├── LICENSE ├── README.md ├── conv_demo.py ├── densenet.py ├── googlenet.py ├── lenet.py ├── mobilenetv1.py ├── mobilenetv2.py ├── requirements.txt ├── resnet.py ├── senet.py ├── shufflenetv1.py ├── shufflenetv2.py ├── utils.py ├── vgg.py └── xception.py /LICENSE: -------------------------------------------------------------------------------- 1 | 2 | Apache License 3 | Version 2.0, January 2004 4 | http://www.apache.org/licenses/ 5 | 6 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 7 | 8 | 1. Definitions. 9 | 10 | "License" shall mean the terms and conditions for use, reproduction, 11 | and distribution as defined by Sections 1 through 9 of this document. 12 | 13 | "Licensor" shall mean the copyright owner or entity authorized by 14 | the copyright owner that is granting the License. 15 | 16 | "Legal Entity" shall mean the union of the acting entity and all 17 | other entities that control, are controlled by, or are under common 18 | control with that entity. For the purposes of this definition, 19 | "control" means (i) the power, direct or indirect, to cause the 20 | direction or management of such entity, whether by contract or 21 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 22 | outstanding shares, or (iii) beneficial ownership of such entity. 23 | 24 | "You" (or "Your") shall mean an individual or Legal Entity 25 | exercising permissions granted by this License. 26 | 27 | "Source" form shall mean the preferred form for making modifications, 28 | including but not limited to software source code, documentation 29 | source, and configuration files. 30 | 31 | "Object" form shall mean any form resulting from mechanical 32 | transformation or translation of a Source form, including but 33 | not limited to compiled object code, generated documentation, 34 | and conversions to other media types. 35 | 36 | "Work" shall mean the work of authorship, whether in Source or 37 | Object form, made available under the License, as indicated by a 38 | copyright notice that is included in or attached to the work 39 | (an example is provided in the Appendix below). 40 | 41 | "Derivative Works" shall mean any work, whether in Source or Object 42 | form, that is based on (or derived from) the Work and for which the 43 | editorial revisions, annotations, elaborations, or other modifications 44 | represent, as a whole, an original work of authorship. For the purposes 45 | of this License, Derivative Works shall not include works that remain 46 | separable from, or merely link (or bind by name) to the interfaces of, 47 | the Work and Derivative Works thereof. 48 | 49 | "Contribution" shall mean any work of authorship, including 50 | the original version of the Work and any modifications or additions 51 | to that Work or Derivative Works thereof, that is intentionally 52 | submitted to Licensor for inclusion in the Work by the copyright owner 53 | or by an individual or Legal Entity authorized to submit on behalf of 54 | the copyright owner. For the purposes of this definition, "submitted" 55 | means any form of electronic, verbal, or written communication sent 56 | to the Licensor or its representatives, including but not limited to 57 | communication on electronic mailing lists, source code control systems, 58 | and issue tracking systems that are managed by, or on behalf of, the 59 | Licensor for the purpose of discussing and improving the Work, but 60 | excluding communication that is conspicuously marked or otherwise 61 | designated in writing by the copyright owner as "Not a Contribution." 62 | 63 | "Contributor" shall mean Licensor and any individual or Legal Entity 64 | on behalf of whom a Contribution has been received by Licensor and 65 | subsequently incorporated within the Work. 66 | 67 | 2. Grant of Copyright License. Subject to the terms and conditions of 68 | this License, each Contributor hereby grants to You a perpetual, 69 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 70 | copyright license to reproduce, prepare Derivative Works of, 71 | publicly display, publicly perform, sublicense, and distribute the 72 | Work and such Derivative Works in Source or Object form. 73 | 74 | 3. Grant of Patent License. Subject to the terms and conditions of 75 | this License, each Contributor hereby grants to You a perpetual, 76 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 77 | (except as stated in this section) patent license to make, have made, 78 | use, offer to sell, sell, import, and otherwise transfer the Work, 79 | where such license applies only to those patent claims licensable 80 | by such Contributor that are necessarily infringed by their 81 | Contribution(s) alone or by combination of their Contribution(s) 82 | with the Work to which such Contribution(s) was submitted. If You 83 | institute patent litigation against any entity (including a 84 | cross-claim or counterclaim in a lawsuit) alleging that the Work 85 | or a Contribution incorporated within the Work constitutes direct 86 | or contributory patent infringement, then any patent licenses 87 | granted to You under this License for that Work shall terminate 88 | as of the date such litigation is filed. 