├── .gitignore ├── LICENSE ├── README.md ├── function.py ├── model.py ├── module.py ├── post_training_quantize.py ├── quantization_aware_training.py └── train.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | pip-wheel-metadata/ 24 | share/python-wheels/ 25 | *.egg-info/ 26 | .installed.cfg 27 | *.egg 28 | MANIFEST 29 | 30 | # PyInstaller 31 | # Usually these files are written by a python script from a template 32 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 33 | *.manifest 34 | *.spec 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .nox/ 44 | .coverage 45 | .coverage.* 46 | .cache 47 | nosetests.xml 48 | coverage.xml 49 | *.cover 50 | *.py,cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | target/ 76 | 77 | # Jupyter Notebook 78 | .ipynb_checkpoints 79 | 80 | # IPython 81 | profile_default/ 82 | ipython_config.py 83 | 84 | # pyenv 85 | .python-version 86 | 87 | # pipenv 88 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 89 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 90 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 91 | # install all needed dependencies. 92 | #Pipfile.lock 93 | 94 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 95 | __pypackages__/ 96 | 97 | # Celery stuff 98 | celerybeat-schedule 99 | celerybeat.pid 100 | 101 | # SageMath parsed files 102 | *.sage.py 103 | 104 | # Environments 105 | .env 106 | .venv 107 | env/ 108 | venv/ 109 | ENV/ 110 | env.bak/ 111 | venv.bak/ 112 | 113 | # Spyder project settings 114 | .spyderproject 115 | .spyproject 116 | 117 | # Rope project settings 118 | .ropeproject 119 | 120 | # mkdocs documentation 121 | /site 122 | 123 | # mypy 124 | .mypy_cache/ 125 | .dmypy.json 126 | dmypy.json 127 | 128 | # Pyre type checker 129 | .pyre/ 130 | 131 | # Pytorch weight 132 | *.pt 133 | 134 | # data dir 135 | /data -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # pytorch-quantization-demo 2 | A simple network quantization demo using pytorch from scratch. This is the code for my [tutorial](https://mp.weixin.qq.com/s?__biz=Mzg4ODA3MDkyMA==&mid=2247483692&idx=1&sn=3e28db4881d591f4e6a66c83d4213823&chksm=cf81f74bf8f67e5d0f2a98fd7bf7a91864d14010d88a5ed89120b7b4fcd94fc34789f0d0db9a&token=680347690&lang=zh_CN#rd) about network quantization written in Chinese. 3 | 4 | 也欢迎感兴趣的读者关注我的知乎专栏：[大白话模型量化](https://zhuanlan.zhihu.com/c_1258047709686231040) -------------------------------------------------------------------------------- /function.py: -------------------------------------------------------------------------------- 1 | from torch.autograd import Function 2 | import torch 3 | 4 | 5 | class FakeQuantize(Function): 6 | 7 | @staticmethod 8 | def forward(ctx, x, qparam): 9 | x = qparam.quantize_tensor(x) 10 | x = qparam.dequantize_tensor(x) 11 | return x 12 | 13 | @staticmethod 14 | def backward(ctx, grad_output): 15 | return grad_output, None 16 | 17 | def interp(x: torch.Tensor, xp: torch.Tensor, fp: torch.Tensor) -> torch.Tensor: 18 | x_ = x.reshape(x.size(0), -1) 19 | xp = xp.unsqueeze(0) 20 | fp = fp.unsqueeze(0) 21 | 22 | m = (fp[:,1:] - fp[:,:-1]) / (xp[:,1:] - xp[:,:-1]) #slope 23 | b = fp[:, :-1] - (m.mul(xp[:, :-1]) ) 24 | 25 | indicies = torch.sum(torch.ge(x_[:, :, None], xp[:, None, :]), -1) - 1 #torch.ge: x[i] >= xp[i] ? true: false 26 | indicies = torch.clamp(indicies, 0, m.shape[-1] - 1) 27 | 28 | line_idx = torch.linspace(0, indicies.shape[0], 1, device=indicies.device).to(torch.long) 29 | line_idx = line_idx.expand(indicies.shape) 30 | # idx = torch.cat([line_idx, indicies] , 0) 31 | out = m[line_idx, indicies].mul(x_) + b[line_idx, indicies] 32 | out = out.reshape(x.shape) 33 | return out -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | from module import * 6 | 7 | 8 | class Net(nn.Module): 9 | 10 | def __init__(self, num_channels=1): 11 | super(Net, self).