom' in spines:
ax.xaxis.set_ticks_position('bottom')
else:
# no xaxis ticks
ax.xaxis.set_ticks([])
def prepare_plots():
fig = plt.figure(figsize=(8, 4))
# Image subplot
ax0 = plt.subplot(1, 3, 1)
imgobj = ax0.imshow(np.zeros((224, 224, 3), dtype=np.uint8))
ax0.set_axis_off()
# Top 5 results subplot
ax1 = plt.subplot(1, 2, 2)
bar_positions = (0, 1, 2, 3, 4, 6)
rects = ax1.barh(bar_positions, np.zeros((6,)), align='center', height=0.5)
plt.xlim(-0.2, 1.01)
ax1.set(xlim=(-0.2, 1.15), ylim=(-1.5, 12))
ax1.set_yticks(bar_positions)
ax1.invert_yaxis()
ax1.yaxis.set_ticks_position('left')
ax1.xaxis.set_ticks([])
adjust_spines(ax1, 'left')
ax1.add_line(lines.Line2D((0, 0), (-0.5, 6.5), color=(0.0, 0.0, 0.0)))
# Adjust Plot Positions
ax0.set_position([0.05, 0.055, 0.3, 0.9])
l1, b1, w1, h1 = ax1.get_position().bounds
ax1.set_position([l1 * 1.05, b1 + 0.09 * h1, w1, 0.8 * h1])
# Add title box
plt.figtext(0.5,
0.9,
"Imagenet Classification by Akida",
size=20,
ha="center",
va="center",
bbox=dict(boxstyle="round",
ec=(0.5, 0.5, 0.5),
fc=(0.9, 0.9, 1.0)))
return fig, imgobj, ax1, rects
def update_bars_chart(rects, vals, true_label):
counter = 0
for rect, h in zip(rects, yvals):
rect.set_width(h)
if counter < 5:
if top5[counter] == true_label:
if counter == 0:
rect.set_facecolor((0.0, 1.0, 0.0))
else:
rect.set_facecolor((0.0, 0.5, 0.0))
else:
rect.set_facecolor('gray')
elif counter == 5:
rect.set_facecolor('red')
counter += 1
# Prepare plots
fig, imgobj, ax1, rects = prepare_plots()
# Get a random image
img = np.random.randint(num_images)
# Predict image class
potentials_akida = model_akida.predict(np.expand_dims(x_test[img],
axis=0)).squeeze()
# Get top 5 prediction labels and associated names
true_label = int(validation_labels[x_test_files[img]])
top5, yvals, class_name = get_top5(potentials_akida, true_label)
# Draw Plots
imgobj.set_data(x_test[img])
ax1.set_yticklabels(class_name, rotation='horizontal', size=9)
update_bars_chart(rects, yvals, true_label)
fig.canvas.draw()
plt.show()
6. Hardware mapping and performance
6.1. Map on hardware
List Akida available devices and check that an NSoC V2 (production chip) is available
devices = akida.devices()
print(f'Available devices: {[dev.desc for dev in devices]}')
device = devices[0]
assert device.version == akida.NSoC_v2
Available devices: ['PCIe/NSoC_v2/0']
Map the model on the device
model_akida.map(device)
# Check model mapping: NP allocation and binary size
model_akida.summary()
Model Summary
_____________________________________________________
Input shape Output shape Sequences Layers NPs
=====================================================
[224, 224, 3] [1, 1, 1000] 1 15 38
_____________________________________________________
__________________________________________________________________
Layer (type) Output shape Kernel shape NPs
===== HW/conv_0-separable_14 (Hardware) - size: 1241424 bytes ====
conv_0 (InputConv.) [112, 112, 16] (3, 3, 3, 16) N/A
__________________________________________________________________
conv_1 (Conv.) [112, 112, 32] (3, 3, 16, 32) 4
__________________________________________________________________
conv_2 (Conv.) [56, 56, 64] (3, 3, 32, 64) 6
__________________________________________________________________
conv_3 (Conv.) [56, 56, 64] (3, 3, 64, 64) 3
__________________________________________________________________
separable_4 (Sep.Conv.) [28, 28, 128] (3, 3, 64, 1) 3
__________________________________________________________________
(1, 1, 64, 128)
__________________________________________________________________
separable_5 (Sep.Conv.) [28, 28, 128] (3, 3, 128, 1) 2
__________________________________________________________________
(1, 1, 128, 128)
__________________________________________________________________
separable_6 (Sep.Conv.) [14, 14, 256] (3, 3, 128, 1) 2
__________________________________________________________________
(1, 1, 128, 256)
__________________________________________________________________
separable_7 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1) 1
__________________________________________________________________
(1, 1, 256, 256)
__________________________________________________________________
separable_8 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1) 1
__________________________________________________________________
(1, 1, 256, 256)
__________________________________________________________________
separable_9 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1) 1
__________________________________________________________________
(1, 1, 256, 256)
__________________________________________________________________
separable_10 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1) 1
__________________________________________________________________
(1, 1, 256, 256)
__________________________________________________________________
separable_11 (Sep.Conv.) [14, 14, 256] (3, 3, 256, 1) 1
__________________________________________________________________
(1, 1, 256, 256)
__________________________________________________________________
separable_12 (Sep.Conv.) [7, 7, 512] (3, 3, 256, 1) 2
__________________________________________________________________
(1, 1, 256, 512)
__________________________________________________________________
separable_13 (Sep.Conv.) [1, 1, 512] (3, 3, 512, 1) 4
__________________________________________________________________
(1, 1, 512, 512)
__________________________________________________________________
separable_14 (Sep.Conv.) [1, 1, 1000] (3, 3, 512, 1) 7
__________________________________________________________________
(1, 1, 512, 1000)
__________________________________________________________________
6.2. Performances measurement
Power measurement must be enabled on the device’ soc (disabled by default). After sending data for inference, performances measurements are available in the model statistics.
# Enable power measurement
device.soc.power_measurement_enabled = True
# Send data for inference
_ = model_akida.forward(x_test)
# Display floor current
floor_power = device.soc.power_meter.floor
print(f'Floor power: {floor_power:.2f} mW')
# Retrieve statistics
print(model_akida.statistics)
Floor power: 881.10 mW
Average framerate = 22.99 fps
Last inference power range (mW): Avg 1033.83 / Min 881.00 / Max 1127.00 / Std 90.78
Last inference energy consumed (mJ/frame): 44.97
Total running time of the script: ( 0 minutes 14.123 seconds)