89 | 90 | 4. Redistribution. You may reproduce and distribute copies of the 91 | Work or Derivative Works thereof in any medium, with or without 92 | modifications, and in Source or Object form, provided that You 93 | meet the following conditions: 94 | 95 | (a) You must give any other recipients of the Work or 96 | Derivative Works a copy of this License; and 97 | 98 | (b) You must cause any modified files to carry prominent notices 99 | stating that You changed the files; and 100 | 101 | (c) You must retain, in the Source form of any Derivative Works 102 | that You distribute, all copyright, patent, trademark, and 103 | attribution notices from the Source form of the Work, 104 | excluding those notices that do not pertain to any part of 105 | the Derivative Works; and 106 | 107 | (d) If the Work includes a "NOTICE" text file as part of its 108 | distribution, then any Derivative Works that You distribute must 109 | include a readable copy of the attribution notices contained 110 | within such NOTICE file, excluding those notices that do not 111 | pertain to any part of the Derivative Works, in at least one 112 | of the following places: within a NOTICE text file distributed 113 | as part of the Derivative Works; within the Source form or 114 | documentation, if provided along with the Derivative Works; or, 115 | within a display generated by the Derivative Works, if and 116 | wherever such third-party notices normally appear. The contents 117 | of the NOTICE file are for informational purposes only and 118 | do not modify the License. You may add Your own attribution 119 | notices within Derivative Works that You distribute, alongside 120 | or as an addendum to the NOTICE text from the Work, provided 121 | that such additional attribution notices cannot be construed 122 | as modifying the License. 123 | 124 | You may add Your own copyright statement to Your modifications and 125 | may provide additional or different license terms and conditions 126 | for use, reproduction, or distribution of Your modifications, or 127 | for any such Derivative Works as a whole, provided Your use, 128 | reproduction, and distribution of the Work otherwise complies with 129 | the conditions stated in this License. 130 | 131 | 5. Submission of Contributions. Unless You explicitly state otherwise, 132 | any Contribution intentionally submitted for inclusion in the Work 133 | by You to the Licensor shall be under the terms and conditions of 134 | this License, without any additional terms or conditions. 135 | Notwithstanding the above, nothing herein shall supersede or modify 136 | the terms of any separate license agreement you may have executed 137 | with Licensor regarding such Contributions. 138 | 139 | 6. Trademarks. This License does not grant permission to use the trade 140 | names, trademarks, service marks, or product names of the Licensor, 141 | except as required for reasonable and customary use in describing the 142 | origin of the Work and reproducing the content of the NOTICE file. 143 | 144 | 7. Disclaimer of Warranty. Unless required by applicable law or 145 | agreed to in writing, Licensor provides the Work (and each 146 | Contributor provides its Contributions) on an "AS IS" BASIS, 147 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 148 | implied, including, without limitation, any warranties or conditions 149 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 150 | PARTICULAR PURPOSE. You are solely responsible for determining the 151 | appropriateness of using or redistributing the Work and assume any 152 | risks associated with Your exercise of permissions under this License. 153 | 154 | 8. Limitation of Liability. In no event and under no legal theory, 155 | whether in tort (including negligence), contract, or otherwise, 156 | unless required by applicable law (such as deliberate and grossly 157 | negligent acts) or agreed to in writing, shall any Contributor be 158 | liable to You for damages, including any direct, indirect, special, 159 | incidental, or consequential damages of any character arising as a 160 | result of this License or out of the use or inability to use the 161 | Work (including but not limited to damages for loss of goodwill, 162 | work stoppage, computer failure or malfunction, or any and all 163 | other commercial damages or losses), even if such Contributor 164 | has been advised of the possibility of such damages. 165 | 166 | 9. Accepting Warranty or Additional Liability. While redistributing 167 | the Work or Derivative Works thereof, You may choose to offer, 168 | and charge a fee for, acceptance of support, warranty, indemnity, 169 | or other liability obligations and/or rights consistent with this 170 | License. However, in accepting such obligations, You may act only 171 | on Your own behalf and on Your sole responsibility, not on behalf 172 | of any other Contributor, and only if You agree to indemnify, 173 | defend, and hold each Contributor harmless for any liability 174 | incurred by, or claims asserted against, such Contributor by reason 175 | of your accepting any such warranty or additional liability. 176 | 177 | END OF TERMS AND CONDITIONS 178 | 179 | APPENDIX: How to apply the Apache License to your work. 180 | 181 | To apply the Apache License to your work, attach the following 182 | boilerplate notice, with the fields enclosed by brackets "[]" 183 | replaced with your own identifying information. (Don't include 184 | the brackets!) The text should be enclosed in the appropriate 185 | comment syntax for the file format. We also recommend that a 186 | file or class name and description of purpose be included on the 187 | same "printed page" as the copyright notice for easier 188 | identification within third-party archives. 189 | 190 | Copyright [2021] [Mofan Zhou] 191 | 192 | Licensed under the Apache License, Version 2.0 (the "License"); 193 | you may not use this file except in compliance with the License. 