__init__() 12 | self.conv1 = nn.Conv2d(num_channels, 40, 3, 1) 13 | self.conv2 = nn.Conv2d(40, 40, 3, 1, groups=20) 14 | self.fc = nn.Linear(5*5*40, 10) 15 | 16 | def forward(self, x): 17 | x = F.relu(self.conv1(x)) 18 | x = F.max_pool2d(x, 2, 2) 19 | x = F.relu(self.conv2(x)) 20 | x = F.max_pool2d(x, 2, 2) 21 | x = x.view(-1, 5*5*40) 22 | x = self.fc(x) 23 | return x 24 | 25 | def quantize(self, num_bits=8): 26 | self.qconv1 = QConv2d(self.conv1, qi=True, qo=True, num_bits=num_bits) 27 | self.qrelu1 = QReLU() 28 | self.qmaxpool2d_1 = QMaxPooling2d(kernel_size=2, stride=2, padding=0) 29 | self.qconv2 = QConv2d(self.conv2, qi=False, qo=True, num_bits=num_bits) 30 | self.qrelu2 = QReLU() 31 | self.qmaxpool2d_2 = QMaxPooling2d(kernel_size=2, stride=2, padding=0) 32 | self.qfc = QLinear(self.fc, qi=False, qo=True, num_bits=num_bits) 33 | 34 | def quantize_forward(self, x): 35 | x = self.qconv1(x) 36 | x = self.qrelu1(x) 37 | x = self.qmaxpool2d_1(x) 38 | x = self.qconv2(x) 39 | x = self.qrelu2(x) 40 | x = self.qmaxpool2d_2(x) 41 | x = x.view(-1, 5*5*40) 42 | x = self.qfc(x) 43 | return x 44 | 45 | def freeze(self): 46 | self.qconv1.freeze() 47 | self.qrelu1.freeze(self.qconv1.qo) 48 | self.qmaxpool2d_1.freeze(self.qconv1.qo) 49 | self.qconv2.freeze(qi=self.qconv1.qo) 50 | self.qrelu2.freeze(self.qconv2.qo) 51 | self.qmaxpool2d_2.freeze(self.qconv2.qo) 52 | self.qfc.freeze(qi=self.qconv2.qo) 53 | 54 | def quantize_inference(self, x): 55 | qx = self.qconv1.qi.quantize_tensor(x) 56 | qx = self.qconv1.quantize_inference(qx) 57 | qx = self.qrelu1.quantize_inference(qx) 58 | qx = self.qmaxpool2d_1.quantize_inference(qx) 59 | qx = self.qconv2.quantize_inference(qx) 60 | qx = self.qrelu2.quantize_inference(qx) 61 | qx = self.qmaxpool2d_2.quantize_inference(qx) 62 | qx = qx.view(-1, 5*5*40) 63 | qx = self.qfc.quantize_inference(qx) 64 | out = self.qfc.qo.dequantize_tensor(qx) 65 | return out 66 | 67 | 68 | class NetBN(nn.Module): 69 | 70 | def __init__(self, num_channels=1): 71 | super(NetBN, self).__init__() 72 | self.conv1 = nn.Conv2d(num_channels, 40, 3, 1) 73 | self.bn1 = nn.BatchNorm2d(40) 74 | self.conv2 = nn.Conv2d(40, 40, 3, 1) 75 | self.bn2 = nn.BatchNorm2d(40) 76 | self.fc = nn.Linear(5 * 5 * 40, 10) 77 | 78 | def forward(self, x): 79 | x = self.conv1(x) 80 | x = self.bn1(x) 81 | x = F.relu(x) 82 | x = F.max_pool2d(x, 2, 2) 83 | x = self.conv2(x) 84 | x = self.bn2(x) 85 | x = F.relu(x) 86 | x = F.max_pool2d(x, 2, 2) 87 | x = x.view(-1, 5 * 5 * 40) 88 | x = self.fc(x) 89 | return x 90 | 91 | def quantize(self, num_bits=8): 92 | self.qconv1 = QConvBNReLU(self.conv1, self.bn1, qi=True, qo=True, num_bits=num_bits) 93 | self.qmaxpool2d_1 = QMaxPooling2d(kernel_size=2, stride=2, padding=0) 94 | self.qconv2 = QConvBNReLU(self.conv2, self.bn2, qi=False, qo=True, num_bits=num_bits) 95 | self.qmaxpool2d_2 = QMaxPooling2d(kernel_size=2, stride=2, padding=0) 96 | self.qfc = QLinear(self.fc, qi=False, qo=True, num_bits=num_bits) 97 | 98 | def quantize_forward(self, x): 99 | x = self.qconv1(x) 100 | x = self.qmaxpool2d_1(x) 101 | x = self.qconv2(x) 102 | x = self.qmaxpool2d_2(x) 103 | x = x.view(-1, 5*5*40) 104 | x = self.qfc(x) 105 | return x 106 | 107 | def freeze(self): 108 | self.qconv1.freeze() 109 | self.qmaxpool2d_1.freeze(self.qconv1.qo) 110 | self.qconv2.freeze(qi=self.qconv1.qo) 111 | self.qmaxpool2d_2.freeze(self.qconv2.qo) 112 | self.qfc.freeze(qi=self.qconv2.qo) 113 | 114 | def quantize_inference(self, x): 115 | qx = self.qconv1.qi.quantize_tensor(x) 116 | qx = self.qconv1.quantize_inference(qx) 117 | qx = self.qmaxpool2d_1.quantize_inference(qx) 118 | qx = self.qconv2.quantize_inference(qx) 119 | qx = self.qmaxpool2d_2.quantize_inference(qx) 120 | qx = qx.view(-1, 5*5*40) 121 | 122 | qx = self.qfc.quantize_inference(qx) 123 | 124 | out = self.qfc.qo.dequantize_tensor(qx) 125 | return out 126 | -------------------------------------------------------------------------------- /module.py: -------------------------------------------------------------------------------- 1 | import math 2 | import numpy as np 3 | 4 | import torch 5 | import torch.nn as nn 6 | import torch.nn.functional as F 7 | from torch.autograd import Variable 8 | 9 | from function import FakeQuantize, interp 10 | 11 | 12 | def calcScaleZeroPoint(min_val, max_val, num_bits=8): 13 | qmin = 0. 14 | qmax = 2. ** num_bits - 1. 