Note
Go to the end to download the full example code.
Segmentation tutorial
This example demonstrates image segmentation with an Akida-compatible model as illustrated through person segmentation using the Portrait128 dataset.
Using pre-trained models for quick runtime, this example shows the evolution of model performance for a trained TF-Keras floating-point model, a TF-Keras quantized and Quantization Aware Trained (QAT) model, and an Akida-converted model. Notice that the performance of the original TF-Keras floating-point model is maintained throughout the model conversion flow.
1. Load the dataset
import os
import numpy as np
from akida_models import fetch_file
# Download the validation set from the BrainChip data server, it contains 10% of the original
# dataset
data_path = fetch_file(fname="val.tar.gz",
origin="https://data.brainchip.com/dataset-mirror/portrait128/val.tar.gz",
cache_subdir=os.path.join("datasets", "portrait128"),
extract=True)
data_dir = os.path.join(os.path.dirname(data_path), "val")
x_val = np.load(os.path.join(data_dir, "val_img.npy"))
y_val = np.load(os.path.join(data_dir, "val_msk.npy")).astype('uint8')
batch_size = 32
steps = x_val.shape[0] // 32
# Visualize some data
import matplotlib.pyplot as plt
rng = np.random.default_rng()
id = rng.integers(0, x_val.shape[0] - 2)
fig, axs = plt.subplots(3, 3, constrained_layout=True)
for col in range(3):
axs[0, col].imshow(x_val[id + col] / 255.)
axs[0, col].axis('off')
axs[1, col].imshow(1 - y_val[id + col], cmap='Greys')
axs[1, col].axis('off')
axs[2, col].imshow(x_val[id + col] / 255. * y_val[id + col])
axs[2, col].axis('off')
fig.suptitle('Image, mask and masked image', fontsize=10)
plt.show()
Downloading data from https://data.brainchip.com/dataset-mirror/portrait128/val.tar.gz.
0/267313385 [..............................] - ETA: 0s
253952/267313385 [..............................] - ETA: 54s
1417216/267313385 [..............................] - ETA: 19s
2990080/267313385 [..............................] - ETA: 13s
4628480/267313385 [..............................] - ETA: 11s
5595136/267313385 [..............................] - ETA: 11s
6873088/267313385 [..............................] - ETA: 11s
8167424/267313385 [..............................] - ETA: 11s
9248768/267313385 [>.............................] - ETA: 11s
10797056/267313385 [>.............................] - ETA: 10s
11886592/267313385 [>.............................] - ETA: 10s
13230080/267313385 [>.............................] - ETA: 10s
14524416/267313385 [>.............................] - ETA: 10s
15294464/267313385 [>.............................] - ETA: 10s
16785408/267313385 [>.............................] - ETA: 10s
17899520/267313385 [=>............................] - ETA: 10s
18898944/267313385 [=>............................] - ETA: 10s
20127744/267313385 [=>............................] - ETA: 10s
21454848/267313385 [=>............................] - ETA: 10s
22781952/267313385 [=>............................] - ETA: 10s
24010752/267313385 [=>............................] - ETA: 10s
24952832/267313385 [=>............................] - ETA: 10s
26222592/267313385 [=>............................] - ETA: 10s
27598848/267313385 [==>...........................] - ETA: 10s
28762112/267313385 [==>...........................] - ETA: 10s
29843456/267313385 [==>...........................] - ETA: 10s
31105024/267313385 [==>...........................] - ETA: 10s
32350208/267313385 [==>...........................] - ETA: 10s
33431552/267313385 [==>...........................] - ETA: 10s
34709504/267313385 [==>...........................] - ETA: 9s
35921920/267313385 [===>..........................] - ETA: 9s
37003264/267313385 [===>..........................] - ETA: 9s
38313984/267313385 [===>..........................] - ETA: 9s
39739392/267313385 [===>..........................] - ETA: 9s
40886272/267313385 [===>..........................] - ETA: 9s
41967616/267313385 [===>..........................] - ETA: 9s
43294720/267313385 [===>..........................] - ETA: 9s
44146688/267313385 [===>..........................] - ETA: 9s
45490176/267313385 [====>.........................] - ETA: 9s
46563328/267313385 [====>.........................] - ETA: 9s
47718400/267313385 [====>.........................] - ETA: 9s
49012736/267313385 [====>.........................] - ETA: 9s
49963008/267313385 [====>.........................] - ETA: 9s
51290112/267313385 [====>.........................] - ETA: 9s
52502528/267313385 [====>.........................] - ETA: 9s
52518912/267313385 [====>.........................] - ETA: 16s
53895168/267313385 [=====>........................] - ETA: 16s
54894592/267313385 [=====>........................] - ETA: 15s
55713792/267313385 [=====>........................] - ETA: 15s
57425920/267313385 [=====>........................] - ETA: 15s
58662912/267313385 [=====>........................] - ETA: 15s
59940864/267313385 [=====>........................] - ETA: 14s
60989440/267313385 [=====>........................] - ETA: 14s
62087168/267313385 [=====>........................] - ETA: 14s
63414272/267313385 [======>.......................] - ETA: 14s