194 | You may obtain a copy of the License at 195 | 196 | http://www.apache.org/licenses/LICENSE-2.0 197 | 198 | Unless required by applicable law or agreed to in writing, software 199 | distributed under the License is distributed on an "AS IS" BASIS, 200 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 201 | See the License for the specific language governing permissions and 202 | limitations under the License. -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Tutorials of Computer Vision 2 | 3 | This repo includes some implementations of Computer Vision algorithms using tf2+. Codes are easy to read and follow. 4 | If you can read Chinese, I have a teaching website for studying AI models. 5 | 6 | All toy implementations are organised as following: 7 | 8 | - CNN 9 | - [Numpy Convolution mechanism](#ConvMechanism) 10 | - [LeNet](#LeNet) 11 | - [VGG](#VGG) 12 | - [GoogLeNet](#GoogLeNet) 13 | - [ResNet](#ResNet) 14 | - [DenseNet](#DenseNet) 15 | - [SENet](#SENet) 16 | - [MobileNetV1](#MobileNetV1) 17 | - [MobileNetV2](#MobileNetV2) 18 | - [Xception](#Xception) 19 | - [ShuffleNetV1](#ShuffleNetV1) 20 | - [ShuffleNetV2](#ShuffleNetV2) 21 | 22 | # Installation 23 | ```shell script 24 | $ git clone https://github.com/MorvanZhou/Computer-Vision 25 | $ cd Computer-Vision 26 | $ pip install -r requirements.txt 27 | ``` 28 | # ConvMechanism 29 | Convolution mechanism and feature map 30 | 31 | [code](/conv_demo.py) - [gif result](https://mofanpy.com/static/results/cv/conv_mechanism.gif) 32 | 33 | 34 | net structure 35 | 36 | 37 | # LeNet 38 | [Gradient-Based Learning Applied to Document Recognition](http://www.dengfanxin.cn/wp-content/uploads/2016/03/1998Lecun.pdf) 39 | 40 | [code](/lenet.py) - [net structure](https://mofanpy.com/static/results/cv/LeNet_structure.png) 41 | 42 | 43 | net structure 44 | 45 | 46 | # VGG 47 | [Very Deep Convolutional Networks for Large-Scale Image Recognition](https://arxiv.org/abs/1409.1556) 48 | 49 | Deep stacked CNN. 50 | 51 | [code](/vgg.py) - [net structure](https://mofanpy.com/static/results/cv/VGG_structure.png) 52 | 53 | 54 | net structure 55 | 56 | 57 | # GoogLeNet 58 | [Going Deeper with Convolutions](https://arxiv.org/abs/1409.4842) 59 | 60 | Multi kernel size to capture different local information 61 | 62 | [code](/googlenet.py) - [net structure](https://mofanpy.com/static/results/cv/GoogleLeNet_structure.png) 63 | 64 | 65 | net structure 66 | 67 | 68 | # ResNet 69 | [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) 70 | 71 | Add residual connection for better gradients. 72 | 73 | [code](/resnet.py) - [net structure](https://mofanpy.com/static/results/cv/ResNet_structure.png) 74 | 75 | 76 | net structure 77 | 78 | 79 | # DenseNet 80 | [Densely Connected Convolutional Networks](https://arxiv.org/abs/1608.06993) 81 | 82 | Compared with resnet, it has less filter each conv, sees more previous inputs. 83 | 84 | [code](/densenet.py) - [net structure](https://mofanpy.com/static/results/cv/DenseNet_structure.png) 85 | 86 | 87 | net structure 88 | 89 | 90 | # SENet 91 | [Squeeze-and-Excitation Networks](https://arxiv.org/abs/1709.01507) 92 | 93 | SE is a module that learns to scale each feature map, it can be plugged in many cnn block, 94 | larger reduction_ratio reduce parameter size in FC layers with limited accuracy drop. 95 | 96 | [code](/senet.py) - [net structure](https://mofanpy.com/static/results/cv/SENet_structure.png) 97 | 98 | 99 | net structure 100 | 101 | 102 | # MobileNetV1 103 | [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861) 104 | 105 | Decomposed classical conv to two operations (dw+pw). Small but effective cnn optimized on mobile (cpu). 106 | 107 | [code](/mobilenetv1.py) - [net structure](https://mofanpy.com/static/results/cv/MobileNetV1_structure.png) 108 | 109 | 110 | net structure 111 | 112 | 113 | # MobileNetV2 114 | [MobileNetV2: Inverted Residuals and Linear Bottlenecks](https://arxiv.org/abs/1801.04381) 115 | 116 | MobileNet v2 is v1 with residual block and layer rearrange (residual+pw+dw+pw): 117 | 118 | - mobilenet v1: dw > pw 119 | - mobilenet v2: pw > dw > pw let dw see more feature maps 120 | 121 | [code](/mobilenetv2.py) - [net structure](https://mofanpy.com/static/results/cv/MobileNetV2_structure.png) 122 | 123 | 124 | net structure 125 | 126 | 127 | # Xception 128 | [Xception: Deep Learning with Depthwise Separable Convolutions](https://arxiv.org/abs/1610.02357) 129 | 130 | Just like MobileNetV2 without last pw (residual+pw+dw). 131 | 132 | [code](/xception.py) - [net structure](https://mofanpy.com/static/results/cv/Xception_structure.png) 133 | 134 | 135 | net structure 136 | 137 | 138 | 139 | # ShuffleNetV1 140 | [ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices](https://arxiv.org/abs/1707.01083) 141 | 142 | Shuffle the output from 1x1 conv, and do group conv to reduce connections and speed up computing. 143 | But MobileNet is better in this case, this may caused by group conv cuts off some feature map communications. 144 | 145 | [code](/shufflenetv1.py) - [net structure](https://mofanpy.com/static/results/cv/ShuffleNetV1_structure.png) 146 | 147 | 148 | net structure 149 | 150 | 151 | # ShuffleNetV2 152 | [ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design](https://arxiv.org/abs/1807.11164) 153 | 154 | Further reduces parameters by switching group conv with split+concat, perform shuffle at end of block. Speed up calculation. 155 | But MobileNet is better in this case, this may caused by group conv cuts off some feature map communications. 156 | 157 | [code](/shufflenetv2.py) - [net structure](https://mofanpy.com/static/results/cv/ShuffleNetV2_structure.png) 158 | 159 | 160 | net structure 161 | -------------------------------------------------------------------------------- /conv_demo.