15 | scale = (max_val - min_val) / (qmax - qmin) 16 | 17 | zero_point = qmax - max_val / scale 18 | 19 | if zero_point < qmin: 20 | zero_point = torch.tensor([qmin], dtype=torch.float32).to(min_val.device) 21 | elif zero_point > qmax: 22 | # zero_point = qmax 23 | zero_point = torch.tensor([qmax], dtype=torch.float32).to(max_val.device) 24 | 25 | zero_point.round_() 26 | 27 | return scale, zero_point 28 | 29 | def quantize_tensor(x, scale, zero_point, num_bits=8, signed=False): 30 | if signed: 31 | qmin = - 2. ** (num_bits - 1) 32 | qmax = 2. ** (num_bits - 1) - 1 33 | else: 34 | qmin = 0. 35 | qmax = 2. ** num_bits - 1. 36 | 37 | q_x = zero_point + x / scale 38 | q_x.clamp_(qmin, qmax).round_() 39 | 40 | return q_x 41 | 42 | def dequantize_tensor(q_x, scale, zero_point): 43 | return scale * (q_x - zero_point) 44 | 45 | 46 | def search(M): 47 | P = 7000 48 | n = 1 49 | while True: 50 | Mo = int(round(2 ** n * M)) 51 | # Mo 52 | approx_result = Mo * P >> n 53 | result = int(round(M * P)) 54 | error = approx_result - result 55 | 56 | print("n=%d, Mo=%f, approx=%d, result=%d, error=%f" % \ 57 | (n, Mo, approx_result, result, error)) 58 | 59 | if math.fabs(error) < 1e-9 or n >= 22: 60 | return Mo, n 61 | n += 1 62 | 63 | 64 | class QParam(nn.Module): 65 | 66 | def __init__(self, num_bits=8): 67 | super(QParam, self).__init__() 68 | self.num_bits = num_bits 69 | scale = torch.tensor([], requires_grad=False) 70 | zero_point = torch.tensor([], requires_grad=False) 71 | min = torch.tensor([], requires_grad=False) 72 | max = torch.tensor([], requires_grad=False) 73 | self.register_buffer('scale', scale) 74 | self.register_buffer('zero_point', zero_point) 75 | self.register_buffer('min', min) 76 | self.register_buffer('max', max) 77 | 78 | def update(self, tensor): 79 | if self.max.nelement() == 0 or self.max.data < tensor.max().data: 80 | self.max.data = tensor.max().data 81 | self.max.clamp_(min=0) 82 | 83 | if self.min.nelement() == 0 or self.min.data > tensor.min().data: 84 | self.min.data = tensor.min().data 85 | self.min.clamp_(max=0) 86 | 87 | self.scale, self.zero_point = calcScaleZeroPoint(self.min, self.max, self.num_bits) 88 | 89 | def quantize_tensor(self, tensor): 90 | return quantize_tensor(tensor, self.scale, self.zero_point, num_bits=self.num_bits) 91 | 92 | def dequantize_tensor(self, q_x): 93 | return dequantize_tensor(q_x, self.scale, self.zero_point) 94 | 95 | def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): 96 | key_names = ['scale', 'zero_point', 'min', 'max'] 97 | for key in key_names: 98 | value = getattr(self, key) 99 | value.data = state_dict[prefix + key].data 100 | state_dict.pop(prefix + key) 101 | 102 | def __str__(self): 103 | info = 'scale: %.10f ' % self.scale 104 | info += 'zp: %d ' % self.zero_point 105 | info += 'min: %.6f ' % self.min 106 | info += 'max: %.6f' % self.max 107 | return info 108 | 109 | 110 | 111 | class QModule(nn.Module): 112 | 113 | def __init__(self, qi=True, qo=True, num_bits=8): 114 | super(QModule, self).__init__() 115 | if qi: 116 | self.qi = QParam(num_bits=num_bits) 117 | if qo: 118 | self.qo = QParam(num_bits=num_bits) 119 | 120 | def freeze(self): 121 | pass 122 | 123 | def quantize_inference(self, x): 124 | raise NotImplementedError('quantize_inference should be implemented.') 125 | 126 | 127 | class QConv2d(QModule): 128 | 129 | def __init__(self, conv_module, qi=True, qo=True, num_bits=8): 130 | super(QConv2d, self).__init__(qi=qi, qo=qo, num_bits=num_bits) 131 | self.num_bits = num_bits 132 | self.conv_module = conv_module 133 | self.qw = QParam(num_bits=num_bits) 134 | self.register_buffer('M', torch.tensor([], requires_grad=False)) # 将M注册为buffer 135 | 136 | def freeze(self, qi=None, qo=None): 137 | 138 | if hasattr(self, 'qi') and qi is not None: 139 | raise ValueError('qi has been provided in init function.') 140 | if not hasattr(self, 'qi') and qi is None: 141 | raise ValueError('qi is not existed, should be provided.') 142 | 143 | if hasattr(self, 'qo') and qo is not None: 144 | raise ValueError('qo has been provided in init function.') 145 | if not hasattr(self, 'qo') and qo is None: 146 | raise ValueError('qo is not existed, should be provided.') 