本文针对cifar10 图集进行了DCGAN的复现。
其中库中的SpectralNormalizationKeras需添加至python环境中 该篇代码如下:
1 from keras import backend as K 2 from keras.engine import * 3 from keras.legacy import interfaces 4 from keras import activations 5 from keras import initializers 6 from keras import regularizers 7 from keras import constraints 8 from keras.utils.generic_utils import func_dump 9 from keras.utils.generic_utils import func_load 10 from keras.utils.generic_utils import deserialize_keras_object 11 from keras.utils.generic_utils import has_arg 12 from keras.utils import conv_utils 13 from keras.legacy import interfaces 14 from keras.layers import Dense, Conv1D, Conv2D, Conv3D, Conv2DTranspose, Embedding 15 import tensorflow as tf 16 17 class DenseSN(Dense): 18 def build(self, input_shape): 19 assert len(input_shape) >= 2 20 input_dim = input_shape[-1] 21 self.kernel = self.add_weight(shape=(input_dim, self.units), 22 initializer=self.kernel_initializer, 23 name='kernel', 24 regularizer=self.kernel_regularizer, 25 constraint=self.kernel_constraint) 26 if self.use_bias: 27 self.bias = self.add_weight(shape=(self.units,), 28 initializer=self.bias_initializer, 29 name='bias', 30 regularizer=self.bias_regularizer, 31 constraint=self.bias_constraint) 32 else: 33 self.bias = None 34 self.u = self.add_weight(shape=tuple([1, self.kernel.shape.as_list()[-1]]), 35 initializer=initializers.RandomNormal(0, 1), 36 name='sn', 37 trainable=False) 38 self.input_spec = InputSpec(min_ndim=2, axes={-1: input_dim}) 39 self.built = True 40 41 def call(self, inputs, training=None): 42 def _l2normalize(v, eps=1e-12): 43 return v / (K.sum(v ** 2) ** 0.5 + eps) 44 def power_iteration(W, u): 45 _u = u 46 _v = _l2normalize(K.dot(_u, K.transpose(W))) 47 _u = _l2normalize(K.dot(_v, W)) 48 return _u, _v 49 W_shape = self.kernel.shape.as_list() 50 #Flatten the Tensor 51 W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]]) 52 _u, _v = power_iteration(W_reshaped, self.u) 53 #Calculate Sigma 54 sigma=K.dot(_v, W_reshaped) 55 sigma=K.dot(sigma, K.transpose(_u)) 56 #normalize it 57 W_bar = W_reshaped / sigma 58 #reshape weight tensor 59 if training in {0, False}: 60 W_bar = K.reshape(W_bar, W_shape) 61 else: 62 with tf.control_dependencies([self.u.assign(_u)]): 63 W_bar = K.reshape(W_bar, W_shape) 64 output = K.dot(inputs, W_bar) 65 if self.use_bias: 66 output = K.bias_add(output, self.bias, data_format='channels_last') 67 if self.activation is not None: 68 output = self.activation(output) 69 return output 70 71 class _ConvSN(Layer): 72 73 def __init__(self, rank, 74 filters, 75 kernel_size, 76 strides=1, 77 padding='valid', 78 data_format=None, 79 dilation_rate=1, 80 activation=None, 81 use_bias=True, 82 kernel_initializer='glorot_uniform', 83 bias_initializer='zeros', 84 kernel_regularizer=None, 85 bias_regularizer=None, 86 activity_regularizer=None, 87 kernel_constraint=None, 88 bias_constraint=None, 89 spectral_normalization=True, 90 **kwargs): 91 super(_ConvSN, self).__init__(**kwargs) 92 self.rank = rank 93 self.filters = filters 94 self.kernel_size = conv_utils.normalize_tuple(kernel_size, rank, 'kernel_size') 95 self.strides = conv_utils.normalize_tuple(strides, rank, 'strides') 96 self.padding = conv_utils.normalize_padding(padding) 97 self.data_format = conv_utils.normalize_data_format(data_format) 98 self.dilation_rate = conv_utils.normalize_tuple(dilation_rate, rank, 'dilation_rate') 99 self.activation = activations.get(activation) 100 self.use_bias = use_bias 101 self.kernel_initializer = initializers.get(kernel_initializer) 102 self.bias_initializer = initializers.get(bias_initializer) 103 self.kernel_regularizer = regularizers.get(kernel_regularizer) 104 self.bias_regularizer = regularizers.get(bias_regularizer) 105 self.activity_regularizer = regularizers.get(activity_regularizer) 