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import matplotlib.pyplot as plt 3 | import os 4 | from PIL import Image 5 | from utils import load_mnist 6 | 7 | 8 | def save_gif(imgs_dir): 9 | imgs = [] 10 | imgs_path = [os.path.join(imgs_dir, p) for p in os.listdir(imgs_dir) if p.endswith(".png")] 11 | for f in sorted(imgs_path, key=os.path.getmtime): 12 | if not f.endswith(".png"): 13 | continue 14 | img = Image.open(f) 15 | img = img.resize((img.width, img.height), Image.ANTIALIAS) 16 | imgs.append(img) 17 | path = "{}/conv_mechanism.gif".format(os.path.dirname(imgs_dir)) 18 | if os.path.exists(path): 19 | os.remove(path) 20 | imgs[0].save(path, append_images=imgs[1:], optimize=False, save_all=True, duration=20, loop=0) 21 | print("saved ", path) 22 | 23 | 24 | def show_conv(save_dir, image): 25 | filter = np.array([ 26 | [1, 1, 1], 27 | [0, 0, 0], 28 | [-1, -1, -1]]) 29 | plt.figure(0, figsize=(9, 5)) 30 | ax1 = plt.subplot(121) 31 | ax1.imshow(image, cmap='gray_r') 32 | plt.xticks(()) 33 | plt.yticks(()) 34 | ax2 = plt.subplot(122) 35 | texts = [] 36 | feature_map = np.zeros((26, 26)) 37 | for i in range(26): 38 | for j in range(26): 39 | 40 | if texts: 41 | fm.remove() 42 | for n in range(3): 43 | for m in range(3): 44 | if len(texts) != 9: 45 | texts.append(ax1.text(j+m, i+n, filter[n, m], color='k', size=8, ha='center', va='center')) 46 | else: 47 | texts[n*3+m].set_position((j+m, i+n)) 48 | 49 | feature_map[i, j] = np.sum(filter * image[i:i+3, j:j+3]) 50 | fm = ax2.imshow(feature_map, cmap='gray', vmax=3, vmin=-3) 51 | plt.xticks(()) 52 | plt.yticks(()) 53 | plt.tight_layout() 54 | plt.savefig(os.path.join(save_dir, "i{}j{}.png".format(i, j))) 55 | 56 | 57 | def show_feature_map(save_dir, image): 58 | filters = [ 59 | np.array([ 60 | [1, 1, 1], 61 | [0, 0, 0], 62 | [-1, -1, -1]]), 63 | np.array([ 64 | [-1, -1, -1], 65 | [0, 0, 0], 66 | [1, 1, 1]]), 67 | np.array([ 68 | [1, 0, -1], 69 | [1, 0, -1], 70 | [1, 0, -1]]), 71 | np.array([ 72 | [-1, 0, 1], 73 | [-1, 0, 1], 74 | [-1, 0, 1]]) 75 | ] 76 | 77 | plt.figure(1) 78 | plt.title('Original image') 79 | plt.imshow(image, cmap='gray_r') 80 | plt.xticks(()) 81 | plt.yticks(()) 82 | plt.savefig(os.path.join(save_dir, "original_img.png")) 83 | 84 | plt.figure(2) 85 | for n in range(4): 86 | feature_map = np.zeros((26, 26)) 87 | 88 | for i in range(26): 89 | for j in range(26): 90 | feature_map[i, j] = np.sum(image[i:i + 3, j:j + 3] * filters[n]) 91 | 92 | plt.subplot(3, 4, 1 + n) 93 | plt.title('Filter%i' % n) 94 | plt.imshow(filters[n], cmap='gray', vmax=3, vmin=-3) 95 | plt.xticks(()) 96 | plt.yticks(()) 97 | 98 | plt.subplot(3, 4, 5 + n) 99 | plt.title('Conv%i' % n) 100 | plt.imshow(feature_map, cmap='gray', vmax=3, vmin=-3) 101 | plt.xticks(()) 102 | plt.yticks(()) 103 | 104 | plt.subplot(3, 4, 9 + n) 105 | plt.title('ReLU%i' % n) 106 | feature_map = np.maximum(0, feature_map) 107 | plt.imshow(feature_map, cmap='gray', vmax=3, vmin=-3) 108 | plt.xticks(()) 109 | plt.yticks(()) 110 | 111 | plt.tight_layout() 112 | plt.savefig(os.path.join(save_dir, "feature_map.png")) 113 | 114 | 115 | if __name__ == "__main__": 116 | result_dir = "visual/basic" 117 | conv_dir = os.path.join(result_dir, "conv") 118 | os.makedirs(conv_dir, exist_ok=True) 119 | (x_train, y_train), (x_test, y_test) = load_mnist() 120 | 121 | data = x_train[7].squeeze(axis=-1) 122 | show_feature_map(result_dir, data) 123 | show_conv(conv_dir, data) 124 | save_gif(conv_dir) 125 | -------------------------------------------------------------------------------- /densenet.py: -------------------------------------------------------------------------------- 1 | # [Densely Connected Convolutional Networks](https://arxiv.org/abs/1608.06993) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | 12 | # compared with resnet, it has less filter each conv, sees more previous inputs 13 | def bottleneck(xs, growth_rate): 14 | if len(xs) == 1: 15 | o = xs[0] 16 | else: 17 | o = layers.concatenate(xs, axis=-1) # [n, h, w, c * xs] 18 | o = layers.ReLU()(o) 19 | o = layers.Conv2D(growth_rate, kernel_size=1, strides=1)(o) # [n, h, w, c] 20 | o = layers.ReLU()(o) 21 | o = layers.Conv2D(growth_rate, kernel_size=3, strides=1, padding="same")(o) # [n, h, w, c] 22 | return o 23 | 24 | 25 | def block(x, growth_rate, n_bottleneck=2): 26 | outs = [bottleneck([x], growth_rate)] # [n, h, w, c] 27 | for i in range(1, n_bottleneck): 28 | o = bottleneck(outs, growth_rate) # [n, h, w, c] 29 | outs.append(o) 30 | return layers.concatenate(outs, axis=-1) # [n, h, w, c * n_bottleneck] 31 | 32 | 33 | def build_model(): 34 | inputs = layers.Input(shape=(28, 28, 1), name="img") 35 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 36 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 37 | x = block(x, growth_rate=8, n_bottleneck=2) # [n, 14, 14, 8*2] 38 | x = layers.MaxPool2D(2)(x) # [n, 7, 7, 8*2] 39 | x = block(x, 8, 3) # [n, 7, 7, 8*3] 40 | x = layers.GlobalAveragePooling2D()(x) # [n, 24] 41 | o = layers.Dense(10)(x) # [n, 10] 42 | return keras.Model(inputs, o, name="DenseNet") 43 | 44 | # show model 45 | model = build_model() 46 | model.summary() 47 | save_model_structure(model) 48 | 49 | # define loss and optimizer 50 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 51 | opt = keras.optimizers.Adam(0.001) 52 | accuracy = keras.metrics.SparseCategoricalAccuracy() 53 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 54 | 55 | # training and validation 56 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 57 | 58 | # save