64315392/267313385 [======>.......................] - ETA: 14s
65609728/267313385 [======>.......................] - ETA: 14s
66969600/267313385 [======>.......................] - ETA: 13s
68313088/267313385 [======>.......................] - ETA: 13s
69394432/267313385 [======>.......................] - ETA: 13s
70721536/267313385 [======>.......................] - ETA: 13s
71999488/267313385 [=======>......................] - ETA: 13s
72867840/267313385 [=======>......................] - ETA: 12s
74194944/267313385 [=======>......................] - ETA: 12s
75423744/267313385 [=======>......................] - ETA: 12s
76505088/267313385 [=======>......................] - ETA: 12s
77864960/267313385 [=======>......................] - ETA: 12s
78782464/267313385 [=======>......................] - ETA: 12s
80076800/267313385 [=======>......................] - ETA: 12s
81092608/267313385 [========>.....................] - ETA: 11s
82452480/267313385 [========>.....................] - ETA: 11s
83566592/267313385 [========>.....................] - ETA: 11s
84893696/267313385 [========>.....................] - ETA: 11s
86073344/267313385 [========>.....................] - ETA: 11s
87318528/267313385 [========>.....................] - ETA: 11s
88776704/267313385 [========>.....................] - ETA: 11s
89858048/267313385 [=========>....................] - ETA: 11s
91299840/267313385 [=========>....................] - ETA: 10s
92741632/267313385 [=========>....................] - ETA: 10s
93782016/267313385 [=========>....................] - ETA: 10s
95100928/267313385 [=========>....................] - ETA: 10s
96198656/267313385 [=========>....................] - ETA: 10s
97525760/267313385 [=========>....................] - ETA: 10s
98721792/267313385 [==========>...................] - ETA: 10s
100261888/267313385 [==========>...................] - ETA: 9s
101277696/267313385 [==========>...................] - ETA: 9s
102768640/267313385 [==========>...................] - ETA: 9s
103915520/267313385 [==========>...................] - ETA: 9s
105340928/267313385 [==========>...................] - ETA: 9s
106749952/267313385 [==========>...................] - ETA: 9s
107945984/267313385 [===========>..................] - ETA: 9s
109273088/267313385 [===========>..................] - ETA: 9s
110501888/267313385 [===========>..................] - ETA: 9s
111616000/267313385 [===========>..................] - ETA: 9s
112975872/267313385 [===========>..................] - ETA: 8s
114188288/267313385 [===========>..................] - ETA: 8s
115662848/267313385 [===========>..................] - ETA: 8s
116760576/267313385 [============>.................] - ETA: 8s
118177792/267313385 [============>.................] - ETA: 8s
119414784/267313385 [============>.................] - ETA: 8s
120627200/267313385 [============>.................] - ETA: 8s
122019840/267313385 [============>.................] - ETA: 8s
123101184/267313385 [============>.................] - ETA: 8s
124542976/267313385 [============>.................] - ETA: 7s
125657088/267313385 [=============>................] - ETA: 7s
126984192/267313385 [=============>................] - ETA: 7s
128327680/267313385 [=============>................] - ETA: 7s
129441792/267313385 [=============>................] - ETA: 7s
130965504/267313385 [=============>................] - ETA: 7s
132145152/267313385 [=============>................] - ETA: 7s
133603328/267313385 [=============>................] - ETA: 7s
134586368/267313385 [==============>...............] - ETA: 7s
135290880/267313385 [==============>...............] - ETA: 7s
136896512/267313385 [==============>...............] - ETA: 7s
138076160/267313385 [==============>...............] - ETA: 7s
139304960/267313385 [==============>...............] - ETA: 6s
140582912/267313385 [==============>...............] - ETA: 6s
141877248/267313385 [==============>...............] - ETA: 6s
143138816/267313385 [===============>..............] - ETA: 6s
144318464/267313385 [===============>..............] - ETA: 6s
145301504/267313385 [===============>..............] - ETA: 6s
146497536/267313385 [===============>..............] - ETA: 6s
147759104/267313385 [===============>..............] - ETA: 6s
148889600/267313385 [===============>..............] - ETA: 6s
149921792/267313385 [===============>..............] - ETA: 6s
151150592/267313385 [===============>..............] - ETA: 6s
152477696/267313385 [================>.............] - ETA: 6s
153772032/267313385 [================>.............] - ETA: 6s
154836992/267313385 [================>.............] - ETA: 5s
155967488/267313385 [================>.............] - ETA: 5s
157245440/267313385 [================>.............] - ETA: 5s
158244864/267313385 [================>.............] - ETA: 5s
159424512/267313385 [================>.............] - ETA: 5s
160702464/267313385 [=================>............] - ETA: 5s
161734656/267313385 [=================>............] - ETA: 5s
162947072/267313385 [=================>............] - ETA: 5s
164257792/267313385 [=================>............] - ETA: 5s
165502976/267313385 [=================>............] - ETA: 5s
