我正试图用自定义数据生成器训练一个堆叠卷积自动编码器，因为它是我生成的一个非常大的合成数据集。我一直在关注https://medium.com/@mrgarg.rajat/training-on-large-datasets-tat-dont-fit-in-memory-in-keras-60a974785d71教程，但仍然无法使其工作
我的数据集目录如下：
real_train
- img 1.png
- img 2.png
- ....
这是我的my_Data_Generator类
class My_Data_Generator(keras.utils.Sequence):
def __init__(self, image_filenames, labels, batch_size):
self.image_filenames = image_filenames
self.labels = labels
self.batch_size = batch_size
self.n = 0
def __next__(self):
# Get one batch of data
data = self.__getitem__(self.n)
# Batch index
self.n += 1
# If we have processed the entire dataset then
if self.n >= self.__len__():
self.on_epoch_end
self.n = 0
return data
def __len__(self) :
return (np.ceil(len(self.image_filenames) / float(self.batch_size))).astype(np.int)
def __getitem__(self, idx):
batch_x = self.image_filenames[idx * self.batch_size:(idx + 1) * self.batch_size]
batch_y = self.labels[idx * self.batch_size:(idx + 1) * self.batch_size]
return np.array([
resize(imread('E:/FontRecognition/Dataset_Final/preprocessed/real_train/' + str(file_name)), (105,105,1))
for file_name in batch_x])/255.0, np.array(batch_y)
这是我的代码
# load
X_train = np.load('X_train_filenames.npy')
X_val = np.load('X_val_filenames.npy')
# print(X_train.shape)
# print(X_val.shape)
batch_size = 128
my_training_batch_generator = My_Data_Generator(X_train, X_train, batch_size=batch_size)
my_validation_batch_generator = My_Data_Generator(X_val, X_val, batch_size=batch_size)
images, labels = next(my_training_batch_generator)
print("Train")
print(images.shape)
print(labels.shape)
images, labels = next(my_validation_batch_generator)
print("Val")
print(images.shape)
print(labels.shape)
input_img = Input(shape=(105,105,1))
x = Conv2D(64, kernel_size=(48,48), activation='relu', padding='same', strides=1)(input_img)
x = BatchNormalization()(x)
x = MaxPooling2D(pool_size=(2,2)) (x)
x = Conv2D(128, kernel_size=(24,24), activation='relu', padding='same', strides=1)(x)
x = BatchNormalization()(x)
encoded = MaxPooling2D(pool_size=(2,2))(x)
x = Conv2D(64, kernel_size=(24,24), activation='relu', padding='same', strides=1)(encoded)
x = UpSampling2D(size=(2,2))(x)
x = Conv2D(1, kernel_size=(48,48), activation='relu', padding='same', strides=1)(x)
decoded = UpSampling2D(size=(2,2))(x)
adam = keras.optimizers.Adam(lr=0.01)
autoencoder = Model(input_img, decoded)
autoencoder.compile(optimizer=adam, loss='mean_squared_error')
autoencoder.summary()
num_epochs = 20
autoencoder.fit_generator(generator=my_training_batch_generator,
steps_per_epoch=(int(1836695 // batch_size)),
epochs=num_epochs,
verbose=1,
validation_data=my_validation_batch_generator,
validation_steps=(int(459174 // batch_size))
# use_multiprocessing=True,
# workers=6
)
print("Finished")
我试着运行代码，结果是：
Model: "model_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 105, 105, 1) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 105, 105, 64) 147520
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 52, 52, 64) 0
_________________________________________________________________
batch_normalization_1 (Batch (None, 52, 52, 64) 256
_________________________________________________________________
conv2d_2 (Conv2D) (None, 52, 52, 128) 4718720
_________________________________________________________________
batch_normalization_2 (Batch (None, 52, 52, 128) 512
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 26, 26, 128) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 26, 26, 64) 4718656