model 59 | save_model_weights(model) 60 | -------------------------------------------------------------------------------- /googlenet.py: -------------------------------------------------------------------------------- 1 | # [Going Deeper with Convolutions](https://arxiv.org/abs/1409.4842) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers, activations 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | 12 | # multi kernel size to capture different local information 13 | def inception(x, f1, f2, f3, f4, name): 14 | act = activations.relu 15 | inputs = layers.Input(shape=x.shape[1:]) 16 | 17 | p1 = layers.Conv2D(filters=f1, kernel_size=1, strides=1, activation=act, name="p1")(inputs) 18 | 19 | p2 = layers.Conv2D(f2[0], 1, 1, name="p21")(inputs) 20 | p2 = layers.Conv2D(f2[1], 3, 1, padding="same", activation=act, name="p22")(p2) 21 | 22 | p3 = layers.Conv2D(f3[0], 1, 1, name="p31")(inputs) 23 | p3 = layers.Conv2D(f3[1], 5, 1, padding="same", activation=act, name="p32")(p3) 24 | 25 | p4 = layers.MaxPool2D(pool_size=3, strides=1, padding="same", name="p41")(inputs) 26 | p4 = layers.Conv2D(f4, 1, activation=act, name="p42")(p4) 27 | 28 | p = layers.concatenate((p1, p2, p3, p4), axis=-1) 29 | m = keras.Model(inputs, p, name=name) 30 | return m(x) 31 | 32 | 33 | def build_model(): 34 | inputs = layers.Input(shape=(28, 28, 1), name="img") 35 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 36 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 37 | x = inception(x, f1=4, f2=(2, 4), f3=(2, 4), f4=4, name="inspection1") # [n, 14, 14, 4*4] 38 | x = layers.MaxPool2D(2, 2)(x) # [n, 7, 7, 4*4] 39 | x = inception(x, f1=8, f2=(4, 8), f3=(4, 8), f4=8, name="inspection2") # [n, 7, 7, 8*4] 40 | x = layers.GlobalAveragePooling2D()(x) # [n, 8*4] 41 | o = layers.Dense(10)(x) # [n, 10] 42 | return keras.Model(inputs, o, name="GoogLeNet") 43 | 44 | 45 | # show model 46 | model = build_model() 47 | model.summary() 48 | save_model_structure(model) 49 | 50 | # define loss and optimizer 51 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 52 | opt = keras.optimizers.Adam(0.001) 53 | accuracy = keras.metrics.SparseCategoricalAccuracy() 54 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 55 | 56 | # training and validation 57 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 58 | 59 | # save model 60 | save_model_weights(model) 61 | 62 | 63 | -------------------------------------------------------------------------------- /lenet.py: -------------------------------------------------------------------------------- 1 | # [Gradient-Based Learning Applied to Document Recognition](http://www.dengfanxin.cn/wp-content/uploads/2016/03/1998Lecun.pdf) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | # define model 12 | model = keras.Sequential([ 13 | layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same", input_shape=(28, 28, 1)), # [n, 28, 28, 8] 14 | layers.MaxPool2D(pool_size=2, strides=2), # [n, 14, 14, 8] 15 | layers.Conv2D(16, 3, 1, "same"), # [n, 14, 14, 16] 16 | layers.MaxPool2D(2, 2), # [n, 7, 7, 16] 17 | layers.Flatten(), # [n, 7*7*16] 18 | layers.Dense(10) # [n, 10] 19 | ], name="LeNet") 20 | 21 | # show model 22 | model.summary() 23 | save_model_structure(model) 24 | 25 | # define loss and optimizer 26 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 27 | opt = keras.optimizers.Adam(0.001) 28 | accuracy = keras.metrics.SparseCategoricalAccuracy() 29 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 30 | 31 | # training and validation 32 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 33 | 34 | # save model 35 | save_model_weights(model) 36 | 37 | 38 | -------------------------------------------------------------------------------- /mobilenetv1.py: -------------------------------------------------------------------------------- 1 | # [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | 12 | # (dw+pw). small but effective cnn optimized on mobile (cpu) 13 | def block(x, filters): 14 | # Depthwise Separable convolutions 15 | o = layers.DepthwiseConv2D(kernel_size=3, strides=1, padding="same")(x) # [n, h, w, c] dw 16 | o = layers.ReLU()(o) 17 | o = layers.Conv2D(filters, 1, 1)(o) # [n, h, w, f] pw 18 | o = layers.ReLU()(o) 19 | return o 20 | 21 | 22 | def build_model(): 23 | inputs = layers.Input(shape=(28, 28, 1), name="img") 24 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 25 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 26 | x = block(x, filters=20) # [n, 14, 14, 20] 27 | x = layers.MaxPool2D(2, 2)(x) # [n, 7, 7, 20] 28 | x = block(x, 40) # [n, 7, 7, 40] 29 | x = layers.GlobalAveragePooling2D()(x) # [n, 40] 30 | o = layers.Dense(10)(x) # [n, 10] 31 | return keras.Model(inputs, o, name="MobileNetV1") 32 | 33 | 34 | # show model 35 | model = build_model() 36 | model.summary() 37 | save_model_structure(model) 38 | 39 | # define loss and optimizer 40 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 41 | opt = keras.optimizers.Adam(0.001) 42 | accuracy = keras.metrics.SparseCategoricalAccuracy() 43 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 44 | 45 | # training and validation 46 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 47 | 48 | # save model 49 | save_model_weights(model) 50 | -------------------------------------------------------------------------------- /mobilenetv2.py: -------------------------------------------------------------------------------- 1 | # [MobileNetV2: Inverted Residuals and Linear Bottlenecks](https://arxiv.org/abs/1801.04381) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | 12 | # (residual+pw+dw+pw). mobilenet v1 with