├── LICENSE ├── README.md ├── setup.py └── xml2pytorch ├── __init__.py ├── converter.py └── xml_examples └── simpleCNN.xml /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2019 yfzhoucs 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # xml2pytorch 2 | Using xml to define pytorch neural networks 3 | ## What can it Do 4 | With xml2pytorch, you can easily define neural networks in xml, and then declare them in pytorch. 5 | 6 | RNN and LSTM are not supported currently. 7 | ## Installation 8 | ### Environment 9 | OS independent. Python3. (Not tested on Python2, but it should work.) 10 | ### Install requirements 11 | torch>=0.4.1 12 | numpy>=1.15.1 13 | ### Installing by pip3 14 | pip3 install xml2pytorch 15 | ## Quick Start 16 | ### How to declare the CNN defined by a xml file 17 | ``` 18 | import torch 19 | import xml2pytorch as xm 20 | 21 | # declare the net defined in .xml 22 | net = xm.convertXML(xml_filename) 23 | 24 | # input a random tensor 25 | x = torch.randn(1, 3, 32, 32) 26 | y = net(x) 27 | print(y) 28 | ``` 29 | ### How to define a simple CNN in xml 30 | ``` 31 | 32 | 33 | 34 | Conv2d 35 | 3 36 | 6 37 | 5 38 | 39 | 40 | ELU 41 | 42 | 43 | MaxPool2d 44 | 2 45 | 2 46 | logsigmoid 47 | 48 | 49 | Conv2d 50 | 6 51 | 16 52 | 5 53 | relu 54 | 55 | 56 | MaxPool2d 57 | 2 58 | 2 59 | relu 60 | 61 | 62 | reshape 63 | [-1, 16*5*5] 64 | 65 | 66 | Linear 67 | 400 68 | 120 69 | tanh 70 | 71 | 72 | Linear 73 | 120 74 | 84 75 | sigmoid 76 | 77 | 78 | Linear 79 | 84 80 | 10 81 | softmax 82 | 83 | 84 | 85 | ``` 86 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | import setuptools 2 | 3 | with open("./README.md", "r") as fh: 4 | long_description = fh.read() 5 | 6 | setuptools.setup( 7 | name="xml2pytorch", 8 | version="0.0.9", 9 | author="Yifan Zhou", 10 | author_email="yfzhou.cs@gmail.com", 11 | description="Using xml to define pytorch neural networks", 12 | long_description=long_description, 13 | long_description_content_type="text/markdown", 14 | url="https://github.com/yfzhoucs/xml2pytorch", 15 | packages=setuptools.find_packages(), 16 | install_requires = [ 17 | "torch>=0.4.1", 18 | "numpy>=1.15.1" 19 | ], 20 | classifiers=[ 21 | "Programming Language :: Python :: 3", 22 | "License :: OSI Approved :: MIT License", 23 | "Operating System :: OS Independent", 24 | ], 25 | ) -------------------------------------------------------------------------------- /xml2pytorch/__init__.py: -------------------------------------------------------------------------------- 1 | from xml2pytorch.converter import convertXML 2 | name = "xml2pytorch" -------------------------------------------------------------------------------- /xml2pytorch/converter.py: -------------------------------------------------------------------------------- 1 | import collections 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | import torch 5 | from xml.dom.minidom import parse 6 | import xml.dom.minidom 7 | import numpy as np 8 | 9 | def main(): 10 | simple_cnn = convertXML("./xml_examples/simpleCNN.xml") 11 | print(simple_cnn) 12 | x = torch.randn(1, 3, 32, 32) 13 | y = simple_cnn(x) 14 | print(y) 15 | 16 | 17 | def convertXML(xml_filename): 18 | # 使用minidom解析器打开 XML 文档 19 | DOMTree = xml.dom.minidom.parse(xml_filename) 20 | graph = DOMTree.documentElement 21 | 22 | nets = graph.getElementsByTagName("net") 23 | net = nets[0] 24 | net_model = Net(net) 25 | 26 | return net_model 27 | 28 | 29 | class Net(nn.Module): 30 | def __init__(self, net_xmlNode): 31 | super(Net, self).