106 self.kernel_constraint = constraints.get(kernel_constraint) 107 self.bias_constraint = constraints.get(bias_constraint) 108 self.input_spec = InputSpec(ndim=self.rank + 2) 109 self.spectral_normalization = spectral_normalization 110 self.u = None 111 112 def _l2normalize(self, v, eps=1e-12): 113 return v / (K.sum(v ** 2) ** 0.5 + eps) 114 115 def power_iteration(self, u, W): 116 ''' 117 Accroding the paper, we only need to do power iteration one time. 118 ''' 119 v = self._l2normalize(K.dot(u, K.transpose(W))) 120 u = self._l2normalize(K.dot(v, W)) 121 return u, v 122 def build(self, input_shape): 123 if self.data_format == 'channels_first': 124 channel_axis = 1 125 else: 126 channel_axis = -1 127 if input_shape[channel_axis] is None: 128 raise ValueError('The channel dimension of the inputs ' 129 'should be defined. Found `None`.') 130 input_dim = input_shape[channel_axis] 131 kernel_shape = self.kernel_size + (input_dim, self.filters) 132 133 self.kernel = self.add_weight(shape=kernel_shape, 134 initializer=self.kernel_initializer, 135 name='kernel', 136 regularizer=self.kernel_regularizer, 137 constraint=self.kernel_constraint) 138 139 #Spectral Normalization 140 if self.spectral_normalization: 141 self.u = self.add_weight(shape=tuple([1, self.kernel.shape.as_list()[-1]]), 142 initializer=initializers.RandomNormal(0, 1), 143 name='sn', 144 trainable=False) 145 146 if self.use_bias: 147 self.bias = self.add_weight(shape=(self.filters,), 148 initializer=self.bias_initializer, 149 name='bias', 150 regularizer=self.bias_regularizer, 151 constraint=self.bias_constraint) 152 else: 153 self.bias = None 154 # Set input spec. 155 self.input_spec = InputSpec(ndim=self.rank + 2, 156 axes={channel_axis: input_dim}) 157 self.built = True 158 159 def call(self, inputs): 160 def _l2normalize(v, eps=1e-12): 161 return v / (K.sum(v ** 2) ** 0.5 + eps) 162 def power_iteration(W, u): 163 _u = u 164 _v = _l2normalize(K.dot(_u, K.transpose(W))) 165 _u = _l2normalize(K.dot(_v, W)) 166 return _u, _v 167 168 if self.spectral_normalization: 169 W_shape = self.kernel.shape.as_list() 170 #Flatten the Tensor 171 W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]]) 172 _u, _v = power_iteration(W_reshaped, self.u) 173 #Calculate Sigma 174 sigma=K.dot(_v, W_reshaped) 175 sigma=K.dot(sigma, K.transpose(_u)) 176 #normalize it 177 W_bar = W_reshaped / sigma 178 #reshape weight tensor 179 if training in {0, False}: 180 W_bar = K.reshape(W_bar, W_shape) 181 else: 182 with tf.control_dependencies([self.u.assign(_u)]): 183 W_bar = K.reshape(W_bar, W_shape) 184 185 #update weitht 186 self.kernel = W_bar 187 188 if self.rank == 1: 189 outputs = K.conv1d( 190 inputs, 191 self.kernel, 192 strides=self.strides[0], 193 padding=self.padding, 194 data_format=self.data_format, 195 dilation_rate=self.dilation_rate[0]) 196 if self.rank == 2: 197 outputs = K.conv2d( 198 inputs, 199 self.kernel, 200 strides=self.strides, 201 padding=self.padding, 202 data_format=self.data_format, 203 dilation_rate=self.dilation_rate) 204 if self.rank == 3: 205 outputs = K.conv3d( 206 inputs, 207 self.kernel, 208 strides=self.strides, 209 padding=self.padding, 210 data_format=self.data_format, 211 dilation_rate=self.dilation_rate) 212 213 if self.use_bias: 214 outputs = K.bias_add( 215 outputs, 216 self.bias, 217 data_format=self.data_format) 218 219 if self.activation is not None: 220 return self.activation(outputs) 221 return outputs 222 223 def compute_output_shape(self, input_shape): 224 if self.data_format == 'channels_last': 225 space = input_shape[1:-1] 226 new_space = [] 227 for i in range(len(space)): 228 new_dim = conv_utils.conv_output_length( 229 space[i], 230 self.kernel_size[i], 231 padding=self.padding, 232 stride=self.strides[i], 233 dilation=self.dilation_rate[i]) 234 