residual block and layer rearrange 13 | def block(x, filters, expand_ratio=4): 14 | # mobilenet v1: dw > pw 15 | # mobilenet v2: pw > dw > pw let dw see more feature maps 16 | o = layers.Conv2D(int(filters*expand_ratio), 1, 1)(x) # [n, h, w, c*e] pw expansion 17 | o = layers.ReLU()(o) 18 | o = layers.DepthwiseConv2D(kernel_size=3, strides=1, padding="same")(o) # [n, h/s, w/s, c*e] dw 19 | o = layers.ReLU()(o) 20 | o = layers.Conv2D(filters, 1, 1)(o) # [n, h, w, c] pw 21 | if x.shape[-1] == filters: 22 | o = layers.add((o, x)) # residual connection 23 | return o 24 | 25 | 26 | def build_model(): 27 | inputs = layers.Input(shape=(28, 28, 1), name="img") 28 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 29 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 30 | x = block(x, filters=5) # [n, 14, 14, 5] 31 | x = layers.MaxPool2D(2, 2)(x) # [n, 7, 7, 5] 32 | x = block(x, 10) # [n, 7, 7, 10] 33 | x = layers.GlobalAveragePooling2D()(x) # [n, 10] 34 | o = layers.Dense(10)(x) # [n, 10] 35 | return keras.Model(inputs, o, name="MobileNetV2") 36 | 37 | 38 | # show model 39 | model = build_model() 40 | model.summary() 41 | save_model_structure(model) 42 | 43 | # define loss and optimizer 44 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 45 | opt = keras.optimizers.Adam(0.001) 46 | accuracy = keras.metrics.SparseCategoricalAccuracy() 47 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 48 | 49 | # training and validation 50 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 51 | 52 | # save model 53 | save_model_weights(model) 54 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | matplotlib==3.3.3 2 | numpy==1.18.5 3 | Pillow==8.0.1 4 | tensorflow==2.3.0 5 | pydot==1.4.1 6 | -------------------------------------------------------------------------------- /resnet.py: -------------------------------------------------------------------------------- 1 | # [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | 12 | # add residual connection for better gradients 13 | def bottleneck(x, filters, strides=1): 14 | c = x.shape[-1] 15 | if c != filters: 16 | shortcut = layers.Conv2D(filters, 1, strides, "same")(x) 17 | else: 18 | shortcut = x 19 | o = layers.ReLU()(x) 20 | o = layers.Conv2D(c, kernel_size=3, strides=1, padding="same")(o) # [n, h, w, c] 21 | o = layers.ReLU()(o) 22 | o = layers.Conv2D(filters, 3, strides, "same")(o) # [n, h/s, w/s, f] 23 | o = layers.add((o, shortcut)) 24 | return o 25 | 26 | 27 | def block(x, filters, strides=1, n_bottleneck=2): 28 | o = bottleneck(x, filters, strides) # [n, h/s, w/s, f] 29 | for i in range(1, n_bottleneck): 30 | o = bottleneck(o, filters, 1) # [n, h/s, w/s, f] 31 | return o 32 | 33 | 34 | def build_model(): 35 | inputs = layers.Input(shape=(28, 28, 1), name="img") 36 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 37 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 38 | x = block(x, filters=8, strides=1, n_bottleneck=2) # [n, 14, 14, 8] 39 | x = block(x, 16, 2, 1) # [n, 7, 7, 16] 40 | x = layers.GlobalAveragePooling2D()(x) # [n, 16] 41 | o = layers.Dense(10)(x) # [n, 10] 42 | return keras.Model(inputs, o, name="ResNet") 43 | 44 | # show model 45 | model = build_model() 46 | model.summary() 47 | save_model_structure(model) 48 | 49 | # define loss and optimizer 50 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 51 | opt = keras.optimizers.Adam(0.001) 52 | accuracy = keras.metrics.SparseCategoricalAccuracy() 53 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 54 | 55 | # training and validation 56 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 57 | 58 | # save model 59 | save_model_weights(model) 60 | -------------------------------------------------------------------------------- /senet.py: -------------------------------------------------------------------------------- 1 | # [Squeeze-and-Excitation Networks](https://arxiv.org/abs/1709.01507) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | 12 | # SE is a module that learns to scale each feature map, we can plug in a resnet block 13 | # larger reduction_ratio reduce parameter size in FC layers with limited accuracy drop 14 | def se(x, reduction_ratio=4): 15 | c = x.shape[-1] 16 | s = layers.GlobalAveragePooling2D()(x) 17 | s = layers.Dense(c // reduction_ratio)(s) 18 | s = layers.ReLU()(s) 19 | s = layers.Dense(c, activation=keras.activations.sigmoid)(s) 20 | return x * layers.Reshape((1, 1, c))(s) 21 | 22 | 23 | def bottleneck(x, filters, strides=1, reduction_ratio=4): 24 | in_channels = x.shape[-1] 25 | if in_channels != filters: 26 | shortcut = layers.Conv2D(filters, 1, strides, "same", name="projection")(x) 27 | else: 28 | shortcut = x 29 | o = layers.ReLU()(x) 30 | o = layers.Conv2D(in_channels, kernel_size=3, strides=1, padding="same")(o) # [n, h, w, c] 31 | o = layers.ReLU()(o) 32 | o = layers.Conv2D(filters, 3, strides, "same")(o) # [n, h/s, w/s, f] 33 | o = se(o, reduction_ratio) 34 | o = layers.add((o, shortcut)) 35 | return o 36 | 37 | 38 | def block(x, filters, strides=1, n_bottleneck=2): 39 | o = bottleneck(x, filters, strides) # [n, h/s, w/s, f] 40 | for i in range(1, n_bottleneck): 41 | o = bottleneck(o, filters, 1) # [n, h/s, w/s, f] 42 | return o 43 | 44 | 45 | def build_model(): 46 | inputs = layers.Input(shape=(28, 28, 1), name="img") 47 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 48 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 49 | x = block(x, filters=8, strides=1, n_bottleneck=2) # [n, 14, 14, 8] 50 | x = block(x, 16, 2, 1) # [n, 7, 7, 16] 51 | x = layers.GlobalAveragePooling2D()(x) # [n, 16] 52 | o = layers.Dense(10)(x) # [n, 10] 53 | return