166567936/267313385 [=================>............] - ETA: 5s
167944192/267313385 [=================>............] - ETA: 5s
169271296/267313385 [=================>............] - ETA: 5s
170336256/267313385 [==================>...........] - ETA: 5s
171499520/267313385 [==================>...........] - ETA: 5s
172859392/267313385 [==================>...........] - ETA: 4s
174088192/267313385 [==================>...........] - ETA: 4s
175251456/267313385 [==================>...........] - ETA: 4s
176627712/267313385 [==================>...........] - ETA: 4s
177848320/267313385 [==================>...........] - ETA: 4s
178855936/267313385 [===================>..........] - ETA: 4s
180183040/267313385 [===================>..........] - ETA: 4s
181207040/267313385 [===================>..........] - ETA: 4s
182394880/267313385 [===================>..........] - ETA: 4s
183590912/267313385 [===================>..........] - ETA: 4s
184606720/267313385 [===================>..........] - ETA: 4s
185982976/267313385 [===================>..........] - ETA: 4s
186900480/267313385 [===================>..........] - ETA: 4s
188227584/267313385 [====================>.........] - ETA: 4s
189587456/267313385 [====================>.........] - ETA: 3s
190586880/267313385 [====================>.........] - ETA: 3s
191848448/267313385 [====================>.........] - ETA: 3s
192995328/267313385 [====================>.........] - ETA: 3s
194093056/267313385 [====================>.........] - ETA: 3s
195420160/267313385 [====================>.........] - ETA: 3s
196354048/267313385 [=====================>........] - ETA: 3s
197681152/267313385 [=====================>........] - ETA: 3s
198598656/267313385 [=====================>........] - ETA: 3s
199942144/267313385 [=====================>........] - ETA: 3s
201236480/267313385 [=====================>........] - ETA: 3s
202366976/267313385 [=====================>........] - ETA: 3s
203710464/267313385 [=====================>........] - ETA: 3s
204840960/267313385 [=====================>........] - ETA: 3s
206086144/267313385 [======================>.......] - ETA: 3s
207495168/267313385 [======================>.......] - ETA: 3s
208723968/267313385 [======================>.......] - ETA: 2s
209977344/267313385 [======================>.......] - ETA: 2s
211443712/267313385 [======================>.......] - ETA: 2s
212885504/267313385 [======================>.......] - ETA: 2s
214081536/267313385 [=======================>......] - ETA: 2s
215375872/267313385 [=======================>......] - ETA: 2s
216686592/267313385 [=======================>......] - ETA: 2s
217669632/267313385 [=======================>......] - ETA: 2s
219111424/267313385 [=======================>......] - ETA: 2s
220422144/267313385 [=======================>......] - ETA: 2s
221618176/267313385 [=======================>......] - ETA: 2s
223125504/267313385 [========================>.....] - ETA: 2s
224403456/267313385 [========================>.....] - ETA: 2s
225599488/267313385 [========================>.....] - ETA: 2s
226992128/267313385 [========================>.....] - ETA: 1s
228409344/267313385 [========================>.....] - ETA: 1s
229466112/267313385 [========================>.....] - ETA: 1s
230809600/267313385 [========================>.....] - ETA: 1s
232169472/267313385 [=========================>....] - ETA: 1s
233152512/267313385 [=========================>....] - ETA: 1s
234561536/267313385 [=========================>....] - ETA: 1s
235659264/267313385 [=========================>....] - ETA: 1s
236953600/267313385 [=========================>....] - ETA: 1s
238264320/267313385 [=========================>....] - ETA: 1s
239329280/267313385 [=========================>....] - ETA: 1s
240771072/267313385 [==========================>...] - ETA: 1s
241778688/267313385 [==========================>...] - ETA: 1s
243097600/267313385 [==========================>...] - ETA: 1s
244424704/267313385 [==========================>...] - ETA: 1s
245563392/267313385 [==========================>...] - ETA: 1s
247046144/267313385 [==========================>...] - ETA: 0s
248061952/267313385 [==========================>...] - ETA: 0s
249470976/267313385 [==========================>...] - ETA: 0s
250896384/267313385 [===========================>..] - ETA: 0s
251994112/267313385 [===========================>..] - ETA: 0s
253386752/267313385 [===========================>..] - ETA: 0s
254631936/267313385 [===========================>..] - ETA: 0s
255877120/267313385 [===========================>..] - ETA: 0s
257318912/267313385 [===========================>..] - ETA: 0s
258351104/267313385 [===========================>..] - ETA: 0s
259743744/267313385 [============================>.] - ETA: 0s
260808704/267313385 [============================>.] - ETA: 0s
262283264/267313385 [============================>.] - ETA: 0s
263446528/267313385 [============================>.] - ETA: 0s
264888320/267313385 [============================>.] - ETA: 0s
266264576/267313385 [============================>.] - ETA: 0s
267313385/267313385 [==============================] - 13s 0us/step
Download complete.