new_space.append(new_dim) 235 return (input_shape[0],) + tuple(new_space) + (self.filters,) 236 if self.data_format == 'channels_first': 237 space = input_shape[2:] 238 new_space = [] 239 for i in range(len(space)): 240 new_dim = conv_utils.conv_output_length( 241 space[i], 242 self.kernel_size[i], 243 padding=self.padding, 244 stride=self.strides[i], 245 dilation=self.dilation_rate[i]) 246 new_space.append(new_dim) 247 return (input_shape[0], self.filters) + tuple(new_space) 248 249 def get_config(self): 250 config = { 251 'rank': self.rank, 252 'filters': self.filters, 253 'kernel_size': self.kernel_size, 254 'strides': self.strides, 255 'padding': self.padding, 256 'data_format': self.data_format, 257 'dilation_rate': self.dilation_rate, 258 'activation': activations.serialize(self.activation), 259 'use_bias': self.use_bias, 260 'kernel_initializer': initializers.serialize(self.kernel_initializer), 261 'bias_initializer': initializers.serialize(self.bias_initializer), 262 'kernel_regularizer': regularizers.serialize(self.kernel_regularizer), 263 'bias_regularizer': regularizers.serialize(self.bias_regularizer), 264 'activity_regularizer': regularizers.serialize(self.activity_regularizer), 265 'kernel_constraint': constraints.serialize(self.kernel_constraint), 266 'bias_constraint': constraints.serialize(self.bias_constraint) 267 } 268 base_config = super(_Conv, self).get_config() 269 return dict(list(base_config.items()) + list(config.items())) 270 271 class ConvSN2D(Conv2D): 272 273 def build(self, input_shape): 274 if self.data_format == 'channels_first': 275 channel_axis = 1 276 else: 277 channel_axis = -1 278 if input_shape[channel_axis] is None: 279 raise ValueError('The channel dimension of the inputs ' 280 'should be defined. Found `None`.') 281 input_dim = input_shape[channel_axis] 282 kernel_shape = self.kernel_size + (input_dim, self.filters) 283 284 self.kernel = self.add_weight(shape=kernel_shape, 285 initializer=self.kernel_initializer, 286 name='kernel', 287 regularizer=self.kernel_regularizer, 288 constraint=self.kernel_constraint) 289 290 if self.use_bias: 291 self.bias = self.add_weight(shape=(self.filters,), 292 initializer=self.bias_initializer, 293 name='bias', 294 regularizer=self.bias_regularizer, 295 constraint=self.bias_constraint) 296 else: 297 self.bias = None 298 299 self.u = self.add_weight(shape=tuple([1, self.kernel.shape.as_list()[-1]]), 300 initializer=initializers.RandomNormal(0, 1), 301 name='sn', 302 trainable=False) 303 304 # Set input spec. 305 self.input_spec = InputSpec(ndim=self.rank + 2, 306 axes={channel_axis: input_dim}) 307 self.built = True 308 def call(self, inputs, training=None): 309 def _l2normalize(v, eps=1e-12): 310 return v / (K.sum(v ** 2) ** 0.5 + eps) 311 def power_iteration(W, u): 312 #Accroding the paper, we only need to do power iteration one time. 313 _u = u 314 _v = _l2normalize(K.dot(_u, K.transpose(W))) 315 _u = _l2normalize(K.dot(_v, W)) 316 return _u, _v 317 #Spectral Normalization 318 W_shape = self.kernel.shape.as_list() 319 #Flatten the Tensor 320 W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]]) 321 _u, _v = power_iteration(W_reshaped, self.u) 322 #Calculate Sigma 323 sigma=K.dot(_v, W_reshaped) 324 sigma=K.dot(sigma, K.transpose(_u)) 325 #normalize it 326 W_bar = W_reshaped / sigma 327 #reshape weight tensor 328 if training in {0, False}: 329 W_bar = K.reshape(W_bar, W_shape) 330 else: 331 with tf.control_dependencies([self.u.assign(_u)]): 332 W_bar = K.reshape(W_bar, W_shape) 333 334 outputs = K.conv2d( 335 inputs, 336 W_bar, 337 strides=self.strides, 338 padding=self.padding, 339 data_format=self.data_format, 340 dilation_rate=self.dilation_rate) 341 if self.use_bias: 342 outputs = K.bias_add( 343 outputs, 344 self.bias, 345 data_format=self.data_format) 346 if self.activation is not None: 347 return self.activation(outputs) 348 return outputs 349 350 class ConvSN1D(Conv1D): 351 352 def build(self, input_shape): 353 if self.data_format == 'channels_first': 354 channel_axis = 1 355 else: 356 channel_axis = -1 357 if input_shape[channel_axis] is None: 358 raise ValueError('The channel dimension of the inputs ' 359 'should be defined. Found `None`.') 