前提・実現したいこと
kerasをつかって
CIFAR-10の画像を生成するGANの実装をしたいです
https://github.com/oreilly-japan/deep-learning-with-keras-ja/tree/master/ch04
の中の
example_gan_cifar10.py
というコードを修正を加えつつ写しています。
発生している問題・エラーメッセージ
実行しようとしたところ以下のようなエラーメッセージが出てしまいました。
ValueError: Negative dimension size caused by subtracting 4 from 2 for 'sequential_2/average_pooling2d_1/AvgPool' (op: 'AvgPool') with input shapes: [?,2,2,1].
該当のソースコード
# coding: utf-8 # 4.4 CIFAR-10の画像を生成するGANの実装 # cifar10 p121 import os import pandas as pd import numpy as np import matplotlib as mpl mpl.use("Agg") import keras import keras.backend as K # # import keras.backend keras.backend.set_image_dim_ordering("th") from numpy import transpose # from keras.layers import Reshape, Flatten, LeakyReLU, Input, Activation, Dense, BatchNormalization, SpatialDropout2D from keras.layers.convolutional import Conv2D, UpSampling2D, MaxPooling2D, AveragePooling2D from keras.regularizers import L1L2 from keras.models import Sequential, Model from keras.optimizers import Adam from keras.callbacks import TensorBoard from keras.datasets import cifar10 from keras_adversarial.image_grid_callback import ImageGridCallback from keras_adversarial import AdversarialModel, simple_gan, gan_targets, fix_names from keras_adversarial import AdversarialOptimizerSimultaneous, normal_latent_sampling # i tried below # from keras_adversarial.examples.image_utils import * from numpy import * # # # # from image_utils import dim_ordering_unfix # from image_utils import dim_ordering_shape #image_dim_ordering def dim_ordering_fix(x): if keras.backend.image_dim_ordering() == "th": return x else: pass# return numpy.transpose(x, (0, 2, 3, 1)) # return x.transpose((0, 2, 3, 1)) return x.transpose(0, 2, 3, 1) # return numpy.transpose(x, (0, 2, 3, 1)) # めんどいエラー出たERROR def dim_ordering_unfix(x): if keras.backend.image_dim_ordering() == "th": return x else: pass#return numpy.transpose(x, (0, 3, 1, 2)) # return x.transpose((0, 3, 1, 2)) return x.transpose(0, 3, 1, 2) # return numpy.transpose(x, (0, 3, 1, 2)) def dim_ordering_shape(input_shape): if keras.backend.image_dim_ordering() == "th": return input_shape else: return (input_shape[1], input_shape[2], input_shape[0]) def dim_ordering_input(input_shape, name): if keras.backend.image_dim_ordering() == "th": return Input(input_shape, name=name) else: return Input((input_shape[1], input_shape[2], input_shape[0]), name=name) def dim_ordering_reshape(k, w, **kwargs): if keras.backend.image_dim_ordering() =="th": return Reshape((k, w, w), **kwargs) else: return Reshape((w, w, k), **kwargs) def channel_axis(): if keras.backend.image_dim_ordering() =="th": return 1 else: return 3 def model_generator(): model = Sequential() nch = 256 reg = lambda: L1L2(l1=1e-7, l2=1e-7) h = 5 model.add(Dense(nch * 4 * 4, input_dim=100, kernel_regularizer=reg())) pass model.add(BatchNormalization()) model.add( Reshape(dim_ordering_shape((nch, 4, 4))) ) model.add( Conv2D(int(nch/2), (h,h), padding="same", kernel_regularizer=reg(), data_format="channels_first") ) pass model.add(BatchNormalization(axis=1)) model.add(LeakyReLU(0.2)) model.add(UpSampling2D(size=(2,2))) model.add(Conv2D(int(nch/4), (h,h), padding="same", kernel_regularizer=reg(), data_format="channels_first")) pass model.add(BatchNormalization(axis=1)) model.add(LeakyReLU(0.2)) model.add(UpSampling2D(size=(2,2))) model.add(Conv2D(3, (h,h), padding="same", kernel_regularizer=reg(), data_format="channels_first")) pass model.add(Activation("sigmoid")) return model def model_discriminator(): nch=256 h=5 reg = lambda: L1L2(l1=1e-7, l2=1e-7) pass c1 = Conv2D( int(nch / 4), (h,h), padding="same", kernel_regularizer=reg(), input_shape=dim_ordering_shape((3,32, 32)), data_format="channels_first" ) pass c2 = Conv2D( int(nch / 2), (h, h), padding="same", kernel_regularizer=reg(), data_format="channels_first" ) pass c3 = Conv2D( nch, (h,h), padding="same", kernel_regularizer=reg(), data_format="channels_first" ) pass c4 = Conv2D( 