__init__() 32 | self.net_xmlNode = net_xmlNode 33 | self.layers = [] 34 | 35 | layers_xmlNode = net_xmlNode.getElementsByTagName("layer") 36 | 37 | for layer_xmlNode in layers_xmlNode: 38 | net_style = None 39 | args_dict = None 40 | activation = None 41 | 42 | # 去除 xml 里的换行符 43 | for i in range(len(layer_xmlNode.childNodes)-1, -1, -1): 44 | if(layer_xmlNode.childNodes[i].nodeName == "#text"): 45 | layer_xmlNode.removeChild(layer_xmlNode.childNodes[i]) 46 | 47 | # 隐藏层类型 48 | net_styles = layer_xmlNode.getElementsByTagName("net_style") 49 | if len(net_styles) > 0: 50 | net_style = net_styles[0].childNodes[0].data 51 | 52 | # 隐藏层的默认参数 53 | args_dict = self.__default_args_for_layer(net_style) 54 | # 隐藏层，xml 追加的参数 55 | args_dict = self.__check_args_from_xml(args_dict, layer_xmlNode) 56 | 57 | # 判断激活函数 58 | activations = layer_xmlNode.getElementsByTagName("activation") 59 | if len(activations) > 0: 60 | activation = activations[0].childNodes[0].data 61 | 62 | 63 | # 声明网络和激活函数 64 | if net_style is not None: 65 | self.__append_layer(net_style, args_dict) 66 | if activation is not None: 67 | self.__append_activation(activation) 68 | 69 | 70 | def __append_activation(self, activation): 71 | if activation == "relu": 72 | self.layers.append(F.relu) 73 | elif activation == "elu": 74 | self.layers.append(F.elu) 75 | elif activation == "selu": 76 | self.layers.append(F.selu) 77 | elif activation == "celu": 78 | self.layers.append(F.celu) 79 | elif activation == "leaky_relu": 80 | self.layers.append(F.leaky_relu) 81 | elif activation == "logsigmoid": 82 | self.layers.append(F.logsigmoid) 83 | elif activation == "softmax": 84 | self.layers.append(F.softmax) 85 | elif activation == "tanh": 86 | self.layers.append(torch.tanh) 87 | elif activation == "sigmoid": 88 | self.layers.append(torch.sigmoid) 89 | 90 | 91 | def __append_layer(self, net_style, args_dict): 92 | args_values_list = list(args_dict.values()) 93 | if net_style == "Conv2d": 94 | self.layers.append(nn.Conv2d(args_values_list[0], 95 | args_values_list[1], 96 | args_values_list[2], 97 | args_values_list[3], 98 | args_values_list[4], 99 | args_values_list[5], 100 | args_values_list[6], 101 | args_values_list[7])) 102 | elif net_style == "MaxPool2d": 103 | self.layers.append(nn.MaxPool2d(args_values_list[0], 104 | args_values_list[1], 105 | args_values_list[2], 106 | args_values_list[3], 107 | args_values_list[4], 108 | args_values_list[5])) 109 | elif net_style == "Linear": 110 | self.layers.append(nn.Linear(args_values_list[0], 111 | args_values_list[1], 112 | args_values_list[2])) 113 | elif net_style == "reshape": 114 | # 如果是特殊情况 reshape，就直接将目标向量尺寸传入 115 | # print(type(args_values_list[0])) 116 | self.layers.append(args_values_list[0]) 117 | elif net_style == "Conv1d": 118 | self.layers.append(nn.Conv1d(args_values_list[0], 119 | args_values_list[1], 120 | args_values_list[2], 121 | args_values_list[3], 122 | args_values_list[4], 123 | args_values_list[5], 124 | args_values_list[6], 125 | args_values_list[7])) 126 | elif net_style == "Conv3d": 127 | self.layers.append(nn.Conv3d(args_values_list[0], 128 | args_values_list[1], 129 | args_values_list[2], 130 | args_values_list[3], 131 | args_values_list[4], 132 | args_values_list[5], 133 | args_values_list[6], 134 | args_values_list[7])) 135 | elif net_style == "ConvTranspose1d": 136 | self.layers.append(nn.ConvTranspose1d(args_values_list[0], 137 | args_values_list[1], 138 | args_values_list[2], 139 | args_values_list[3], 140 | args_values_list[4], 141 | args_values_list[5], 142 | args_values_list[6], 143 | args_values_list[7], 144 | args_values_list[8])) 145 | elif net_style == "ConvTranspose2d": 146 | self.layers.append(nn.ConvTranspose2d(args_values_list[0], 147 | args_values_list[1], 148 | args_values_list[2], 149 | args_values_list[3], 150 | args_values_list[4], 151 | args_values_list[5], 152 | args_values_list[6], 153 | args_values_list[7], 154 | args_values_list[8])) 155 | elif net_style == "ConvTranspose3d": 156 | self.layers.append(nn.ConvTranspose3d(args_values_list[0], 157 | args_values_list[1], 158 | args_values_list[2], 159 | args_values_list[3], 160 | args_values_list[4], 161 | args_values_list[5], 162 | args_values_list[6], 163 | args_values_list[7], 164 | args_values_list[8])) 165 | elif net_style == "Unfold": 166 | self.layers.append(nn.Unfold(args_values_list[0], 167 | args_values_list[1], 168 | args_values_list[2], 169 | args_values_list[3])) 170 | elif net_style == "Fold": 171 | self.layers.append(nn.Unfold(args_values_list[0], 172 | args_values_list[1], 173 | args_values_list[2], 174 | args_values_list[3], 175 | args_values_list[4])) 176 | elif net_style == "MaxPool1d": 177 | self.layers.append(nn.MaxPool1d(args_values_list[0], 178 | args_values_list[1], 179 | args_values_list[2], 180 | args_values_list[3], 181 | args_values_list[4], 182 | args_values_list[5])) 183 | elif net_style == "MaxPool3d": 184 | self.layers.append(nn.MaxPool3d(args_values_list[0], 185 | args_values_list[1], 186 | args_values_list[2], 187 | args_values_list[3], 188 | args_values_list[4], 189 | args_values_list[5])) 190 | elif net_style == "MaxUnpool1d": 191 | self.layers.append(nn.MaxUnpool1d(args_values_list[0], 192 | args_values_list[1], 193 | args_values_list[2])) 194 | elif net_style == "MaxUnpool2d": 195 | self.layers.append(nn.MaxUnpool2d(args_values_list[0], 196 | args_values_list[1], 197 | args_values_list[2])) 198 | elif net_style == "MaxUnpool3d": 199 | self.layers.append(nn.MaxUnpool3d(args_values_list[0], 200 | args_values_list[1], 201 | args_values_list[2])) 202 | elif net_style == "AvgPool1d": 203 | self.layers.append(nn.AvgPool1d(args_values_list[0], 204 | args_values_list[1], 205 | args_values_list[2], 206 | args_values_list[3], 207 | args_values_list[4])) 208 | elif net_style == "AvgPool2d": 209 | self.layers.append(nn.AvgPool2d(args_values_list[0], 210 | args_values_list[1], 211 | args_values_list[2], 212 | args_values_list[3], 213 | args_values_list[4])) 214 | elif net_style == "AvgPool3d": 215 | self.layers.append(nn.AvgPool3d(args_values_list[0], 216 | args_values_list[1], 217 | args_values_list[2], 218 | args_values_list[3], 219 | args_values_list[4])) 220 | elif net_style == "FractionalMaxPool2d": 221 | self.layers.append(nn.FractionalMaxPool2d(args_values_list[0], 222 | args_values_list[1], 223 | args_values_list[2], 224 | args_values_list[3], 225 | args_values_list[4])) 226 | elif net_style == "LPPool1d": 227 | self.layers.append(nn.LPPool1d(args_values_list[0], 228 | args_values_list[1], 229 | args_values_list[2], 230 | args_values_list[3])) 231 | elif net_style == "LPPool2d": 232 | self.layers.append(nn.LPPool2d(args_values_list[0], 233 | args_values_list[1], 234 | args_values_list[2], 235 | args_values_list[3])) 236 | elif net_style == "AdaptiveMaxPool1d": 237 | self.layers.append(nn.AdaptiveMaxPool1d(args_values_list[0], 238 | args_values_list[1])) 239 | elif net_style == "AdaptiveMaxPool2d": 240 | self.layers.append(nn.AdaptiveMaxPool2d(args_values_list[0], 241 | args_values_list[1])) 242 | elif net_style == "AdaptiveMaxPool3d": 243 | self.layers.append(nn.AdaptiveMaxPool3d(args_values_list[0], 244 | args_values_list[1])) 245 | elif net_style == "AdaptiveAvgPool1d": 246 | self.layers.append(nn.AdaptiveAvgPool1d(args_values_list[0])) 247 | elif net_style == "AdaptiveAvgPool2d": 248 | self.layers.append(nn.AdaptiveAvgPool2d(args_values_list[0])) 249 | elif net_style == "AdaptiveAvgPool3d": 250 | self.layers.append(nn.AdaptiveAvgPool3d(args_values_list[0])) 251 | elif net_style == "ReflectionPad1d": 252 | self.layers.append(nn.ReflectionPad1d(args_values_list[0])) 253 | elif net_style == "ReflectionPad2d": 254 | self.layers.append(nn.ReflectionPad2d(args_values_list[0])) 255 | elif net_style == "ReplicationPad1d": 256 | self.layers.append(nn.ReplicationPad1d(args_values_list[0])) 257 | elif net_style == "ReplicationPad2d": 258 | self.layers.append(nn.ReplicationPad2d(args_values_list[0])) 259 | elif net_style == "ReplicationPad3d": 260 | self.layers.append(nn.ReplicationPad3d(args_values_list[0])) 261 | elif net_style == "ZeroPad2d": 262 | self.layers.append(nn.ZeroPad2d(args_values_list[0])) 263 | elif net_style == "ConstantPad1d": 264 | self.layers.append(nn.ConstantPad1d(args_values_list[0], 265 | args_values_list[1])) 266 | elif net_style == "ConstantPad2d": 267 | self.layers.append(nn.ConstantPad2d(args_values_list[0], 268 | args_values_list[1])) 269 | elif net_style == "ConstantPad3d": 270 | self.layers.append(nn.ConstantPad3d(args_values_list[0], 271 | args_values_list[1])) 272 | elif net_style == "ELU": 273 | self.layers.append(nn.ELU(args_values_list[0], 274 | args_values_list[1])) 275 | elif net_style == "Hardshrink": 276 | self.layers.append(nn.Hardshrink(args_values_list[0])) 277 | elif net_style == "Hardtanh": 278 | self.layers.append(nn.Hardtanh(args_values_list[0], 279 | args_values_list[1], 280 | args_values_list[2], 281 | args_values_list[3], 282 | args_values_list[4])) 283 | elif net_style == "LeakyReLU": 284 | self.layers.append(nn.LeakyReLU(args_values_list[0], 285 | args_values_list[1])) 286 | elif net_style == "LogSigmoid": 287 | self.layers.append(nn.LogSigmoid()) 288 | elif net_style == "PReLU": 289 | self.layers.append(nn.PReLU(args_values_list[0], 290 | args_values_list[1])) 291 | elif net_style == "ReLU": 292 | self.layers.append(nn.ReLU(args_values_list[0])) 293 | elif net_style == "ReLU6": 294 | self.layers.append(nn.ReLU6(args_values_list[0])) 295 | elif net_style == "RReLU": 296 | self.layers.append(nn.RReLU(args_values_list[0], 297 | args_values_list[1], 298 | args_values_list[2])) 299 | elif net_style == "SELU": 300 | self.layers.append(nn.SELU(args_values_list[0])) 301 | elif net_style == "CELU": 302 | self.layers.append(nn.CELU(args_values_list[0], 303 | args_values_list[1])) 304 | elif net_style == "Sigmoid": 305 | self.layers.append(nn.Sigmoid()) 306 | elif net_style == "Softplus": 307 | self.layers.append(nn.Softplus(args_values_list[0], 308 | args_values_list[1])) 309 | elif net_style == "Softshrink": 310 | self.layers.append(nn.Softshrink(args_values_list[0])) 311 | elif net_style == "Softsign": 312 | self.layers.append(nn.Softsign()) 313 | elif net_style == "Tanh": 314 | self.layers.append(nn.Tanh()) 315 | elif net_style == "Tanhshrink": 316 | self.layers.append(nn.Tanhshrink()) 317 | elif net_style == "Threshold": 318 | self.layers.append(nn.Threshold(args_values_list[0], 319 | args_values_list[1], 320 | args_values_list[2])) 321 | elif net_style == "Softmin": 322 | self.layers.append(nn.Softmin(args_values_list[0])) 323 | elif net_style == "Softmax": 324 | self.layers.append(nn.Softmax(args_values_list[0])) 325 | elif net_style == "Softmax2d": 326 | self.layers.append(nn.Softmax2d()) 327 | elif net_style == "LogSoftmax": 328 | self.layers.append(nn.LogSoftmax(args_values_list[0])) 329 | elif net_style == "AdaptiveLogSoftmaxWithLoss": 330 | self.layers.append(nn.AdaptiveLogSoftmaxWithLoss(args_values_list[0], 331 | args_values_list[1], 332 | args_values_list[2], 333 | args_values_list[3], 334 | args_values_list[4])) 335 | elif net_style == "BatchNorm1d": 336 | self.layers.append(nn.BatchNorm1d(args_values_list[0], 337 | args_values_list[1], 338 | args_values_list[2], 339 | args_values_list[3], 340 | args_values_list[4])) 341 | elif net_style == "BatchNorm2d": 342 | self.layers.append(nn.BatchNorm2d(args_values_list[0], 343 | args_values_list[1], 344 | args_values_list[2], 345 | args_values_list[3], 346 | args_values_list[4])) 347 | elif net_style == "BatchNorm3d": 348 | self.layers.append(nn.BatchNorm3d(args_values_list[0], 349 | args_values_list[1], 350 | args_values_list[2], 351 | args_values_list[3], 352 | args_values_list[4])) 353 | elif net_style == "GroupNorm": 354 | self.layers.append(nn.GroupNorm(args_values_list[0], 355 | args_values_list[1], 356 | args_values_list[2], 357 | args_values_list[3])) 358 | elif net_style == "InstanceNorm1d": 359 | self.layers.append(nn.InstanceNorm1d(args_values_list[0], 360 | args_values_list[1], 361 | args_values_list[2], 362 | args_values_list[3], 363 | args_values_list[4])) 364 | elif net_style == "InstanceNorm2d": 365 | self.layers.append(nn.InstanceNorm2d(args_values_list[0], 366 | args_values_list[1], 367 | args_values_list[2], 368 | args_values_list[3], 369 | args_values_list[4])) 370 | elif net_style == "InstanceNorm3d": 371 | self.layers.append(nn.InstanceNorm3d(args_values_list[0], 372 | args_values_list[1], 373 | args_values_list[2], 374 | args_values_list[3], 375 | args_values_list[4])) 376 | elif net_style == "LayerNorm": 377 | self.layers.append(nn.LayerNorm(args_values_list[0], 378 | args_values_list[1], 379 | args_values_list[2])) 380 | elif net_style == "LocalResponseNorm": 381 | self.layers.append(nn.LocalResponseNorm(args_values_list[0], 382 | args_values_list[1], 383 | args_values_list[2], 384 | args_values_list[3])) 385 | elif net_style == "Linear": 386 | self.layers.append(nn.Linear(args_values_list[0], 387 | args_values_list[1], 388 | args_values_list[2])) 389 | elif net_style == "Dropout": 390 | self.layers.append(nn.Dropout(args_values_list[0], 391 | args_values_list[1])) 392 | elif net_style == "Dropout2d": 393 | self.layers.append(nn.Dropout2d(args_values_list[0], 394 | args_values_list[1])) 395 | elif net_style == "Dropout3d": 396 | self.layers.append(nn.Dropout3d(args_values_list[0], 397 | args_values_list[1])) 398 | elif net_style == "AlphaDropout": 399 | self.layers.append(nn.AlphaDropout(args_values_list[0], 400 | args_values_list[1])) 401 | 402 | 403 | # 从 xml 中读取参数并覆盖默认参数 404 | def __check_args_from_xml(self, args_dict, layer_xml_node): 405 | args_keys = list(args_dict.keys()) 406 | for args_key in args_keys: 407 | args_value_xml_node_list = layer_xml_node.getElementsByTagName(args_key) 408 | if len(args_value_xml_node_list) > 0: 409 | args_value_string = args_value_xml_node_list[0].childNodes[0].data 410 | args_value = eval(args_value_string) 411 | args_dict[args_key] = args_value 412 | return args_dict 413 | 414 | 415 | # 网络的默认参数 416 | def __default_args_for_layer(self, net_style): 417 | args_dict = collections.OrderedDict() 418 | if net_style == "Conv2d": 419 | args_dict["in_channels"] = None 420 | args_dict["out_channels"] = None 421 | args_dict["kernel_size"] = None 422 | args_dict["stride"] = 1 423 | args_dict["padding"] = 0 424 | args_dict["dilation"] = 1 425 | args_dict["groups"] = 1 426 | args_dict["bias"] = True 427 | elif net_style == "MaxPool2d": 428 | args_dict["kernel_size"] = None 429 | args_dict["stride"] = None 430 | args_dict["padding"] = 0 431 | args_dict["dilation"] = 1 432 | args_dict["return_indices"] = False 433 | args_dict["ceil_mode"] = False 434 | elif net_style == "Linear": 435 | args_dict["in_features"] = None 436 | args_dict["out_features"] = None 437 | args_dict["bias"] = True 438 | elif net_style == "reshape": 439 | args_dict["dimensions"] = None 440 | elif net_style == "Conv1d": 441 | args_dict["in_channels"] = None 442 | args_dict["out_channels"] = None 443 | args_dict["kernel_size"] = None 444 | args_dict["stride"] = 1 445 | args_dict["padding"] = 0 446 | args_dict["dilation"] = 1 447 | args_dict["groups"] = 1 448 | args_dict["bias"] = True 449 | elif net_style == "Conv3d": 450 | args_dict["in_channels"] = None 451 | args_dict["out_channels"] = None 452 | args_dict["kernel_size"] = None 453 | args_dict["stride"] = 1 454 | args_dict["padding"] = 0 455 | args_dict["dilation"] = 1 456 | args_dict["groups"] = 1 457 | args_dict["bias"] = True 458 | elif net_style == "ConvTranspose1d": 