147 | 148 | if qi is not None: 149 | self.qi = qi 150 | if qo is not None: 151 | self.qo = qo 152 | self.M.data = (self.qw.scale * self.qi.scale / self.qo.scale).data 153 | 154 | self.conv_module.weight.data = self.qw.quantize_tensor(self.conv_module.weight.data) 155 | self.conv_module.weight.data = self.conv_module.weight.data - self.qw.zero_point 156 | 157 | self.conv_module.bias.data = quantize_tensor(self.conv_module.bias.data, scale=self.qi.scale * self.qw.scale, 158 | zero_point=0, num_bits=32, signed=True) 159 | 160 | def forward(self, x): 161 | if hasattr(self, 'qi'): 162 | self.qi.update(x) 163 | x = FakeQuantize.apply(x, self.qi) 164 | 165 | self.qw.update(self.conv_module.weight.data) 166 | 167 | x = F.conv2d(x, FakeQuantize.apply(self.conv_module.weight, self.qw), self.conv_module.bias, 168 | stride=self.conv_module.stride, 169 | padding=self.conv_module.padding, dilation=self.conv_module.dilation, 170 | groups=self.conv_module.groups) 171 | 172 | if hasattr(self, 'qo'): 173 | self.qo.update(x) 174 | x = FakeQuantize.apply(x, self.qo) 175 | 176 | return x 177 | 178 | def quantize_inference(self, x): 179 | x = x - self.qi.zero_point 180 | x = self.conv_module(x) 181 | x = self.M * x 182 | x.round_() 183 | x = x + self.qo.zero_point 184 | x.clamp_(0., 2.**self.num_bits-1.).round_() 185 | return x 186 | 187 | 188 | class QLinear(QModule): 189 | 190 | def __init__(self, fc_module, qi=True, qo=True, num_bits=8): 191 | super(QLinear, self).__init__(qi=qi, qo=qo, num_bits=num_bits) 192 | self.num_bits = num_bits 193 | self.fc_module = fc_module 194 | self.qw = QParam(num_bits=num_bits) 195 | self.register_buffer('M', torch.tensor([], requires_grad=False)) # 将M注册为buffer 196 | 197 | def freeze(self, qi=None, qo=None): 198 | 199 | if hasattr(self, 'qi') and qi is not None: 200 | raise ValueError('qi has been provided in init function.') 201 | if not hasattr(self, 'qi') and qi is None: 202 | raise ValueError('qi is not existed, should be provided.') 203 | 204 | if hasattr(self, 'qo') and qo is not None: 205 | raise ValueError('qo has been provided in init function.') 206 | if not hasattr(self, 'qo') and qo is None: 207 | raise ValueError('qo is not existed, should be provided.') 208 | 209 | if qi is not None: 210 | self.qi = qi 211 | if qo is not None: 212 | self.qo = qo 213 | self.M.data = (self.qw.scale * self.qi.scale / self.qo.scale).data 214 | 215 | self.fc_module.weight.data = self.qw.quantize_tensor(self.fc_module.weight.data) 216 | self.fc_module.weight.data = self.fc_module.weight.data - self.qw.zero_point 217 | self.fc_module.bias.data = quantize_tensor(self.fc_module.bias.data, scale=self.qi.scale * self.qw.scale, 218 | zero_point=0, num_bits=32, signed=True) 219 | 220 | def forward(self, x): 221 | if hasattr(self, 'qi'): 222 | self.qi.update(x) 223 | x = FakeQuantize.apply(x, self.qi) 224 | 225 | self.qw.update(self.fc_module.weight.data) 226 | 227 | x = F.linear(x, FakeQuantize.apply(self.fc_module.weight, self.qw), self.fc_module.bias) 228 | 229 | if hasattr(self, 'qo'): 230 | self.qo.update(x) 231 | x = FakeQuantize.apply(x, self.qo) 232 | 233 | return x 234 | 235 | def quantize_inference(self, x): 236 | x = x - self.qi.zero_point 237 | x = self.fc_module(x) 238 | x = self.M * x 239 | x.round_() 240 | x = x + self.qo.zero_point 241 | x.clamp_(0., 2.**self.num_bits-1.).round_() 242 | return x 243 | 244 | 245 | class QReLU(QModule): 246 | 247 | def __init__(self, qi=False, num_bits=None): 248 | super(QReLU, self).__init__(qi=qi, num_bits=num_bits) 249 | 250 | def freeze(self, qi=None): 251 | 252 | if hasattr(self, 'qi') and qi is not None: 253 | raise ValueError('qi has been provided in init function.') 254 | if not hasattr(self, 'qi') and qi is None: 255 | raise ValueError('qi is not existed, should be provided.') 256 | 257 | if qi is not None: 258 | self.qi = qi 259 | 260 | def forward(self, x): 261 | if hasattr(self, 'qi'): 262 | self.qi.update(x) 263 | x = FakeQuantize.apply(x, self.qi) 264 | 265 | x = F.relu(x) 266 | 267 | return x 268 | 269 | def quantize_inference(self, x): 270 | x = x.clone() 271 | x[x < self.qi.zero_point] = self.qi.zero_point 272 | return x 273 | 274 | class QMaxPooling2d(QModule): 275 | 276 | def __init__(self, kernel_size=3, stride=1, padding=0, qi=False, num_bits=None): 277 | super(QMaxPooling2d, self).