keras.Model(inputs, o, name="SENet") 54 | 55 | # show model 56 | model = build_model() 57 | model.summary() 58 | save_model_structure(model) 59 | 60 | # define loss and optimizer 61 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 62 | opt = keras.optimizers.Adam(0.001) 63 | accuracy = keras.metrics.SparseCategoricalAccuracy() 64 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 65 | 66 | # training and validation 67 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 68 | 69 | # save model 70 | save_model_weights(model) 71 | -------------------------------------------------------------------------------- /shufflenetv1.py: -------------------------------------------------------------------------------- 1 | # [ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices](https://arxiv.org/abs/1707.01083) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | import tensorflow as tf 7 | from utils import load_mnist, save_model_structure, save_model_weights 8 | 9 | # get data 10 | (x_train, y_train), (x_test, y_test) = load_mnist() 11 | 12 | 13 | # Grouped 1x1 conv to reduce parameters, but shuffle grouped feature maps to mix isolated feature map. 14 | def shuffle(x, groups): 15 | _, h, w, c = x.shape 16 | gc = c // groups 17 | o = layers.Reshape((h, w, groups, gc))(x) # [n, h, w, g, gc] 18 | o = layers.Permute((1, 2, 4, 3))(o) # shuffle in groups 19 | o = layers.Reshape((h, w, c))(o) 20 | return o 21 | 22 | 23 | def group_conv(x, filters, groups): 24 | assert x.shape[-1] % groups == 0 25 | assert filters % groups == 0 26 | if tf.test.is_built_with_gpu_support(): 27 | o = layers.Conv2D(filters, 1, 1, groups=groups)(x) # [n, h, w, f] pw, groups=groups not works on cpu (tf=2.3.0) 28 | else: 29 | o = layers.concatenate([ 30 | layers.Conv2D(filters // groups, 1, 1)(g) for g in tf.split(x, groups, axis=-1) 31 | ], axis=-1) # this works on cpu [n, h, w, f] 32 | return o 33 | 34 | 35 | def block(x, filters, groups=4): 36 | o = group_conv(x, x.shape[-1], groups) # [n, h, w, c] gpw 37 | o = shuffle(o, groups) 38 | o = layers.ReLU()(o) 39 | o = layers.DepthwiseConv2D(kernel_size=3, strides=1, padding="same")(o) # [n, h, w, c] dw 40 | o = group_conv(o, filters, groups) # [n, h, w, f] gpw 41 | if x.shape[-1] != filters: 42 | x = group_conv(x, filters, groups) # [n, h, w, f] 43 | o = layers.add((o, x)) # residual connection 44 | return o 45 | 46 | 47 | def build_model(): 48 | inputs = layers.Input(shape=(28, 28, 1), name="img") 49 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 50 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 51 | x = block(x, filters=32, groups=4) # [n, 14, 14, 32] 52 | x = layers.MaxPool2D(2, 2)(x) # [n, 7, 7, 32] 53 | x = block(x, 64, 4) # [n, 7, 7, 64] 54 | x = layers.GlobalAveragePooling2D()(x) # [n, 64] 55 | o = layers.Dense(10)(x) # [n, 10] 56 | return keras.Model(inputs, o, name="ShuffleNetV1") 57 | 58 | 59 | # show model 60 | model = build_model() 61 | model.summary() 62 | save_model_structure(model) 63 | 64 | # define loss and optimizer 65 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 66 | opt = keras.optimizers.Adam(0.001) 67 | accuracy = keras.metrics.SparseCategoricalAccuracy() 68 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 69 | 70 | # training and validation 71 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 72 | 73 | # save model 74 | save_model_weights(model) 75 | -------------------------------------------------------------------------------- /shufflenetv2.py: -------------------------------------------------------------------------------- 1 | # [ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design](https://arxiv.org/abs/1807.11164) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | import tensorflow as tf 7 | from utils import load_mnist, save_model_structure, save_model_weights 8 | 9 | # get data 10 | (x_train, y_train), (x_test, y_test) = load_mnist() 11 | 12 | 13 | # Further reduces parameters by switching group conv with split+concat, perform shuffle at end of block. 14 | # Speed up calculation. 15 | def shuffle(x, groups): 16 | _, h, w, c = x.shape 17 | gc = c // groups 18 | o = layers.Reshape((h, w, groups, gc))(x) # [n, h, w, g, gc] 19 | o = layers.Permute((1, 2, 4, 3))(o) # shuffle in groups 20 | o = layers.Reshape((h, w, c))(o) 21 | return o 22 | 23 | 24 | def block(x, filters): 25 | x1, x2 = tf.split(x, 2, axis=-1) # x1 [n, h, w, c/2], x2 [n, h, w, c/2] 26 | c = x.shape[-1] 27 | o = layers.Conv2D(x2.shape[-1], 1, 1)(x2) # [n, h, w, c/2] 28 | o = layers.ReLU()(o) 29 | o = layers.DepthwiseConv2D(kernel_size=3, strides=1, padding="same")(o) # [n, h, w, c/2] 30 | o = layers.ReLU()(o) 31 | o = layers.Conv2D(filters//2, 1, 1)(o) # [n, h, w, f/2] 32 | if filters != c: 33 | x1 = layers.Conv2D(filters//2, 1, 1)(x1) # [n, h, w, f/2] 34 | o = layers.concatenate([x1, o], axis=-1) # [n, h, w, f] 35 | o = shuffle(o, 2) # 36 | return o 37 | 38 | 39 | def build_model(): 40 | inputs = layers.Input(shape=(28, 28, 1), name="img") 41 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 42 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 43 | x = block(x, filters=32) # [n, 14, 14, 32] 44 | x = layers.MaxPool2D(2, 2)(x) # [n, 7, 7, 32] 45 | x = block(x, 64) # [n, 7, 7, 64] 46 | x = layers.GlobalAveragePooling2D()(x) # [n, 64] 47 | o = layers.Dense(10)(x) # [n, 10] 48 | return keras.Model(inputs, o, name="ShuffleNetV2") 49 | 50 | 51 | # show model 52 | model = build_model() 53 | model.summary() 54 | save_model_structure(model) 55 | 56 | # define loss and optimizer 57 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 58 | opt = keras.optimizers.Adam(0.001) 59 | accuracy = keras.metrics.SparseCategoricalAccuracy() 60 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 61 | 62 | # training and validation 63 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 64 | 65 | # save model 66 | save_model_weights(model) 67 | -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import os 3 | from urllib.request import urlretrieve 4 | from tensorflow import keras 5 | 6 | MNIST_PATH = "./mnist.npz" 7 | 8 | 9 | def load_mnist(path="./mnist.npz", norm=True): 10 | if not os.path.isfile(path): 11 | print("not mnist data is found, try downloading...") 