2. Load a pre-trained native TF-Keras model
The model used in this example is AkidaUNet. It has an AkidaNet (0.5) backbone to extract features combined with a succession of separable transposed convolutional blocks to build an image segmentation map. A pre-trained floating-point TF-Keras model is downloaded to save training time.
Note
The “transposed” convolutional feature is new in Akida 2.0.
The “separable transposed” operation is realized through the combination of a QuantizeML custom DepthwiseConv2DTranspose layer with a standard pointwise convolution.
The performance of the model is evaluated using both pixel accuracy and Binary IoU. The pixel accuracy describes how well the model can predict the segmentation mask pixel by pixel, and the Binary IoU takes into account how close the predicted mask is to the ground truth.
from akida_models.model_io import load_model
# Retrieve the model file from the BrainChip data server
model_file = fetch_file(fname="akida_unet_portrait128.h5",
origin="https://data.brainchip.com/models/AkidaV2/akida_unet/akida_unet_portrait128.h5",
cache_subdir='models')
# Load the native TF-Keras pre-trained model
model_keras = load_model(model_file)
model_keras.summary()
Downloading data from https://data.brainchip.com/models/AkidaV2/akida_unet/akida_unet_portrait128.h5.
0/4501976 [..............................] - ETA: 0s
204800/4501976 [>.............................] - ETA: 1s
1236992/4501976 [=======>......................] - ETA: 0s
2629632/4501976 [================>.............] - ETA: 0s
3317760/4501976 [=====================>........] - ETA: 0s
4501976/4501976 [==============================] - 0s 0us/step
Download complete.
Model: "akida_unet"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 128, 128, 3)] 0
rescaling (Rescaling) (None, 128, 128, 3) 0
conv_0 (Conv2D) (None, 64, 64, 16) 432
conv_0/BN (BatchNormalizat (None, 64, 64, 16) 64
ion)
conv_0/relu (ReLU) (None, 64, 64, 16) 0
conv_1 (Conv2D) (None, 64, 64, 32) 4608
conv_1/BN (BatchNormalizat (None, 64, 64, 32) 128
ion)
conv_1/relu (ReLU) (None, 64, 64, 32) 0
conv_2 (Conv2D) (None, 32, 32, 64) 18432
conv_2/BN (BatchNormalizat (None, 32, 32, 64) 256
ion)
conv_2/relu (ReLU) (None, 32, 32, 64) 0
conv_3 (Conv2D) (None, 32, 32, 64) 36864
conv_3/BN (BatchNormalizat (None, 32, 32, 64) 256
ion)
conv_3/relu (ReLU) (None, 32, 32, 64) 0
dw_separable_4 (DepthwiseC (None, 16, 16, 64) 576
onv2D)
pw_separable_4 (Conv2D) (None, 16, 16, 128) 8192
pw_separable_4/BN (BatchNo (None, 16, 16, 128) 512
rmalization)
pw_separable_4/relu (ReLU) (None, 16, 16, 128) 0
dw_separable_5 (DepthwiseC (None, 16, 16, 128) 1152
onv2D)
pw_separable_5 (Conv2D) (None, 16, 16, 128) 16384
pw_separable_5/BN (BatchNo (None, 16, 16, 128) 512
rmalization)
pw_separable_5/relu (ReLU) (None, 16, 16, 128) 0
dw_separable_6 (DepthwiseC (None, 8, 8, 128) 1152
onv2D)
pw_separable_6 (Conv2D) (None, 8, 8, 256) 32768
pw_separable_6/BN (BatchNo (None, 8, 8, 256) 1024
rmalization)
pw_separable_6/relu (ReLU) (None, 8, 8, 256) 0
dw_separable_7 (DepthwiseC (None, 8, 8, 256) 2304
onv2D)
pw_separable_7 (Conv2D) (None, 8, 8, 256) 65536
pw_separable_7/BN (BatchNo (None, 8, 8, 256) 1024
rmalization)
pw_separable_7/relu (ReLU) (None, 8, 8, 256) 0
dw_separable_8 (DepthwiseC (None, 8, 8, 256) 2304
onv2D)
pw_separable_8 (Conv2D) (None, 8, 8, 256) 65536
pw_separable_8/BN (BatchNo (None, 8, 8, 256) 1024
rmalization)
pw_separable_8/relu (ReLU) (None, 8, 8, 256) 0
dw_separable_9 (DepthwiseC (None, 8, 8, 256) 2304
onv2D)
pw_separable_9 (Conv2D) (None, 8, 8, 256) 65536
pw_separable_9/BN (BatchNo (None, 8, 8, 256) 1024
rmalization)
pw_separable_9/relu (ReLU) (None, 8, 8, 256) 0