459 | args_dict["in_channels"] = None 460 | args_dict["out_channels"] = None 461 | args_dict["kernel_size"] = None 462 | args_dict["stride"] = 1 463 | args_dict["padding"] = 0 464 | args_dict["output_padding"] = 0 465 | args_dict["groups"] = 1 466 | args_dict["bias"] = True 467 | args_dict["dilation"] = 1 468 | elif net_style == "ConvTranspose2d": 469 | args_dict["in_channels"] = None 470 | args_dict["out_channels"] = None 471 | args_dict["kernel_size"] = None 472 | args_dict["stride"] = 1 473 | args_dict["padding"] = 0 474 | args_dict["output_padding"] = 0 475 | args_dict["groups"] = 1 476 | args_dict["bias"] = True 477 | args_dict["dilation"] = 1 478 | elif net_style == "ConvTranspose3d": 479 | args_dict["in_channels"] = None 480 | args_dict["out_channels"] = None 481 | args_dict["kernel_size"] = None 482 | args_dict["stride"] = 1 483 | args_dict["padding"] = 0 484 | args_dict["output_padding"] = 0 485 | args_dict["groups"] = 1 486 | args_dict["bias"] = True 487 | args_dict["dilation"] = 1 488 | elif net_style == "Unfold": 489 | args_dict["kernel_size"] = None 490 | args_dict["dilation"] = 1 491 | args_dict["padding"] = 0 492 | args_dict["stride"] = 1 493 | elif net_style == "Fold": 494 | args_dict["output_size"] = None 495 | args_dict["kernel_size"] = None 496 | args_dict["dilation"] = 1 497 | args_dict["padding"] = 0 498 | args_dict["stride"] = 1 499 | elif net_style == "MaxPool1d": 500 | args_dict["kernel_size"] = None 501 | args_dict["stride"] = None 502 | args_dict["padding"] = 0 503 | args_dict["dilation"] = 1 504 | args_dict["return_indices"] = False 505 | args_dict["ceil_mode"] = False 506 | elif net_style == "MaxPool3d": 507 | args_dict["kernel_size"] = None 508 | args_dict["stride"] = None 509 | args_dict["padding"] = 0 510 | args_dict["dilation"] = 1 511 | args_dict["return_indices"] = False 512 | args_dict["ceil_mode"] = False 513 | elif net_style == "MaxUnpool1d": 514 | args_dict["kernel_size"] = None 515 | args_dict["stride"] = None 516 | args_dict["padding"] = 0 517 | elif net_style == "MaxUnpool2d": 518 | args_dict["kernel_size"] = None 519 | args_dict["stride"] = None 520 | args_dict["padding"] = 0 521 | elif net_style == "MaxUnpool3d": 522 | args_dict["kernel_size"] = None 523 | args_dict["stride"] = None 524 | args_dict["padding"] = 0 525 | elif net_style == "AvgPool1d": 526 | args_dict["kernel_size"] = None 527 | args_dict["stride"] = None 528 | args_dict["padding"] = 0 529 | args_dict["ceil_mode"] = False 530 | args_dict["count_include_pad"] = True 531 | elif net_style == "AvgPool2d": 532 | args_dict["kernel_size"] = None 533 | args_dict["stride"] = None 534 | args_dict["padding"] = 0 535 | args_dict["ceil_mode"] = False 536 | args_dict["count_include_pad"] = True 537 | elif net_style == "AvgPool3d": 538 | args_dict["kernel_size"] = None 539 | args_dict["stride"] = None 540 | args_dict["padding"] = 0 541 | args_dict["ceil_mode"] = False 542 | args_dict["count_include_pad"] = True 543 | elif net_style == "FractionalMaxPool2d": 544 | args_dict["kernel_size"] = None 545 | args_dict["output_size"] = None 546 | args_dict["output_ratio"] = None 547 | args_dict["return_indices"] = False 548 | args_dict["_random_samples"] = None 549 | elif net_style == "LPPool1d": 550 | args_dict["norm_type"] = None 551 | args_dict["kernel_size"] = None 552 | args_dict["stride"] = None 553 | args_dict["ceil_mode"] = False 554 | elif net_style == "LPPool2d": 555 | args_dict["norm_type"] = None 556 | args_dict["kernel_size"] = None 557 | args_dict["stride"] = None 558 | args_dict["ceil_mode"] = False 559 | elif net_style == "AdaptiveMaxPool1d": 560 | args_dict["output_size"] = None 561 | args_dict["return_indices"] = False 562 | elif net_style == "AdaptiveMaxPool2d": 563 | args_dict["output_size"] = None 564 | args_dict["return_indices"] = False 