12 | urlretrieve("https://s3.amazonaws.com/img-datasets/mnist.npz", path) 13 | with np.load(path, allow_pickle=True) as f: 14 | x_train = f['x_train'].astype(np.float32)[:, :, :, None] 15 | y_train = f['y_train'].astype(np.float32)[:, None] 16 | x_test = f['x_test'].astype(np.float32)[:, :, :, None] 17 | y_test = f['y_test'].astype(np.float32)[:, None] 18 | if norm: 19 | x_train /= 255 20 | x_test /= 255 21 | return (x_train, y_train), (x_test, y_test) 22 | 23 | 24 | def save_model_structure(model: keras.Model, path=None): 25 | if path is None: 26 | path = "visual/{}/{}_structure.png".format(model.name, model.name) 27 | os.makedirs(os.path.dirname(path), exist_ok=True) 28 | try: 29 | keras.utils.plot_model(model, show_shapes=True, expand_nested=True, dpi=150, to_file=path) 30 | except Exception as e: 31 | print(e) 32 | 33 | 34 | def save_model_weights(model: keras.Model, path=None): 35 | if path is None: 36 | path = "visual/{}/model/net".format(model.name) 37 | os.makedirs(os.path.dirname(path), exist_ok=True) 38 | model.save_weights(path) 39 | 40 | 41 | def load_model_weights(model: keras.Model, path=None): 42 | if path is None: 43 | path = "visual/{}/model/net".format(model.name) 44 | model.load_weights(path) -------------------------------------------------------------------------------- /vgg.py: -------------------------------------------------------------------------------- 1 | # [Very Deep Convolutional Networks for Large-Scale Image Recognition](https://arxiv.org/abs/1409.1556) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | # define model 12 | # like LeNet with more layers and activations 13 | model = keras.Sequential([ 14 | layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same", input_shape=(28, 28, 1)), # [n, 28, 28, 8] 15 | layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same", input_shape=(28, 28, 1)), # [n, 28, 28, 8] 16 | layers.ReLU(), 17 | layers.MaxPool2D(pool_size=2, strides=2), # [n, 14, 14, 8] 18 | layers.Conv2D(16, 3, 1, "same"), # [n, 14, 14, 16] 19 | layers.Conv2D(16, 3, 1, "same"), # [n, 14, 14, 16] 20 | layers.ReLU(), 21 | layers.MaxPool2D(2, 2), # [n, 7, 7, 16] 22 | layers.Flatten(), # [n, 7*7*16] 23 | layers.Dense(32), # [n, 32] 24 | layers.ReLU(), 25 | layers.Dense(10) # [n, 32] 26 | ], name="VGG") 27 | 28 | # show model 29 | model.summary() 30 | save_model_structure(model) 31 | 32 | # define loss and optimizer 33 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 34 | opt = keras.optimizers.Adam(0.001) 35 | accuracy = keras.metrics.SparseCategoricalAccuracy() 36 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 37 | 38 | # training and validation 39 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 40 | 41 | # save model 42 | save_model_weights(model) 43 | -------------------------------------------------------------------------------- /xception.py: -------------------------------------------------------------------------------- 1 | # [Xception: Deep Learning with Depthwise Separable Convolutions](https://arxiv.org/abs/1610.02357) 2 | # dependency file: https://github.com/MorvanZhou/Computer-Vision/requirements.txt 3 | 4 | from tensorflow import keras 5 | from tensorflow.keras import layers 6 | from utils import load_mnist, save_model_structure, save_model_weights 7 | 8 | # get data 9 | (x_train, y_train), (x_test, y_test) = load_mnist() 10 | 11 | 12 | # (residual+pw+dw) just like mobilenetv2 without last pw 13 | def block(x, filters): 14 | if x.shape[-1] != filters: 15 | shortcut = layers.Conv2D(filters, 1, 1)(x) 16 | else: 17 | shortcut = x 18 | o = layers.ReLU()(x) 19 | o = layers.Conv2D(filters, 1, 1)(o) # [n, h, w, f] pw 20 | o = layers.ReLU()(o) 21 | o = layers.DepthwiseConv2D(3, 1, "same")(o) # [n, h, w, f] dw 22 | o = layers.add((o, shortcut)) # residual connection 23 | return o 24 | 25 | 26 | def build_model(): 27 | inputs = layers.Input(shape=(28, 28, 1), name="img") 28 | x = layers.Conv2D(filters=8, kernel_size=3, strides=1, padding="same")(inputs) # [n, 28, 28, 8] 29 | x = layers.MaxPool2D(pool_size=2, strides=2)(x) # [n, 14, 14, 8] 30 | x = block(x, filters=16) # [n, 14, 14, 16] 31 | x = layers.MaxPool2D(2, 2)(x) # [n, 7, 7, 16] 32 | x = block(x, 32) # [n, 7, 7, 32] 33 | x = layers.GlobalAveragePooling2D()(x) # [n, 32] 34 | o = layers.Dense(10)(x) # [n, 10] 35 | return keras.Model(inputs, o, name="Xception") 36 | 37 | 38 | # show model 39 | model = build_model() 40 | model.summary() 41 | save_model_structure(model) 42 | 43 | # define loss and optimizer 44 | loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True) 45 | opt = keras.optimizers.Adam(0.001) 46 | accuracy = keras.metrics.SparseCategoricalAccuracy() 47 | model.compile(optimizer=opt, loss=loss, metrics=[accuracy]) 48 | 49 | # training and validation 50 | model.fit(x=x_train, y=y_train, batch_size=32, epochs=3, validation_data=(x_test, y_test)) 51 | 52 | # save model 53 | save_model_weights(model) 54 | --------------------------------------------------------------------------------