1, (h, h), padding="same", kernel_regularizer=reg(), data_format="channels_first" ) def m(dropout): model = Sequential() model.add(c1) model.add(SpatialDropout2D(dropout)) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(LeakyReLU(0.2)) pass model.add(c2) model.add(SpatialDropout2D(dropout)) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(LeakyReLU(0.2)) pass model.add(c3) model.add(SpatialDropout2D(dropout)) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(LeakyReLU(0.2)) pass model.add(c4) model.add(AveragePooling2D(pool_size=(4, 4), padding="valid")) model.add(Flatten()) model.add(Activation("sigmoid")) return model pass return m def cifar10_process(x): x = x.astype(np.float32) / 255.0 return x def cifar10_data(): (xtrain, ytrain), (xtest, ytest) = cifar10.load_data() return cifar10_process(xtrain), cifar10_process(xtest) # def example_gan def example_gan( adversarial_optimizer, path, opt_g, opt_d, nb_epoch, generator, discriminator, latent_dim, targets=gan_targets, loss="binary_crossentropy" ): csvpath = os.path.join(path, "history.csv") if os.path.exists(csvpath): print("Already exists:{}".format(csvpath)) return print("Training: {}".format(csvpath)) pass # gan (x -> yfake, yreal), z is gaussian generated on GPU pass # can also experiment with uniform_latent_sampling d_g = discriminator(0) d_d = discriminator(0.5) generator.summary() d_d.summary() gan_g = simple_gan(generator, d_g, None) gan_d = simple_gan(generator, d_d, None) x = gan_g.inputs[1] z = normal_latent_sampling((latent_dim,))(x) pass # eliminate z from inputs gan_g = Model( [x], fix_names(gan_g([z,x]),gan_g.output_names) ) gan_d = Model( [x], fix_names(gan_d([z,x]),gan_d.output_names) ) pass # build adversarial model model = AdversarialModel( player_models=[gan_g, gan_d], player_params=[generator.traineble_weights, d_d.trainable_weights], player_names=["generator", "discriminator"] ) model.adversarial_compile( adversarial_optimizer=adversarial_optimizer, player_optimizer=[opt_g, opt_d], loss=loss ) pass # create callback to generate images zsample = numpy.random.normal(size=(10 * 10, latent_dim)) pass def generator_sampler(): xpred = generator.predict(zsamples) xpred = dim_ordering_unfix( xpred(transpose((0, 2, 3, 1))) ) return xpred.reshape((10, 10) + xpred.shape[1:]) pass generator_cb = ImageGridCallback( os.path.join(path, "epoch-{:03d}.png"), generator_sampler, cmap=None ) pass callbacks = [generator_cb] if keras.backend.backend() == "tensorflow": callbacks.append( TensorBoard(log_dir=os.path.join(path, "logs"), histogram_freq=0, write_graph=True, write_images=True) ) pass #train model xtrain, xtest = cifar10_data() y = targets(xtrain.shape[0]) pass y_test = targets(xtest.shape[0]) history = model.fit( x=xtrain, y=y, validation_data=(xtest, ytest), callbacks=callbacks, epochs=epoch, batch_size=32 ) pass # save history to csv df = pandas.DataFrame(history.history) df.to_csv(csvpath) pass #save models generator.save( os.path.join(path, "generator.h5") ) d_d.save(os.path.join(path, "discriminator.h5")) def main (): pass # z \in R^100 latent_dim = 100 pass # x \in R^{28x28} pass # generator (z -> x) generator = model_generator() pass # discriminator (x->y) discriminator = model_discriminator() if not os.path.exists("output/gan-cifar10"): os.makedirs("output/gan-cifar10") example_gan( AdversarialOptimizerSimultaneous(), "output/gan-cifar10", opt_g=Adam(1e-4, decay=1e-5), opt_d=Adam(1e-3, decay=1e-5), nb_epoch=1, generator=generator, discriminator=discriminator, latent_dim=latent_dim ) print("A") if __name__=="__main__": main() # dim_ordering_shape dim_ordering_unfix print("A") if __name__ == "__main__": print("B") #
試したこと
githubのisssuesを参考にして、
keras.backend.set_image_dim_ordering("th")
を追記してみたり、
model.add(Conv2D(int(nch/4), (h,h), padding="same", kernel_regularizer=reg())
を以下のようにdata_format="channel_first"を追加したりしました。
model.add(Conv2D(int(nch/4), (h,h), padding="same", kernel_regularizer=reg(), data_format="channels_first"))
補足情報(FW/ツールのバージョンなど)
あなたの回答
tips
プレビュー