【发布时间】：2017-01-31 20:12:27
【问题描述】：
我的输入数据形状是 [n, 3, 64, 64]
我在 Stampede 上运行代码后得到了这个。
Using Theano backend.
Using gpu device 0: Tesla K20m (CNMeM is disabled, cuDNN not available)
ValueError: GpuCorrMM images and kernel must have the same stack size
Apply node that caused the error: GpuCorrMM{half, (1, 1)}(GpuContiguous.0, GpuContiguous.0)
Toposort index: 115
Inputs types: [CudaNdarrayType(float32, 4D), CudaNdarrayType(float32, 4D)]
Inputs shapes: [(32, 8, 16, 1024), (256, 512, 5, 5)]
Inputs strides: [(131072, 16384, 1024, 1), (12800, 25, 5, 1)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [[GpuElemwise{Add}[(0, 0)](GpuCorrMM{half, (1, 1)}.0, GpuReshape{4}.0)]]
代码发生了什么以及如何解决这个问题？谢谢
我的代码：
g_input = Input(shape=(100,))
generator = Sequential()
generator.add(Dense(1024 * 4 * 4, input_shape=(100,)))
generator.add(BatchNormalization(mode=2))
generator.add(Activation('relu'))
generator.add(Reshape([1024, 4, 4]))
generator.add(UpSampling2D(size=(2, 2), dim_ordering='th'))
generator.add(Convolution2D(512, 5, 5, border_mode='same', dim_ordering='th'))
generator.add(BatchNormalization(mode=2))
generator.add(Activation('relu'))
generator.add(UpSampling2D(size=(2, 2), dim_ordering='th'))
generator.add(Convolution2D(256, 5, 5, border_mode='same', dim_ordering='th'))
generator.add(BatchNormalization(mode=2))
generator.add(Activation('relu'))
generator.add(UpSampling2D(size=(2, 2), dim_ordering='th'))
generator.add(Convolution2D(128, 5, 5, border_mode='same', dim_ordering='th'))
generator.add(BatchNormalization(mode=2))
generator.add(Activation('relu'))
generator.add(UpSampling2D(size=(2, 2), dim_ordering='th'))
generator.add(Convolution2D(64, 5, 5, border_mode='same', dim_ordering='th'))
generator.add(BatchNormalization(mode=2))
generator.add(Activation('relu'))
generator.add(Convolution2D(3, 5, 5, border_mode='same', dim_ordering='th'))
generator.add(Activation('sigmoid'))
generator.compile(loss='binary_crossentropy', optimizer=Adam(lr=0.0002, beta_1=0.5))
generator.summary()
# discriminative model
discriminator = Sequential()
discriminator.add(Convolution2D(64, 5, 5, subsample=(2, 2), border_mode='same', dim_ordering='th', input_shape=X_train.shape[1:]))
discriminator.add(LeakyReLU(0.2))
discriminator.add(Convolution2D(128, 5, 5, subsample=(2, 2), border_mode='same', dim_ordering='th'))
discriminator.add(LeakyReLU(0.2))
discriminator.add(Convolution2D(256, 5, 5, subsample=(2, 2), border_mode='same', dim_ordering='th'))
discriminator.add(LeakyReLU(0.2))
discriminator.add(Convolution2D(512, 5, 5, subsample=(2, 2), border_mode='same', dim_ordering='th'))
discriminator.add(LeakyReLU(0.2))
discriminator.add(Flatten())
discriminator.add(Dense(1024))
discriminator.add(LeakyReLU(0.2))
discriminator.add(Dropout(0.5))
discriminator.add(Dense(2, activation='softmax'))
discriminator.compile(loss='categorical_crossentropy', optimizer=Adam(lr=0.0002, beta_1=0.5))
discriminator.summary()
# GAN Model
gan_input = Input(shape=(100,))
gan_output = discriminator(generator(gan_input))
gan_model = Model(gan_input, gan_output)
gan_model.compile(loss='categorical_crossentropy', optimizer=Adam(lr=0.0002, beta_1=0.5))
gan_model.summary()
print("Pre-training generator...")
noise_gen = np.random.uniform(0, 1, size=(14000, 100)) # at (0,1) creates 10000 points
generated_images = generator.predict(noise_gen)
print('generated_images shape ----', generated_images.shape)
X = np.concatenate((X_train[:14000, :, :, :], generated_images))
y = np.zeros([28000, 2])
y[:14000, 1] = 1
y[14000:, 0] = 1
discriminator.fit(X, y, nb_epoch=1, batch_size=128)
y_hat = discriminator.predict(X)
# set up loss storage vector
losses = {"d": [], "g": []}
def train_for_n(nb_epoch=28000, batch_size=128):
for e in range(nb_epoch):
# Make generative images
train_idx = np.random.randint(0, X_train.shape[0], size=batch_size) # 0 <= train_idx <= X_train.shape[0]
mini_batch = X_train[train_idx]
noise_gen = np.random.normal(0, 1, size=(batch_size, 100))
generated_images = generator.predict(noise_gen)
# Train discriminator on generated images
X = np.concatenate((mini_batch, generated_images))
y = np.zeros([2 * batch_size, 2])
y[:batch_size, 1] = 1
y[batch_size:, 0] = 1
discriminator.trainable = True
for layer in discriminator.layers:
layer.trainable = True
d_loss = discriminator.train_on_batch(X, y)
losses["d"].append(d_loss)
noise_tr = np.random.uniform(0, 1, size=(batch_size, 100))
y2 = np.zeros([batch_size, 2])
y2[:, 1] = 1
discriminator.trainable = False
for layer in discriminator.layers:
layer.trainable = False
g_loss = gan_model.train_on_batch(noise_tr, y2)
losses["g"].append(g_loss)
if e % 10 == 9:
generator.save_weights('G0_weights.h5')
discriminator.save_weights('D0_weights.h5')
noise = np.random.uniform(0, 1, size=(100, 100))
generated_images = generator.predict(noise)
np.save('/Users/zhangguanghua/Desktop/Stampede/generated_images_0.npy', generated_images)
print(("Iteration: {0} / {1}, G-Loss: {2:.4f}".format(e, nb_epoch, float(g_loss))))
train_for_n(nb_epoch=2000, batch_size=128)
另外，有人知道什么是输入形状：[(32, 8, 16, 1024), (256, 512, 5, 5)] 代表什么？我该如何解决这个问题？
谢谢
【问题讨论】：
-
你能在你的帖子中添加一个 gan_model.summary() 结果吗？
-
感谢您的快速回复。添加 gan_model.summary()result 是什么意思？我不明白。谢谢
-
此函数打印项目的摘要。它应该打印它的拓扑
-
太长了，我在答案中添加了 summary() 结果。
-
你能检查一下 Keras.json 中的 dim_ordering 吗？
标签： python theano keras gpuimage