__init__(qi=qi, num_bits=num_bits) 278 | self.kernel_size = kernel_size 279 | self.stride = stride 280 | self.padding = padding 281 | 282 | def freeze(self, qi=None): 283 | if hasattr(self, 'qi') and qi is not None: 284 | raise ValueError('qi has been provided in init function.') 285 | if not hasattr(self, 'qi') and qi is None: 286 | raise ValueError('qi is not existed, should be provided.') 287 | if qi is not None: 288 | self.qi = qi 289 | 290 | def forward(self, x): 291 | if hasattr(self, 'qi'): 292 | self.qi.update(x) 293 | x = FakeQuantize.apply(x, self.qi) 294 | 295 | x = F.max_pool2d(x, self.kernel_size, self.stride, self.padding) 296 | 297 | return x 298 | 299 | def quantize_inference(self, x): 300 | return F.max_pool2d(x, self.kernel_size, self.stride, self.padding) 301 | 302 | 303 | class QConvBNReLU(QModule): 304 | 305 | def __init__(self, conv_module, bn_module, qi=True, qo=True, num_bits=8): 306 | super(QConvBNReLU, self).__init__(qi=qi, qo=qo, num_bits=num_bits) 307 | self.num_bits = num_bits 308 | self.conv_module = conv_module 309 | self.bn_module = bn_module 310 | self.qw = QParam(num_bits=num_bits) 311 | self.qb = QParam(num_bits=32) 312 | self.register_buffer('M', torch.tensor([], requires_grad=False)) # 将M注册为buffer 313 | 314 | def fold_bn(self, mean, std): 315 | if self.bn_module.affine: 316 | gamma_ = self.bn_module.weight / std 317 | weight = self.conv_module.weight * gamma_.view(self.conv_module.out_channels, 1, 1, 1) 318 | if self.conv_module.bias is not None: 319 | bias = gamma_ * self.conv_module.bias - gamma_ * mean + self.bn_module.bias 320 | else: 321 | bias = self.bn_module.bias - gamma_ * mean 322 | else: 323 | gamma_ = 1 / std 324 | weight = self.conv_module.weight * gamma_ 325 | if self.conv_module.bias is not None: 326 | bias = gamma_ * self.conv_module.bias - gamma_ * mean 327 | else: 328 | bias = -gamma_ * mean 329 | 330 | return weight, bias 331 | 332 | 333 | def forward(self, x): 334 | 335 | if hasattr(self, 'qi'): 336 | self.qi.update(x) 337 | x = FakeQuantize.apply(x, self.qi) 338 | 339 | if self.training: 340 | y = F.conv2d(x, self.conv_module.weight, self.conv_module.bias, 341 | stride=self.conv_module.stride, 342 | padding=self.conv_module.padding, 343 | dilation=self.conv_module.dilation, 344 | groups=self.conv_module.groups) 345 | y = y.permute(1, 0, 2, 3) # NCHW -> CNHW 346 | y = y.contiguous().view(self.conv_module.out_channels, -1) # CNHW -> C,NHW 347 | # mean = y.mean(1) 348 | # var = y.var(1) 349 | mean = y.mean(1).detach() 350 | var = y.var(1).detach() 351 | self.bn_module.running_mean = \ 352 | (1 - self.bn_module.momentum) * self.bn_module.running_mean + \ 353 | self.bn_module.momentum * mean 354 | self.bn_module.running_var = \ 355 | (1 - self.bn_module.momentum) * self.bn_module.running_var + \ 356 | self.bn_module.momentum * var 357 | else: 358 | mean = Variable(self.bn_module.running_mean) 359 | var = Variable(self.bn_module.running_var) 360 | 361 | std = torch.sqrt(var + self.bn_module.eps) 362 | 363 | weight, bias = self.fold_bn(mean, std) 364 | 365 | self.qw.update(weight.data) 366 | 367 | x = F.conv2d(x, FakeQuantize.apply(weight, self.qw), bias, 368 | stride=self.conv_module.stride, 369 | padding=self.conv_module.padding, dilation=self.conv_module.dilation, 370 | groups=self.conv_module.groups) 371 | 372 | x = F.relu(x) 373 | 374 | if hasattr(self, 'qo'): 375 | self.qo.update(x) 376 | x = FakeQuantize.apply(x, self.qo) 377 | 378 | return x 379 | 380 | def freeze(self, qi=None, qo=None): 381 | if hasattr(self, 'qi') and qi is not None: 382 | raise ValueError('qi has been provided in init function.') 383 | if not hasattr(self, 'qi') and qi is None: 384 | raise ValueError('qi is not existed, should be provided.') 385 | 386 | if hasattr(self, 'qo') and qo is not None: 387 | raise ValueError('qo has been provided in init function.') 388 | if not hasattr(self, 'qo') and qo is None: 389 | raise ValueError('qo is not existed, should be provided.') 