dw_separable_10 (Depthwise (None, 8, 8, 256) 2304
Conv2D)
pw_separable_10 (Conv2D) (None, 8, 8, 256) 65536
pw_separable_10/BN (BatchN (None, 8, 8, 256) 1024
ormalization)
pw_separable_10/relu (ReLU (None, 8, 8, 256) 0
)
dw_separable_11 (Depthwise (None, 8, 8, 256) 2304
Conv2D)
pw_separable_11 (Conv2D) (None, 8, 8, 256) 65536
pw_separable_11/BN (BatchN (None, 8, 8, 256) 1024
ormalization)
pw_separable_11/relu (ReLU (None, 8, 8, 256) 0
)
dw_separable_12 (Depthwise (None, 4, 4, 256) 2304
Conv2D)
pw_separable_12 (Conv2D) (None, 4, 4, 512) 131072
pw_separable_12/BN (BatchN (None, 4, 4, 512) 2048
ormalization)
pw_separable_12/relu (ReLU (None, 4, 4, 512) 0
)
dw_separable_13 (Depthwise (None, 4, 4, 512) 4608
Conv2D)
pw_separable_13 (Conv2D) (None, 4, 4, 512) 262144
pw_separable_13/BN (BatchN (None, 4, 4, 512) 2048
ormalization)
pw_separable_13/relu (ReLU (None, 4, 4, 512) 0
)
dw_sepconv_t_0 (DepthwiseC (None, 8, 8, 512) 5120
onv2DTranspose)
pw_sepconv_t_0 (Conv2D) (None, 8, 8, 256) 131328
pw_sepconv_t_0/BN (BatchNo (None, 8, 8, 256) 1024
rmalization)
pw_sepconv_t_0/relu (ReLU) (None, 8, 8, 256) 0
dropout (Dropout) (None, 8, 8, 256) 0
dw_sepconv_t_1 (DepthwiseC (None, 16, 16, 256) 2560
onv2DTranspose)
pw_sepconv_t_1 (Conv2D) (None, 16, 16, 128) 32896
pw_sepconv_t_1/BN (BatchNo (None, 16, 16, 128) 512
rmalization)
pw_sepconv_t_1/relu (ReLU) (None, 16, 16, 128) 0
dropout_1 (Dropout) (None, 16, 16, 128) 0
dw_sepconv_t_2 (DepthwiseC (None, 32, 32, 128) 1280
onv2DTranspose)
pw_sepconv_t_2 (Conv2D) (None, 32, 32, 64) 8256
pw_sepconv_t_2/BN (BatchNo (None, 32, 32, 64) 256
rmalization)
pw_sepconv_t_2/relu (ReLU) (None, 32, 32, 64) 0
dropout_2 (Dropout) (None, 32, 32, 64) 0
dw_sepconv_t_3 (DepthwiseC (None, 64, 64, 64) 640
onv2DTranspose)
pw_sepconv_t_3 (Conv2D) (None, 64, 64, 32) 2080
pw_sepconv_t_3/BN (BatchNo (None, 64, 64, 32) 128
rmalization)
pw_sepconv_t_3/relu (ReLU) (None, 64, 64, 32) 0
dropout_3 (Dropout) (None, 64, 64, 32) 0
dw_sepconv_t_4 (DepthwiseC (None, 128, 128, 32) 320
onv2DTranspose)
pw_sepconv_t_4 (Conv2D) (None, 128, 128, 16) 528
pw_sepconv_t_4/BN (BatchNo (None, 128, 128, 16) 64
rmalization)
pw_sepconv_t_4/relu (ReLU) (None, 128, 128, 16) 0
dropout_4 (Dropout) (None, 128, 128, 16) 0
head (Conv2D) (None, 128, 128, 1) 17
=================================================================
Total params: 1058865 (4.04 MB)
Trainable params: 1051889 (4.01 MB)
Non-trainable params: 6976 (27.25 KB)
_________________________________________________________________
from tf_keras.metrics import BinaryIoU
# Compile the native TF-Keras model (required to evaluate the metrics)
model_keras.compile(loss='binary_crossentropy', metrics=[BinaryIoU(), 'accuracy'])
# Check Keras model performance
_, biou, acc = model_keras.evaluate(x_val, y_val, steps=steps, verbose=0)
print(f"TF-Keras binary IoU / pixel accuracy: {biou:.4f} / {100*acc:.2f}%")
TF-Keras binary IoU / pixel accuracy: 0.9356 / 96.78%
3. Load a pre-trained quantized TF-Keras model
The next step is to quantize and potentially perform Quantization Aware Training (QAT) on the TF-Keras model from the previous step. After the TF-Keras model is quantized to 8-bit for all weights and activations, QAT is used to maintain the performance of the quantized model. Again, a pre-trained model is downloaded to save runtime.
from akida_models import akida_unet_portrait128_pretrained
# Load the pre-trained quantized model
model_quantized_keras = akida_unet_portrait128_pretrained()
model_quantized_keras.summary()
Downloading data from https://data.brainchip.com/models/AkidaV2/akida_unet/akida_unet_portrait128_i8_w8_a8.h5.