360 input_dim = input_shape[channel_axis] 361 kernel_shape = self.kernel_size + (input_dim, self.filters) 362 363 self.kernel = self.add_weight(shape=kernel_shape, 364 initializer=self.kernel_initializer, 365 name='kernel', 366 regularizer=self.kernel_regularizer, 367 constraint=self.kernel_constraint) 368 369 if self.use_bias: 370 self.bias = self.add_weight(shape=(self.filters,), 371 initializer=self.bias_initializer, 372 name='bias', 373 regularizer=self.bias_regularizer, 374 constraint=self.bias_constraint) 375 else: 376 self.bias = None 377 378 self.u = self.add_weight(shape=tuple([1, self.kernel.shape.as_list()[-1]]), 379 initializer=initializers.RandomNormal(0, 1), 380 name='sn', 381 trainable=False) 382 # Set input spec. 383 self.input_spec = InputSpec(ndim=self.rank + 2, 384 axes={channel_axis: input_dim}) 385 self.built = True 386 387 def call(self, inputs, training=None): 388 def _l2normalize(v, eps=1e-12): 389 return v / (K.sum(v ** 2) ** 0.5 + eps) 390 def power_iteration(W, u): 391 #Accroding the paper, we only need to do power iteration one time. 392 _u = u 393 _v = _l2normalize(K.dot(_u, K.transpose(W))) 394 _u = _l2normalize(K.dot(_v, W)) 395 return _u, _v 396 #Spectral Normalization 397 W_shape = self.kernel.shape.as_list() 398 #Flatten the Tensor 399 W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]]) 400 _u, _v = power_iteration(W_reshaped, self.u) 401 #Calculate Sigma 402 sigma=K.dot(_v, W_reshaped) 403 sigma=K.dot(sigma, K.transpose(_u)) 404 #normalize it 405 W_bar = W_reshaped / sigma 406 #reshape weight tensor 407 if training in {0, False}: 408 W_bar = K.reshape(W_bar, W_shape) 409 else: 410 with tf.control_dependencies([self.u.assign(_u)]): 411 W_bar = K.reshape(W_bar, W_shape) 412 413 outputs = K.conv1d( 414 inputs, 415 W_bar, 416 strides=self.strides, 417 padding=self.padding, 418 data_format=self.data_format, 419 dilation_rate=self.dilation_rate) 420 if self.use_bias: 421 outputs = K.bias_add( 422 outputs, 423 self.bias, 424 data_format=self.data_format) 425 if self.activation is not None: 426 return self.activation(outputs) 427 return outputs 428 429 class ConvSN3D(Conv3D): 430 def build(self, input_shape): 431 if self.data_format == 'channels_first': 432 channel_axis = 1 433 else: 434 channel_axis = -1 435 if input_shape[channel_axis] is None: 436 raise ValueError('The channel dimension of the inputs ' 437 'should be defined. Found `None`.') 438 input_dim = input_shape[channel_axis] 439 kernel_shape = self.kernel_size + (input_dim, self.filters) 440 441 self.kernel = self.add_weight(shape=kernel_shape, 442 initializer=self.kernel_initializer, 443 name='kernel', 444 regularizer=self.kernel_regularizer, 445 constraint=self.kernel_constraint) 446 447 self.u = self.add_weight(shape=tuple([1, self.kernel.shape.as_list()[-1]]), 448 initializer=initializers.RandomNormal(0, 1), 449 name='sn', 450 trainable=False) 451 452 if self.use_bias: 453 self.bias = self.add_weight(shape=(self.filters,), 454 initializer=self.bias_initializer, 455 name='bias', 456 regularizer=self.bias_regularizer, 457 constraint=self.bias_constraint) 458 else: 459 self.bias = None 460 # Set input spec. 461 self.input_spec = InputSpec(ndim=self.rank + 2, 462 axes={channel_axis: input_dim}) 463 self.built = True 464 465 def call(self, inputs, training=None): 466 def _l2normalize(v, eps=1e-12): 467 return v / (K.sum(v ** 2) ** 0.5 + eps) 468 def power_iteration(W, u): 