390 | 391 | if qi is not None: 392 | self.qi = qi 393 | if qo is not None: 394 | self.qo = qo 395 | self.M.data = (self.qw.scale * self.qi.scale / self.qo.scale).data 396 | 397 | std = torch.sqrt(self.bn_module.running_var + self.bn_module.eps) 398 | 399 | weight, bias = self.fold_bn(self.bn_module.running_mean, std) 400 | self.conv_module.weight.data = self.qw.quantize_tensor(weight.data) 401 | self.conv_module.weight.data = self.conv_module.weight.data - self.qw.zero_point 402 | 403 | self.conv_module.bias.data = quantize_tensor(bias, scale=self.qi.scale * self.qw.scale, 404 | zero_point=0, num_bits=32, signed=True) 405 | 406 | def quantize_inference(self, x): 407 | x = x - self.qi.zero_point 408 | x = self.conv_module(x) 409 | x = self.M * x 410 | x.round_() 411 | x = x + self.qo.zero_point 412 | x.clamp_(0., 2.**self.num_bits-1.).round_() 413 | return x 414 | 415 | 416 | class QSigmoid(QModule): 417 | 418 | def __init__(self, qi=True, qo=True, num_bits=8, lut_size=64): 419 | super(QSigmoid, self).__init__(qi=qi, qo=qo, num_bits=num_bits) 420 | self.num_bits = num_bits 421 | self.lut_size = lut_size 422 | 423 | def forward(self, x): 424 | if hasattr(self, 'qi'): 425 | self.qi.update(x) 426 | x = FakeQuantize.apply(x, self.qi) 427 | 428 | x = torch.sigmoid(x) 429 | 430 | if hasattr(self, 'qo'): 431 | self.qo.update(x) 432 | x = FakeQuantize.apply(x, self.qo) 433 | 434 | return x 435 | 436 | def freeze(self, qi=None, qo=None): 437 | if hasattr(self, 'qi') and qi is not None: 438 | raise ValueError('qi has been provided in init function.') 439 | if not hasattr(self, 'qi') and qi is None: 440 | raise ValueError('qi is not existed, should be provided.') 441 | 442 | if hasattr(self, 'qo') and qo is not None: 443 | raise ValueError('qo has been provided in init function.') 444 | if not hasattr(self, 'qo') and qo is None: 445 | raise ValueError('qo is not existed, should be provided.') 446 | 447 | if qi is not None: 448 | self.qi = qi 449 | if qo is not None: 450 | self.qo = qo 451 | 452 | lut_qx = torch.tensor(np.linspace(0, 2 ** self.num_bits - 1, self.lut_size), dtype=torch.uint8) 453 | lut_x = self.qi.dequantize_tensor(lut_qx) 454 | lut_y = torch.sigmoid(lut_x) 455 | lut_qy = self.qo.quantize_tensor(lut_y) 456 | 457 | self.register_buffer('lut_qy', lut_qy) 458 | self.register_buffer('lut_qx', lut_qx) 459 | 460 | 461 | def quantize_inference(self, x): 462 | y = interp(x, self.lut_qx, self.lut_qy) 463 | y = y.round_().clamp_(0., 2.**self.num_bits-1.) 464 | return y -------------------------------------------------------------------------------- /post_training_quantize.py: -------------------------------------------------------------------------------- 1 | from torch.serialization import load 2 | from model import * 3 | 4 | import argparse 5 | import torch 6 | import torch.nn as nn 7 | import torch.optim as optim 8 | from torchvision import datasets, transforms 9 | import os 10 | import os.path as osp 11 | 12 | 13 | def direct_quantize(model, test_loader): 14 | for i, (data, target) in enumerate(test_loader, 1): 15 | output = model.quantize_forward(data) 16 | if i % 500 == 0: 17 | break 18 | print('direct quantization finish') 19 | 20 | 21 | def full_inference(model, test_loader): 22 | correct = 0 23 | for i, (data, target) in enumerate(test_loader, 1): 24 | output = model(data) 25 | pred = output.argmax(dim=1, keepdim=True) 26 | correct += pred.eq(target.view_as(pred)).sum().item() 27 | print('\nTest set: Full Model Accuracy: {:.0f}%\n'.format(100. * correct / len(test_loader.dataset))) 28 | 29 | 30 | def quantize_inference(model, test_loader): 31 | correct = 0 32 | for i, (data, target) in enumerate(test_loader, 1): 33 | output = model.quantize_inference(data) 34 | pred = output.argmax(dim=1, keepdim=True) 35 | correct += pred.eq(target.view_as(pred)).sum().item() 36 | print('\nTest set: Quant Model Accuracy: {:.0f}%\n'.format(100. * correct / len(test_loader.dataset))) 37 | 38 | 39 | if __name__ == "__main__": 40 | batch_size = 64 41 | using_bn = True 42 | load_quant_model_file = None 43 | # load_model_file = None 44 | 45 | train_loader = torch.utils.data.DataLoader( 46 | datasets.MNIST('data', train=True, download=True, 47 | transform=transforms.Compose([ 48 | transforms.ToTensor(), 49 | transforms.Normalize((0.1307,), (0.3081,)) 