0/4527816 [..............................] - ETA: 0s
253952/4527816 [>.............................] - ETA: 0s
1515520/4527816 [=========>....................] - ETA: 0s
2973696/4527816 [==================>...........] - ETA: 0s
4527816/4527816 [==============================] - 0s 0us/step
Download complete.
Model: "akida_unet"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 128, 128, 3)] 0
rescaling (QuantizedRescal (None, 128, 128, 3) 0
ing)
conv_0 (QuantizedConv2D) (None, 64, 64, 16) 448
conv_0/relu (QuantizedReLU (None, 64, 64, 16) 32
)
conv_1 (QuantizedConv2D) (None, 64, 64, 32) 4640
conv_1/relu (QuantizedReLU (None, 64, 64, 32) 64
)
conv_2 (QuantizedConv2D) (None, 32, 32, 64) 18496
conv_2/relu (QuantizedReLU (None, 32, 32, 64) 128
)
conv_3 (QuantizedConv2D) (None, 32, 32, 64) 36928
conv_3/relu (QuantizedReLU (None, 32, 32, 64) 128
)
dw_separable_4 (QuantizedD (None, 16, 16, 64) 704
epthwiseConv2D)
pw_separable_4 (QuantizedC (None, 16, 16, 128) 8320
onv2D)
pw_separable_4/relu (Quant (None, 16, 16, 128) 256
izedReLU)
dw_separable_5 (QuantizedD (None, 16, 16, 128) 1408
epthwiseConv2D)
pw_separable_5 (QuantizedC (None, 16, 16, 128) 16512
onv2D)
pw_separable_5/relu (Quant (None, 16, 16, 128) 256
izedReLU)
dw_separable_6 (QuantizedD (None, 8, 8, 128) 1408
epthwiseConv2D)
pw_separable_6 (QuantizedC (None, 8, 8, 256) 33024
onv2D)
pw_separable_6/relu (Quant (None, 8, 8, 256) 512
izedReLU)
dw_separable_7 (QuantizedD (None, 8, 8, 256) 2816
epthwiseConv2D)
pw_separable_7 (QuantizedC (None, 8, 8, 256) 65792
onv2D)
pw_separable_7/relu (Quant (None, 8, 8, 256) 512
izedReLU)
dw_separable_8 (QuantizedD (None, 8, 8, 256) 2816
epthwiseConv2D)
pw_separable_8 (QuantizedC (None, 8, 8, 256) 65792
onv2D)
pw_separable_8/relu (Quant (None, 8, 8, 256) 512
izedReLU)
dw_separable_9 (QuantizedD (None, 8, 8, 256) 2816
epthwiseConv2D)
pw_separable_9 (QuantizedC (None, 8, 8, 256) 65792
onv2D)
pw_separable_9/relu (Quant (None, 8, 8, 256) 512
izedReLU)
dw_separable_10 (Quantized (None, 8, 8, 256) 2816
DepthwiseConv2D)
pw_separable_10 (Quantized (None, 8, 8, 256) 65792
Conv2D)
pw_separable_10/relu (Quan (None, 8, 8, 256) 512
tizedReLU)
dw_separable_11 (Quantized (None, 8, 8, 256) 2816
DepthwiseConv2D)
pw_separable_11 (Quantized (None, 8, 8, 256) 65792
Conv2D)
pw_separable_11/relu (Quan (None, 8, 8, 256) 512
tizedReLU)
dw_separable_12 (Quantized (None, 4, 4, 256) 2816
DepthwiseConv2D)
pw_separable_12 (Quantized (None, 4, 4, 512) 131584
Conv2D)
pw_separable_12/relu (Quan (None, 4, 4, 512) 1024
tizedReLU)
dw_separable_13 (Quantized (None, 4, 4, 512) 5632
DepthwiseConv2D)
pw_separable_13 (Quantized (None, 4, 4, 512) 262656
Conv2D)
pw_separable_13/relu (Quan (None, 4, 4, 512) 1024
tizedReLU)
dw_sepconv_t_0 (QuantizedD (None, 8, 8, 512) 6144
epthwiseConv2DTranspose)
pw_sepconv_t_0 (QuantizedC (None, 8, 8, 256) 131328
onv2D)
pw_sepconv_t_0/relu (Quant (None, 8, 8, 256) 512
izedReLU)
dropout (QuantizedDropout) (None, 8, 8, 256) 0
dw_sepconv_t_1 (QuantizedD (None, 16, 16, 256) 3072
epthwiseConv2DTranspose)
pw_sepconv_t_1 (QuantizedC (None, 16, 16, 128) 32896
onv2D)
pw_sepconv_t_1/relu (Quant (None, 16, 16, 128) 256
izedReLU)
dropout_1 (QuantizedDropou (None, 16, 16, 128) 0
t)
dw_sepconv_t_2 (QuantizedD (None, 32, 32, 128) 1536
epthwiseConv2DTranspose)
pw_sepconv_t_2 (QuantizedC (None, 32, 32, 64) 8256
onv2D)
pw_sepconv_t_2/relu (Quant (None, 32, 32, 64) 128
izedReLU)
dropout_2 (QuantizedDropou (None, 32, 32, 64) 0
t)
dw_sepconv_t_3 (QuantizedD (None, 64, 64, 64) 768
epthwiseConv2DTranspose)
pw_sepconv_t_3 (QuantizedC (None, 64, 64, 32) 2080
onv2D)
pw_sepconv_t_3/relu (Quant (None, 64, 64, 32) 64
izedReLU)
dropout_3 (QuantizedDropou (None, 64, 64, 32) 0
t)
dw_sepconv_t_4 (QuantizedD (None, 128, 128, 32) 384
epthwiseConv2DTranspose)
pw_sepconv_t_4 (QuantizedC (None, 128, 128, 16) 528
onv2D)
pw_sepconv_t_4/relu (Quant (None, 128, 128, 16) 32
izedReLU)
dropout_4 (QuantizedDropou (None, 128, 128, 16) 0
t)
head (QuantizedConv2D) (None, 128, 128, 1) 19
dequantizer (Dequantizer) (None, 128, 128, 1) 0
=================================================================