469 #Accroding the paper, we only need to do power iteration one time. 470 _u = u 471 _v = _l2normalize(K.dot(_u, K.transpose(W))) 472 _u = _l2normalize(K.dot(_v, W)) 473 return _u, _v 474 #Spectral Normalization 475 W_shape = self.kernel.shape.as_list() 476 #Flatten the Tensor 477 W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]]) 478 _u, _v = power_iteration(W_reshaped, self.u) 479 #Calculate Sigma 480 sigma=K.dot(_v, W_reshaped) 481 sigma=K.dot(sigma, K.transpose(_u)) 482 #normalize it 483 W_bar = W_reshaped / sigma 484 #reshape weight tensor 485 if training in {0, False}: 486 W_bar = K.reshape(W_bar, W_shape) 487 else: 488 with tf.control_dependencies([self.u.assign(_u)]): 489 W_bar = K.reshape(W_bar, W_shape) 490 491 outputs = K.conv3d( 492 inputs, 493 W_bar, 494 strides=self.strides, 495 padding=self.padding, 496 data_format=self.data_format, 497 dilation_rate=self.dilation_rate) 498 if self.use_bias: 499 outputs = K.bias_add( 500 outputs, 501 self.bias, 502 data_format=self.data_format) 503 if self.activation is not None: 504 return self.activation(outputs) 505 return outputs 506 507 508 class EmbeddingSN(Embedding): 509 510 def build(self, input_shape): 511 self.embeddings = self.add_weight( 512 shape=(self.input_dim, self.output_dim), 513 initializer=self.embeddings_initializer, 514 name='embeddings', 515 regularizer=self.embeddings_regularizer, 516 constraint=self.embeddings_constraint, 517 dtype=self.dtype) 518 519 self.u = self.add_weight(shape=tuple([1, self.embeddings.shape.as_list()[-1]]), 520 initializer=initializers.RandomNormal(0, 1), 521 name='sn', 522 trainable=False) 523 524 self.built = True 525 526 def call(self, inputs): 527 if K.dtype(inputs) != 'int32': 528 inputs = K.cast(inputs, 'int32') 529 530 def _l2normalize(v, eps=1e-12): 531 return v / (K.sum(v ** 2) ** 0.5 + eps) 532 def power_iteration(W, u): 533 #Accroding the paper, we only need to do power iteration one time. 534 _u = u 535 _v = _l2normalize(K.dot(_u, K.transpose(W))) 536 _u = _l2normalize(K.dot(_v, W)) 537 return _u, _v 538 W_shape = self.embeddings.shape.as_list() 539 #Flatten the Tensor 540 W_reshaped = K.reshape(self.embeddings, [-1, W_shape[-1]]) 541 _u, _v = power_iteration(W_reshaped, self.u) 542 #Calculate Sigma 543 sigma=K.dot(_v, W_reshaped) 544 sigma=K.dot(sigma, K.transpose(_u)) 545 #normalize it 546 W_bar = W_reshaped / sigma 547 #reshape weight tensor 548 if training in {0, False}: 549 W_bar = K.reshape(W_bar, W_shape) 550 else: 551 with tf.control_dependencies([self.u.assign(_u)]): 552 W_bar = K.reshape(W_bar, W_shape) 553 self.embeddings = W_bar 554 555 out = K.gather(self.embeddings, inputs) 556 return out 557 558 class ConvSN2DTranspose(Conv2DTranspose): 559 560 def build(self, input_shape): 561 if len(input_shape) != 4: 562 raise ValueError('Inputs should have rank ' + 563 str(4) + 564 '; Received input shape:', str(input_shape)) 565 if self.data_format == 'channels_first': 566 channel_axis = 1 567 else: 568 channel_axis = -1 569 if input_shape[channel_axis] is None: 570 raise ValueError('The channel dimension of the inputs ' 571 'should be defined. Found `None`.') 572 input_dim = input_shape[channel_axis] 573 kernel_shape = self.kernel_size + (self.filters, input_dim) 574 575 self.kernel = self.add_weight(shape=kernel_shape, 576 initializer=self.kernel_initializer, 577 name='kernel', 578 regularizer=self.kernel_regularizer, 579 constraint=self.kernel_constraint) 580 if self.use_bias: 581 self.bias = self.add_weight(shape=(self.filters,), 582 initializer=self.bias_initializer, 583 name='bias', 584 regularizer=self.bias_regularizer, 585 constraint=self.bias_constraint) 586 else: 587 self.bias = None 588 589 self.u = self.add_weight(shape=tuple([1, self.kernel.shape.as_list()[-1]]), 590 initializer=initializers.RandomNormal(0, 