50 | ])), 51 | batch_size=batch_size, shuffle=True, num_workers=1, pin_memory=True 52 | ) 53 | 54 | test_loader = torch.utils.data.DataLoader( 55 | datasets.MNIST('data', train=False, transform=transforms.Compose([ 56 | transforms.ToTensor(), 57 | transforms.Normalize((0.1307,), (0.3081,)) 58 | ])), 59 | batch_size=batch_size, shuffle=False, num_workers=1, pin_memory=True 60 | ) 61 | 62 | if using_bn: 63 | model = NetBN() 64 | model.load_state_dict(torch.load('ckpt/mnist_cnnbn.pt', map_location='cpu')) 65 | save_file = "ckpt/mnist_cnnbn_ptq.pt" 66 | else: 67 | model = Net() 68 | model.load_state_dict(torch.load('ckpt/mnist_cnn.pt', map_location='cpu')) 69 | save_file = "ckpt/mnist_cnn_ptq.pt" 70 | 71 | model.eval() 72 | full_inference(model, test_loader) 73 | 74 | num_bits = 8 75 | model.quantize(num_bits=num_bits) 76 | model.eval() 77 | print('Quantization bit: %d' % num_bits) 78 | 79 | 80 | if load_quant_model_file is not None: 81 | model.load_state_dict(torch.load(load_quant_model_file)) 82 | print("Successfully load quantized model %s" % load_quant_model_file) 83 | 84 | 85 | direct_quantize(model, train_loader) 86 | 87 | torch.save(model.state_dict(), save_file) 88 | model.freeze() 89 | 90 | # 测试是否设备转移是否正确 91 | # model.cuda() 92 | # print(model.qconv1.M.device) 93 | # model.cpu() 94 | # print(model.qconv1.M.device) 95 | 96 | quantize_inference(model, test_loader) 97 | 98 | 99 | 100 | 101 | 102 | 103 | -------------------------------------------------------------------------------- /quantization_aware_training.py: -------------------------------------------------------------------------------- 1 | from model import * 2 | 3 | import argparse 4 | import torch 5 | import torch.nn as nn 6 | import torch.optim as optim 7 | from torchvision import datasets, transforms 8 | import os 9 | import os.path as osp 10 | 11 | 12 | def quantize_aware_training(model, device, train_loader, optimizer, epoch): 13 | lossLayer = torch.nn.CrossEntropyLoss() 14 | for batch_idx, (data, target) in enumerate(train_loader, 1): 15 | data, target = data.to(device), target.to(device) 16 | optimizer.zero_grad() 17 | output = model.quantize_forward(data) 18 | loss = lossLayer(output, target) 19 | loss.backward() 20 | optimizer.step() 21 | 22 | if batch_idx % 50 == 0: 23 | print('Quantize Aware Training Epoch: {} [{}/{}]\tLoss: {:.6f}'.format( 24 | epoch, batch_idx * len(data), len(train_loader.dataset), loss.item() 25 | )) 26 | 27 | 28 | def full_inference(model, test_loader): 29 | correct = 0 30 | for i, (data, target) in enumerate(test_loader, 1): 31 | data, target = data.to(device), target.to(device) 32 | output = model(data) 33 | pred = output.argmax(dim=1, keepdim=True) 34 | correct += pred.eq(target.view_as(pred)).sum().item() 35 | print('\nTest set: Full Model Accuracy: {:.0f}%\n'.format(100. * correct / len(test_loader.dataset))) 36 | 37 | 38 | def quantize_inference(model, test_loader): 39 | correct = 0 40 | for i, (data, target) in enumerate(test_loader, 1): 41 | data, target = data.to(device), target.to(device) 42 | output = model.quantize_inference(data) 43 | pred = output.argmax(dim=1, keepdim=True) 44 | correct += pred.eq(target.view_as(pred)).sum().item() 45 | print('\nTest set: Quant Model Accuracy: {:.0f}%\n'.format(100. * correct / len(test_loader.dataset))) 46 | 47 | 48 | if __name__ == "__main__": 49 | batch_size = 64 50 | seed = 1 51 | epochs = 3 52 | lr = 0.01 53 | momentum = 0.5 54 | using_bn = True 55 | load_quant_model_file = None 56 | # load_quant_model_file = "ckpt/mnist_cnnbn_qat.pt" 57 | 58 | torch.manual_seed(seed) 59 | 60 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 61 | 62 | train_loader = torch.utils.data.DataLoader( 63 | datasets.MNIST('data', train=True, download=True, 64 | transform=transforms.Compose([ 65 | transforms.ToTensor(), 66 | transforms.Normalize((0.1307,), (0.3081,)) 67 | ])), 68 | batch_size=batch_size, shuffle=True, num_workers=1, pin_memory=False 69 | ) 70 | 71 | test_loader = torch.utils.data.DataLoader( 72 | datasets.MNIST('data', train=False, transform=transforms.Compose([ 73 | transforms.ToTensor(), 74 | transforms.Normalize((0.1307,), (0.3081,)) 