Total params: 1061603 (4.05 MB)
Trainable params: 1047905 (4.00 MB)
Non-trainable params: 13698 (53.51 KB)
_________________________________________________________________
# Compile the quantized TF-Keras model (required to evaluate the metrics)
model_quantized_keras.compile(loss='binary_crossentropy', metrics=[BinaryIoU(), 'accuracy'])
# Check Keras model performance
_, biou, acc = model_quantized_keras.evaluate(x_val, y_val, steps=steps, verbose=0)
print(f"TF-Keras quantized binary IoU / pixel accuracy: {biou:.4f} / {100*acc:.2f}%")
TF-Keras quantized binary IoU / pixel accuracy: 0.9344 / 96.77%
4. Conversion to Akida
Finally, the quantized TF-Keras model from the previous step is converted into an Akida model and its performance is evaluated. Note that the original performance of the TF-Keras floating-point model is maintained throughout the conversion process in this example.
from cnn2snn import convert
# Convert the model
model_akida = convert(model_quantized_keras)
model_akida.summary()
Model Summary
_________________________________________________
Input shape Output shape Sequences Layers
=================================================
[128, 128, 3] [128, 128, 1] 1 36
_________________________________________________
_____________________________________________________________________________
Layer (type) Output shape Kernel shape
====================== SW/conv_0-dequantizer (Software) =====================
conv_0 (InputConv2D) [64, 64, 16] (3, 3, 3, 16)
_____________________________________________________________________________
conv_1 (Conv2D) [64, 64, 32] (3, 3, 16, 32)
_____________________________________________________________________________
conv_2 (Conv2D) [32, 32, 64] (3, 3, 32, 64)
_____________________________________________________________________________
conv_3 (Conv2D) [32, 32, 64] (3, 3, 64, 64)
_____________________________________________________________________________
dw_separable_4 (DepthwiseConv2D) [16, 16, 64] (3, 3, 64, 1)
_____________________________________________________________________________
pw_separable_4 (Conv2D) [16, 16, 128] (1, 1, 64, 128)
_____________________________________________________________________________
dw_separable_5 (DepthwiseConv2D) [16, 16, 128] (3, 3, 128, 1)
_____________________________________________________________________________
pw_separable_5 (Conv2D) [16, 16, 128] (1, 1, 128, 128)
_____________________________________________________________________________
dw_separable_6 (DepthwiseConv2D) [8, 8, 128] (3, 3, 128, 1)
_____________________________________________________________________________
pw_separable_6 (Conv2D) [8, 8, 256] (1, 1, 128, 256)
_____________________________________________________________________________
dw_separable_7 (DepthwiseConv2D) [8, 8, 256] (3, 3, 256, 1)
_____________________________________________________________________________
pw_separable_7 (Conv2D) [8, 8, 256] (1, 1, 256, 256)
_____________________________________________________________________________
dw_separable_8 (DepthwiseConv2D) [8, 8, 256] (3, 3, 256, 1)
_____________________________________________________________________________
pw_separable_8 (Conv2D) [8, 8, 256] (1, 1, 256, 256)
_____________________________________________________________________________
dw_separable_9 (DepthwiseConv2D) [8, 8, 256] (3, 3, 256, 1)
_____________________________________________________________________________
pw_separable_9 (Conv2D) [8, 8, 256] (1, 1, 256, 256)
_____________________________________________________________________________
dw_separable_10 (DepthwiseConv2D) [8, 8, 256] (3, 3, 256, 1)
_____________________________________________________________________________
pw_separable_10 (Conv2D) [8, 8, 256] (1, 1, 256, 256)
_____________________________________________________________________________
dw_separable_11 (DepthwiseConv2D) [8, 8, 256] (3, 3, 256, 1)
_____________________________________________________________________________
pw_separable_11 (Conv2D) [8, 8, 256] (1, 1, 256, 256)
_____________________________________________________________________________
dw_separable_12 (DepthwiseConv2D) [4, 4, 256] (3, 3, 256, 1)