1), 591 name='sn', 592 trainable=False) 593 594 # Set input spec. 595 self.input_spec = InputSpec(ndim=4, axes={channel_axis: input_dim}) 596 self.built = True 597 598 def call(self, inputs): 599 input_shape = K.shape(inputs) 600 batch_size = input_shape[0] 601 if self.data_format == 'channels_first': 602 h_axis, w_axis = 2, 3 603 else: 604 h_axis, w_axis = 1, 2 605 606 height, width = input_shape[h_axis], input_shape[w_axis] 607 kernel_h, kernel_w = self.kernel_size 608 stride_h, stride_w = self.strides 609 if self.output_padding is None: 610 out_pad_h = out_pad_w = None 611 else: 612 out_pad_h, out_pad_w = self.output_padding 613 614 # Infer the dynamic output shape: 615 out_height = conv_utils.deconv_length(height, 616 stride_h, kernel_h, 617 self.padding, 618 out_pad_h) 619 out_width = conv_utils.deconv_length(width, 620 stride_w, kernel_w, 621 self.padding, 622 out_pad_w) 623 if self.data_format == 'channels_first': 624 output_shape = (batch_size, self.filters, out_height, out_width) 625 else: 626 output_shape = (batch_size, out_height, out_width, self.filters) 627 628 #Spectral Normalization 629 def _l2normalize(v, eps=1e-12): 630 return v / (K.sum(v ** 2) ** 0.5 + eps) 631 def power_iteration(W, u): 632 #Accroding the paper, we only need to do power iteration one time. 633 _u = u 634 _v = _l2normalize(K.dot(_u, K.transpose(W))) 635 _u = _l2normalize(K.dot(_v, W)) 636 return _u, _v 637 W_shape = self.kernel.shape.as_list() 638 #Flatten the Tensor 639 W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]]) 640 _u, _v = power_iteration(W_reshaped, self.u) 641 #Calculate Sigma 642 sigma=K.dot(_v, W_reshaped) 643 sigma=K.dot(sigma, K.transpose(_u)) 644 #normalize it 645 W_bar = W_reshaped / sigma 646 #reshape weight tensor 647 if training in {0, False}: 648 W_bar = K.reshape(W_bar, W_shape) 649 else: 650 with tf.control_dependencies([self.u.assign(_u)]): 651 W_bar = K.reshape(W_bar, W_shape) 652 self.kernel = W_bar 653 654 outputs = K.conv2d_transpose( 655 inputs, 656 self.kernel, 657 output_shape, 658 self.strides, 659 padding=self.padding, 660 data_format=self.data_format) 661 662 if self.use_bias: 663 outputs = K.bias_add( 664 outputs, 665 self.bias, 666 data_format=self.data_format) 667 668 if self.activation is not None: 669 return self.activation(outputs) 670 return outputs
完成了该部之后开始正文。
首先是导入数据集：
1 # 导入CIFAR10数据集 2 # 读取数据 3 def unpickle(file): 4 import pickle 5 with open(file, 'rb') as fo: 6 dict = pickle.load(fo, encoding='bytes') 7 return dict 8 9 cifar={} 10 for i in range(5): 11 cifar1=unpickle('data_batch_'+str(i+1)) 12 if i==0: 13 cifar[b'data']=cifar1[b'data'] 14 cifar[b'labels']=cifar1[b'labels'] 15 else: 16 cifar[b'data']=np.vstack([cifar1[b'data'],cifar[b'data']]) 17 cifar[b'labels']=np.hstack([cifar1[b'labels'],cifar[b'labels']]) 18 target_name=unpickle('batches.meta') 19 cifar[b'label_names']=target_name[b'label_names'] 20 del cifar1 21 22 # 定义数据格式 23 blank_image= np.zeros((len(cifar[b'data']),32,32,3), np.uint8) 24 for i in range(len(cifar[b'data'])): 25 blank_image[i] = np.zeros((32,32,3), np.uint8) 26 blank_image[i][:,:,0]=cifar[b'data'][i][0:1024].reshape(32,32) 27 blank_image[i][:,:,1]=cifar[b'data'][i][1024:1024*2].reshape(32,32) 28 blank_image[i][:,:,2]=cifar[b'data'][i][1024*2:1024*3].reshape(32,32) 29 cifar[b'data']=blank_image 30 31 cifar_test=unpickle('test_batch') 32 cifar_test[b'labels']=np.array(cifar_test[b'labels']) 33 blank_image= np.zeros((len(cifar_test[b'data']),32,32,3), np.uint8) 34 for i in range(len(cifar_test[b'data'])): 35 blank_image[i] = np.zeros((32,32,3), np.uint8) 36 blank_image[i][:,:,0]=cifar_test[b'data'][i][0:1024].reshape(32,32) 37 blank_image[i][:,:,1]=cifar_test[b'data'][i][1024:1024*2].reshape(32,32) 38 blank_image[i][:,:,2]=cifar_test[b'data'][i][1024*2:1024*3].reshape(32,32) 39 cifar_test[b'data']=blank_image 40 41 42 x_train=cifar[b'data'] 43 x_test=cifar[b'data'] 44 x_test=cifar_test[b'data'] 45 y_test=cifar_test[b'labels'] 46 X = np.concatenate((x_test,x_train))