75 | ])), 76 | batch_size=batch_size, shuffle=True, num_workers=1, pin_memory=False 77 | ) 78 | 79 | if using_bn: 80 | model = NetBN() 81 | model.load_state_dict(torch.load('ckpt/mnist_cnnbn.pt', map_location='cpu')) 82 | save_file = "ckpt/mnist_cnnbn_qat.pt" 83 | else: 84 | model = Net() 85 | model.load_state_dict(torch.load('ckpt/mnist_cnn.pt', map_location='cpu')) 86 | save_file = "ckpt/mnist_cnn_qat.pt" 87 | model.to(device) 88 | 89 | optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum) 90 | 91 | model.eval() 92 | 93 | full_inference(model, test_loader) 94 | 95 | num_bits = 8 96 | model.quantize(num_bits=num_bits) 97 | print('Quantization bit: %d' % num_bits) 98 | 99 | if load_quant_model_file is not None: 100 | model.load_state_dict(torch.load(load_quant_model_file)) 101 | print("Successfully load quantized model %s" % load_quant_model_file) 102 | 103 | model.train() 104 | 105 | for epoch in range(1, epochs + 1): 106 | quantize_aware_training(model, device, train_loader, optimizer, epoch) 107 | 108 | model.eval() 109 | torch.save(model.state_dict(), save_file) 110 | 111 | model.freeze() 112 | 113 | quantize_inference(model, test_loader) 114 | 115 | 116 | 117 | 118 | 119 | 120 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | from model import * 2 | 3 | import torch 4 | import torch.nn as nn 5 | import torch.optim as optim 6 | from torchvision import datasets, transforms 7 | import os 8 | import os.path as osp 9 | 10 | 11 | def train(model, device, train_loader, optimizer, epoch): 12 | model.train() 13 | lossLayer = torch.nn.CrossEntropyLoss() 14 | for batch_idx, (data, target) in enumerate(train_loader): 15 | data, target = data.to(device), target.to(device) 16 | optimizer.zero_grad() 17 | output = model(data) 18 | loss = lossLayer(output, target) 19 | loss.backward() 20 | optimizer.step() 21 | 22 | if batch_idx % 50 == 0: 23 | print('Train Epoch: {} [{}/{}]\tLoss: {:.6f}'.format( 24 | epoch, batch_idx * len(data), len(train_loader.dataset), loss.item() 25 | )) 26 | 27 | def test(model, device, test_loader): 28 | model.eval() 29 | test_loss = 0 30 | correct = 0 31 | lossLayer = torch.nn.CrossEntropyLoss(reduction='sum') 32 | for data, target in test_loader: 33 | data, target = data.to(device), target.to(device) 34 | output = model(data) 35 | test_loss += lossLayer(output, target).item() 36 | pred = output.argmax(dim=1, keepdim=True) 37 | correct += pred.eq(target.view_as(pred)).sum().item() 38 | 39 | test_loss /= len(test_loader.dataset) 40 | 41 | print('\nTest set: Average loss: {:.4f}, Accuracy: {:.0f}%\n'.format( 42 | test_loss, 100. * correct / len(test_loader.dataset) 43 | )) 44 | 45 | 46 | if __name__ == "__main__": 47 | batch_size = 64 48 | test_batch_size = 64 49 | seed = 1 50 | epochs = 15 51 | lr = 0.01 52 | momentum = 0.5 53 | save_model = True 54 | using_bn = True 55 | 56 | torch.manual_seed(seed) 57 | 58 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 59 | 60 | train_loader = torch.utils.data.DataLoader( 61 | datasets.MNIST('data', train=True, download=True, 62 | transform=transforms.Compose([ 63 | transforms.ToTensor(), 64 | transforms.Normalize((0.1307,), (0.3081,)) 65 | ])), 66 | batch_size=batch_size, shuffle=True, num_workers=1, pin_memory=True 67 | ) 68 | 69 | test_loader = torch.utils.data.DataLoader( 70 | datasets.MNIST('data', train=False, transform=transforms.Compose([ 71 | transforms.ToTensor(), 72 | transforms.Normalize((0.1307,), (0.3081,)) 73 | ])), 74 | batch_size=test_batch_size, shuffle=True, num_workers=1, pin_memory=True 75 | ) 76 | 77 | if using_bn: 78 | model = NetBN().to(device) 79 | else: 80 | model = Net().to(device) 81 | 82 | optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum) 83 | 84 | for epoch in range(1, epochs + 1): 85 | train(model, device, train_loader, optimizer, epoch) 86 | test(model, device, test_loader) 87 | 88 | if save_model: 89 | if not osp.exists('ckpt'): 90 | os.makedirs('ckpt') 91 | if using_bn: 92 | torch.save(model.state_dict(), 'ckpt/mnist_cnnbn.pt') 93 | else: 94 | torch.save(model.state_dict(), 'ckpt/mnist_cnn.pt') --------------------------------------------------------------------------------