_____________________________________________________________________________
pw_separable_12 (Conv2D) [4, 4, 512] (1, 1, 256, 512)
_____________________________________________________________________________
dw_separable_13 (DepthwiseConv2D) [4, 4, 512] (3, 3, 512, 1)
_____________________________________________________________________________
pw_separable_13 (Conv2D) [4, 4, 512] (1, 1, 512, 512)
_____________________________________________________________________________
dw_sepconv_t_0 (DepthwiseConv2DTranspose) [8, 8, 512] (3, 3, 512, 1)
_____________________________________________________________________________
pw_sepconv_t_0 (Conv2D) [8, 8, 256] (1, 1, 512, 256)
_____________________________________________________________________________
dw_sepconv_t_1 (DepthwiseConv2DTranspose) [16, 16, 256] (3, 3, 256, 1)
_____________________________________________________________________________
pw_sepconv_t_1 (Conv2D) [16, 16, 128] (1, 1, 256, 128)
_____________________________________________________________________________
dw_sepconv_t_2 (DepthwiseConv2DTranspose) [32, 32, 128] (3, 3, 128, 1)
_____________________________________________________________________________
pw_sepconv_t_2 (Conv2D) [32, 32, 64] (1, 1, 128, 64)
_____________________________________________________________________________
dw_sepconv_t_3 (DepthwiseConv2DTranspose) [64, 64, 64] (3, 3, 64, 1)
_____________________________________________________________________________
pw_sepconv_t_3 (Conv2D) [64, 64, 32] (1, 1, 64, 32)
_____________________________________________________________________________
dw_sepconv_t_4 (DepthwiseConv2DTranspose) [128, 128, 32] (3, 3, 32, 1)
_____________________________________________________________________________
pw_sepconv_t_4 (Conv2D) [128, 128, 16] (1, 1, 32, 16)
_____________________________________________________________________________
head (Conv2D) [128, 128, 1] (1, 1, 16, 1)
_____________________________________________________________________________
dequantizer (Dequantizer) [128, 128, 1] N/A
_____________________________________________________________________________
import tf_keras as keras
# Check Akida model performance
labels, pots = None, None
for s in range(steps):
batch = x_val[s * batch_size: (s + 1) * batch_size, :]
label_batch = y_val[s * batch_size: (s + 1) * batch_size, :]
pots_batch = model_akida.predict(batch.astype('uint8'))
if labels is None:
labels = label_batch
pots = pots_batch
else:
labels = np.concatenate((labels, label_batch))
pots = np.concatenate((pots, pots_batch))
preds = keras.activations.sigmoid(pots)
m_binary_iou = keras.metrics.BinaryIoU(target_class_ids=[0, 1], threshold=0.5)
m_binary_iou.update_state(labels, preds)
binary_iou = m_binary_iou.result().numpy()
m_accuracy = keras.metrics.Accuracy()
m_accuracy.update_state(labels, preds > 0.5)
accuracy = m_accuracy.result().numpy()
print(f"Akida binary IoU / pixel accuracy: {binary_iou:.4f} / {100*accuracy:.2f}%")
# For non-regression purposes
assert binary_iou > 0.9
Akida binary IoU / pixel accuracy: 0.9249 / 96.71%
5. Segment a single image
For visualization of the person segmentation performed by the Akida model, display a single image along with the segmentation produced by the original floating-point model and the ground truth segmentation.
import matplotlib.pyplot as plt
# Run inference on a random single image and display TF-Keras and Akida outputs
sample = np.expand_dims(x_val[id, :], 0)
keras_out = model_keras(sample)
akida_out = keras.activations.sigmoid(model_akida.forward(sample.astype('uint8')))
fig, axs = plt.subplots(1, 3, constrained_layout=True)
axs[0].imshow(keras_out[0] * sample[0] / 255.)
axs[0].set_title('Keras segmentation', fontsize=10)
axs[0].axis('off')
axs[1].imshow(akida_out[0] * sample[0] / 255.)
axs[1].set_title('Akida segmentation', fontsize=10)
axs[1].axis('off')
axs[2].imshow(y_val[id] * sample[0] / 255.)
axs[2].set_title('Expected segmentation', fontsize=10)
axs[2].axis('off')
plt.show()
Total running time of the script: (2 minutes 4.373 seconds)