以上是在cifar 10 官方网站下载的数据文件。也可以使用keras官方的cifar10导入代码:
from keras.datasets import cifar100, cifar10 (x_train, y_train), (x_test, y_test) = cifar10.load_data()
接下来是生成器与判别器的构造:
1 # Hyperperemeter 2 # 生成器与判别器构造 3 BATCHSIZE=64 4 LEARNING_RATE = 0.0002 5 TRAINING_RATIO = 1 6 BETA_1 = 0.0 7 BETA_2 = 0.9 8 EPOCHS = 500 9 BN_MIMENTUM = 0.1 10 BN_EPSILON = 0.00002 11 SAVE_DIR = 'img/generated_img_CIFAR10_DCGAN/' 12 13 GENERATE_ROW_NUM = 8 14 GENERATE_BATCHSIZE = GENERATE_ROW_NUM*GENERATE_ROW_NUM 15 16 def BuildGenerator(summary=True): 17 model = Sequential() 18 model.add(Dense(4*4*512, kernel_initializer='glorot_uniform' , input_dim=128)) 19 model.add(Reshape((4,4,512))) 20 model.add(Conv2DTranspose(256, kernel_size=4, strides=2, padding='same', activation='relu',kernel_initializer='glorot_uniform')) 21 model.add(BatchNormalization(epsilon=BN_EPSILON, momentum=BN_MIMENTUM)) 22 model.add(Conv2DTranspose(128, kernel_size=4, strides=2, padding='same', activation='relu',kernel_initializer='glorot_uniform')) 23 model.add(BatchNormalization(epsilon=BN_EPSILON, momentum=BN_MIMENTUM)) 24 model.add(Conv2DTranspose(64, kernel_size=4, strides=2, padding='same', activation='relu',kernel_initializer='glorot_uniform')) 25 model.add(BatchNormalization(epsilon=BN_EPSILON, momentum=BN_MIMENTUM)) 26 model.add(Conv2DTranspose(3, kernel_size=3, strides=1, padding='same', activation='tanh')) 27 if summary: 28 print("Generator") 29 model.summary() 30 return model 31 32 def BuildDiscriminator(summary=True, spectral_normalization=True): 33 if spectral_normalization: 34 model = Sequential() 35 model.add(ConvSN2D(64, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same', input_shape=(32,32,3) )) 36 model.add(LeakyReLU(0.1)) 37 model.add(ConvSN2D(64, kernel_size=4, strides=2,kernel_initializer='glorot_uniform', padding='same')) 38 model.add(LeakyReLU(0.1)) 39 model.add(ConvSN2D(128, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same')) 40 model.add(LeakyReLU(0.1)) 41 model.add(ConvSN2D(128, kernel_size=4, strides=2,kernel_initializer='glorot_uniform', padding='same')) 42 model.add(LeakyReLU(0.1)) 43 model.add(ConvSN2D(256, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same')) 44 model.add(LeakyReLU(0.1)) 45 model.add(ConvSN2D(256, kernel_size=4, strides=2,kernel_initializer='glorot_uniform', padding='same')) 46 model.add(LeakyReLU(0.1)) 47 model.add(ConvSN2D(512, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same')) 48 model.add(LeakyReLU(0.1)) 49 model.add(Flatten()) 50 model.add(DenseSN(1,kernel_initializer='glorot_uniform')) 51 else: 52 model = Sequential() 53 model.add(Conv2D(64, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same', input_shape=(32,32,3) )) 54 model.add(LeakyReLU(0.1)) 55 model.add(Conv2D(64, kernel_size=4, strides=2,kernel_initializer='glorot_uniform', padding='same')) 56 model.add(LeakyReLU(0.1)) 57 model.add(Conv2D(128, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same')) 58 model.add(LeakyReLU(0.1)) 59 model.add(Conv2D(128, kernel_size=4, strides=2,kernel_initializer='glorot_uniform', padding='same')) 60 model.add(LeakyReLU(0.1)) 61 model.add(Conv2D(256, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same')) 62 model.add(LeakyReLU(0.1)) 63 model.add(Conv2D(256, kernel_size=4, strides=2,kernel_initializer='glorot_uniform', padding='same')) 64 model.add(LeakyReLU(0.1)) 65 model.add(Conv2D(512, kernel_size=3, strides=1,kernel_initializer='glorot_uniform', padding='same')) 66 model.add(LeakyReLU(0.1)) 67 model.add(Flatten()) 68 model.add(Dense(1,kernel_initializer='glorot_uniform')) 69 if summary: 70 print('Discriminator') 71 print('Spectral Normalization: {}'.format(spectral_normalization)) 72 model.summary() 73 return model 74 75 def wasserstein_loss(y_true, y_pred): 76 return K.mean(y_true*y_pred) 77 78 generator = BuildGenerator() 79 discriminator = BuildDiscriminator()