565 | elif net_style == "AdaptiveMaxPool3d": 566 | args_dict["output_size"] = None 567 | args_dict["return_indices"] = False 568 | elif net_style == "AdaptiveAvgPool1d": 569 | args_dict["output_size"] = None 570 | elif net_style == "AdaptiveAvgPool2d": 571 | args_dict["output_size"] = None 572 | elif net_style == "AdaptiveAvgPool3d": 573 | args_dict["output_size"] = None 574 | elif net_style == "ReflectionPad1d": 575 | args_dict["padding"] = None 576 | elif net_style == "ReflectionPad2d": 577 | args_dict["padding"] = None 578 | elif net_style == "ReplicationPad1d": 579 | args_dict["padding"] = None 580 | elif net_style == "ReplicationPad2d": 581 | args_dict["padding"] = None 582 | elif net_style == "ReplicationPad3d": 583 | args_dict["padding"] = None 584 | elif net_style == "ZeroPad2d": 585 | args_dict["padding"] = None 586 | elif net_style == "ConstantPad1d": 587 | args_dict["padding"] = None 588 | args_dict["value"] = None 589 | elif net_style == "ConstantPad2d": 590 | args_dict["padding"] = None 591 | args_dict["value"] = None 592 | elif net_style == "ConstantPad3d": 593 | args_dict["padding"] = None 594 | args_dict["value"] = None 595 | elif net_style == "ELU": 596 | args_dict["alpha"] = 1.0 597 | args_dict["inplace"] = False 598 | elif net_style == "Hardshrink": 599 | args_dict["lambd"] = 0.5 600 | elif net_style == "Hardtanh": 601 | args_dict["min_val"] = -1.0 602 | args_dict["max_val"] = 1.0 603 | args_dict["inplace"] = False 604 | args_dict["min_value"] = None 605 | args_dict["max_value"] = None 606 | elif net_style == "LeakyReLU": 607 | args_dict["negative_slope"] = 0.01 608 | args_dict["inplace"] = False 609 | elif net_style == "LogSigmoid": 610 | args_dict = None 611 | elif net_style == "PReLU": 612 | args_dict["num_parameters"] = 1 613 | args_dict["init"] = 0.25 614 | elif net_style == "ReLU": 615 | args_dict["inplace"] = False 616 | elif net_style == "ReLU6": 617 | args_dict["inplace"] = False 618 | elif net_style == "RReLU": 619 | args_dict["lower"] = 0.125 620 | args_dict["upper"] = 1.0 / 3.0 621 | args_dict["inplace"] = False 622 | elif net_style == "SELU": 623 | args_dict["inplace"] = False 624 | elif net_style == "CELU": 625 | args_dict["alpha"] = 1.0 626 | args_dict["inplace"] = False 627 | elif net_style == "Sigmoid": 628 | args_dict = None 629 | elif net_style == "Softplus": 630 | args_dict["beta"] = 1 631 | args_dict["threshold"] = 20 632 | elif net_style == "Softshrink": 633 | args_dict["lambd"] = 0.5 634 | elif net_style == "Softsign": 635 | args_dict = None 636 | elif net_style == "Tanh": 637 | args_dict = None 638 | elif net_style == "Tanhshrink": 639 | args_dict = None 640 | elif net_style == "Threshold": 641 | args_dict["threshold"] = None 642 | args_dict["value"] = None 643 | args_dict["inplace"] = False 644 | elif net_style == "Softmin": 645 | args_dict["dim"] = None 646 | elif net_style == "Softmax": 647 | args_dict["dim"] = None 648 | elif net_style == "Softmax2d": 649 | args_dict = None 650 | elif net_style == "LogSoftmax": 651 | args_dict["dim"] = None 652 | elif net_style == "AdaptiveLogSoftmaxWithLoss": 653 | args_dict["in_features"] = None 654 | args_dict["n_classes"] = None 655 | args_dict["cutoffs"] = None 656 | args_dict["div_value"] = 4.0 657 | args_dict["head_bias"] = False 658 | elif net_style == "BatchNorm1d": 659 | args_dict["num_features"] = None 660 | args_dict["eps"] = 1e-05 661 | args_dict["momentum"] = 0.1 662 | args_dict["affine"] = True 663 | args_dict["track_running_stats"] = True 664 | elif net_style == "BatchNorm2d": 665 | args_dict["num_features"] = None 666 | args_dict["eps"] = 1e-05 667 | args_dict["momentum"] = 0.1 668 | args_dict["affine"] = True 669 | args_dict["track_running_stats"] = True 670 | elif net_style == "BatchNorm3d": 671 | args_dict["num_features"] = None 672 | args_dict["eps"] = 1e-05 673 | args_dict["momentum"] = 0.1 674 | args_dict["affine"] = True 675 | args_dict["track_running_stats"] = True 676 | elif net_style == "GroupNorm": 677 | args_dict["num_groups"] = None 678 | args_dict["num_channels"] = None 679 | args_dict["eps"] = 1e-05 680 | args_dict["affine"] = True 681 | elif net_style == "InstanceNorm1d": 682 | args_dict["num_features"] = None 683 | args_dict["eps"] = 1e-05 684 | args_dict["momentum"] = 0.1 685 | args_dict["affine"] = True 686 | args_dict["track_running_stats"] = False 687 | elif net_style == "InstanceNorm2d": 688 | args_dict["num_features"] = None 689 | args_dict["eps"] = 1e-05 690 | args_dict["momentum"] = 0.1 691 | args_dict["affine"] = True 692 | args_dict["track_running_stats"] = False 693 | elif net_style == "InstanceNorm3d": 694 | args_dict["num_features"] = None 695 | args_dict["eps"] = 1e-05 696 | args_dict["momentum"] = 0.1 697 | args_dict["affine"] = True 698 | args_dict["track_running_stats"] = False 699 | elif net_style == "LayerNorm": 700 | args_dict["normalized_shape"] = None 701 | args_dict["eps"] = 1e-05 702 | args_dict["elementwise_affine"] = True 703 | elif net_style == "LocalResponseNorm": 704 | args_dict["size"] = None 705 | args_dict["alpha"] = 0.0001 706 | args_dict["beta"] = 0.75 707 | args_dict["k"] = 1.0 708 | elif net_style == "Linear": 709 | args_dict["in_features"] = None 710 | args_dict["out_features"] = None 711 | args_dict["bias"] = True 712 | elif net_style == "Dropout": 713 | args_dict["p"] = 0.5 714 | args_dict["inplace"] = False 715 | elif net_style == "Dropout2d": 716 | args_dict["p"] = 0.5 717 | args_dict["inplace"] = False 718 | elif net_style == "Dropout3d": 719 | args_dict["p"] = 0.5 720 | args_dict["inplace"] = False 721 | elif net_style == "AlphaDropout": 722 | args_dict["p"] = 0.5 723 | args_dict["inplace"] = False 724 | # elif net_style == "RNNCell": 725 | # args_dict["input_size"] = None 726 | # args_dict["hidden_size"] = None 727 | # args_dict["bias"] = True 728 | # args_dict["nonlinearity"] = 'tanh' 729 | 730 | return args_dict 731 | 732 | def forward(self, x): 733 | for i in range(len(self.layers)): 734 | # 是.view，转换 Tensor 的形状 735 | if isinstance(self.layers[i], list): 736 | # print(x.shape) 737 | # print(self.layers[i]) 738 | x = x.view(self.layers[i]) 739 | # 是一个隐藏层 740 | else: 741 | # print(x.shape) 742 | x = self.layers[i](x) 743 | return x 744 | 745 | 746 | 747 | class Net2(nn.Module): 748 | def __init__(self): 749 | super(Net2, self).__init__() 750 | self.conv1 = nn.Conv2d(3, 6, 5) 751 | self.pool = nn.MaxPool2d(2, 2) 752 | self.conv2 = nn.Conv2d(6, 16, 5) 753 | self.fc1 = nn.Linear(16 * 5 * 5, 120) 754 | self.fc2 = nn.Linear(120, 84) 755 | self.fc3 = nn.Linear(84, 10) 756 | 757 | def forward(self, x): 758 | x = self.pool(F.relu(self.conv1(x))) 759 | print(x.shape) 760 | x = self.pool(F.relu(self.conv2(x))) 761 | print(x.shape) 762 | x = x.view([-1, 16 * 5 * 5]) 763 | print(x.shape) 764 | x = F.relu(self.fc1(x)) 765 | print(x.shape) 766 | x = F.relu(self.fc2(x)) 767 | print(x.shape) 768 | x = self.fc3(x) 769 | return x 770 | 771 | if __name__ == "__main__": 772 | main() -------------------------------------------------------------------------------- /xml2pytorch/xml_examples/simpleCNN.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | Conv2d 5 | 3 6 | 6 7 | 5 8 | 9 | 10 | ELU 11 | 12 | 13 | MaxPool2d 14 | 2 15 | 2 16 | logsigmoid 17 | 18 | 19 | Conv2d 20 | 6 21 | 16 22 | 5 23 | relu 24 | 25 | 26 | MaxPool2d 27 | 2 28 | 2 29 | relu 30 | 31 | 32 | reshape 33 | [-1, 16*5*5] 34 | 35 | 36 | Linear 37 | 400 38 | 120 39 | tanh 40 | 41 | 42 | Linear 43 | 120 44 | 84 45 | sigmoid 46 | 47 | 48 | Linear 49 | 84 50 | 10 51 | softmax 52 | 53 | 54 | --------------------------------------------------------------------------------