然后是训练器的构造:
1 # 生成器训练模型 2 Noise_input_for_training_generator = Input(shape=(128,)) 3 Generated_image = generator(Noise_input_for_training_generator) 4 Discriminator_output = discriminator(Generated_image) 5 model_for_training_generator = Model(Noise_input_for_training_generator, Discriminator_output) 6 print("model_for_training_generator") 7 model_for_training_generator.summary() 8 9 discriminator.trainable = False 10 model_for_training_generator.summary() 11 12 model_for_training_generator.compile(optimizer=Adam(LEARNING_RATE, beta_1=BETA_1, beta_2=BETA_2), loss=wasserstein_loss) 13 14 15 # 判别器训练模型 16 Real_image = Input(shape=(32,32,3)) 17 Noise_input_for_training_discriminator = Input(shape=(128,)) 18 Fake_image = generator(Noise_input_for_training_discriminator) 19 Discriminator_output_for_real = discriminator(Real_image) 20 Discriminator_output_for_fake = discriminator(Fake_image) 21 22 model_for_training_discriminator = Model([Real_image, 23 Noise_input_for_training_discriminator], 24 [Discriminator_output_for_real, 25 Discriminator_output_for_fake]) 26 print("model_for_training_discriminator") 27 generator.trainable = False 28 discriminator.trainable = True 29 model_for_training_discriminator.compile(optimizer=Adam(LEARNING_RATE, beta_1=BETA_1, beta_2=BETA_2), loss=[wasserstein_loss, wasserstein_loss]) 30 model_for_training_discriminator.summary() 31 32 33 real_y = np.ones((BATCHSIZE, 1), dtype=np.float32) 34 fake_y = -real_y 35 36 X = X/255*2-1 37 38 plt.imshow((X[8787]+1)/2)
最后是重复训练：
1 test_noise = np.random.randn(GENERATE_BATCHSIZE, 128) 2 W_loss = [] 3 discriminator_loss = [] 4 generator_loss = [] 5 for epoch in range(EPOCHS): 6 np.random.shuffle(X) 7 8 print("epoch {} of {}".format(epoch+1, EPOCHS)) 9 num_batches = int(X.shape[0] // BATCHSIZE) 10 11 print("number of batches: {}".format(int(X.shape[0] // (BATCHSIZE)))) 12 13 progress_bar = Progbar(target=int(X.shape[0] // (BATCHSIZE * TRAINING_RATIO))) 14 minibatches_size = BATCHSIZE * TRAINING_RATIO 15 16 start_time = time() 17 for index in range(int(X.shape[0] // (BATCHSIZE * TRAINING_RATIO))): 18 progress_bar.update(index) 19 discriminator_minibatches = X[index * minibatches_size:(index + 1) * minibatches_size] 20 21 for j in range(TRAINING_RATIO): 22 image_batch = discriminator_minibatches[j * BATCHSIZE : (j + 1) * BATCHSIZE] 23 noise = np.random.randn(BATCHSIZE, 128).astype(np.float32) 24 discriminator.trainable = True 25 generator.trainable = False 26 discriminator_loss.append(model_for_training_discriminator.train_on_batch([image_batch, noise], 27 [real_y, fake_y])) 28 discriminator.trainable = False 29 generator.trainable = True 30 generator_loss.append(model_for_training_generator.train_on_batch(np.random.randn(BATCHSIZE, 128), real_y)) 31 32 print('\nepoch time: {}'.format(time()-start_time)) 33 34 W_real = model_for_training_generator.evaluate(test_noise, real_y) 35 print(W_real) 36 W_fake = model_for_training_generator.evaluate(test_noise, fake_y) 37 print(W_fake) 38 W_l = W_real+W_fake 39 print('wasserstein_loss: {}'.format(W_l)) 40 W_loss.append(W_l) 41 #Generate image 42 generated_image = generator.predict(test_noise) 43 generated_image = (generated_image+1)/2 44 for i in range(GENERATE_ROW_NUM): 45 new = generated_image[i*GENERATE_ROW_NUM:i*GENERATE_ROW_NUM+GENERATE_ROW_NUM].reshape(32*GENERATE_ROW_NUM,32,3) 46 if i!=0: 47 old = np.concatenate((old,new),axis=1) 48 else: 49 old = new 50 print('plot generated_image') 51 plt.imsave('{}/SN_epoch_{}.png'.format(SAVE_DIR, epoch), old)
训练十轮之后生成的图片有显著的提升，结果如下：
第一轮：
第10轮：