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 52, 52, 64) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 52, 52, 1) 147457
_________________________________________________________________
up_sampling2d_2 (UpSampling2 (None, 104, 104, 1) 0
=================================================================
Total params: 9,733,121
Trainable params: 9,732,737
Non-trainable params: 384
_________________________________________________________________
Epoch 1/20
Traceback (most recent call last):
File "SCAE_train.py", line 142, in <module>
validation_steps=(int(459174 // batch_size))
File "C:MyProgramFilesAnaconda3envstf_gpulibsite-packageskeraslegacyinterfaces.py", line 91, in wrapper
return func(*args, **kwargs)
File "C:MyProgramFilesAnaconda3envstf_gpulibsite-packageskerasenginetraining.py", line 1732, in fit_generator
initial_epoch=initial_epoch)
File "C:MyProgramFilesAnaconda3envstf_gpulibsite-packageskerasenginetraining_generator.py", line 221, in fit_generator
reset_metrics=False)
File "C:MyProgramFilesAnaconda3envstf_gpulibsite-packageskerasenginetraining.py", line 1508, in train_on_batch
class_weight=class_weight)
File "C:MyProgramFilesAnaconda3envstf_gpulibsite-packageskerasenginetraining.py", line 621, in _standardize_user_data
exception_prefix='target')
File "C:MyProgramFilesAnaconda3envstf_gpulibsite-packageskerasenginetraining_utils.py", line 135, in standardize_input_data
'with shape ' + str(data_shape))
ValueError: Error when checking target: expected up_sampling2d_2 to have 4 dimensions, but got array with shape (128, 1)
我是keras和python的新手，但我仍然不知道是什么原因导致的。
首先，模型的输入和输出形状不匹配。您的模型输入大小为105x105，而输出大小为104x104。要么使用类似的输入大小，要么在卷积层中调整内核/步长大小。
但为了回答您的问题，请注意，您遵循的教程执行分类，因此使用目标形状(batch_size，number_of_categories(。然而，您使用的是自动编码器，这意味着您应该更改数据生成器以返回适当的目标，即，使(batch_size、HEIGHT、WIDTH、NUM_CHANNELS(的形状与您的输入相同。
您的输入和输出图像是相同的，因此在数据生成器中不需要额外的labels参数，只需读取图像并返回它们的两个副本即可。假设你有正确格式/目录中的图像文件，我已经编辑了你的代码，如下所示：
您的数据生成器：
class My_Custom_Generator(keras.utils.Sequence) :
def __init__(self, image_filenames, batch_size) :
self.image_filenames = image_filenames
self.batch_size = batch_size
def __len__(self) :
return (np.ceil(len(self.image_filenames) / float(self.batch_size))).astype(np.int)
def __getitem__(self, idx) :
batch_x = self.image_filenames[idx * self.batch_size : (idx+1) * self.batch_size]
current_x = np.array(
resize(imread('E:/FontRecognition/Dataset_Final/preprocessed/real_train/' + str(file_name)), (105,105,1))
for file_name in batch_x])/255.0
return current_x, current_x
您的模型和脚本：
# load
X_train = np.load('X_train_filenames.npy')
X_val = np.load('X_val_filenames.npy')
# print(X_train.shape)
# print(X_val.shape)
batch_size = 128
my_training_batch_generator = My_Data_Generator(X_train, batch_size=batch_size)
my_validation_batch_generator = My_Data_Generator(X_val, batch_size=batch_size)
input_img = keras.layers.Input(shape=(104,104,1))
x = keras.layers.Conv2D(64, kernel_size=(48,48), activation='relu', padding='same', strides=1)(input_img)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.MaxPooling2D(pool_size=(2,2), padding='same') (x)
x = keras.layers.Conv2D(128, kernel_size=(24,24), activation='relu', padding='same', strides=1)(x)
x = keras.layers.BatchNormalization()(x)
encoded = keras.layers.MaxPooling2D(pool_size=(2,2))(x)
x = keras.layers.Conv2D(64, kernel_size=(24,24), activation='relu', padding='same', strides=1)(encoded)
x = keras.layers.UpSampling2D(size=(2,2))(x)
x = keras.layers.Conv2D(1, kernel_size=(48,48), activation='relu', padding='same', strides=1)(x)
decoded = keras.layers.UpSampling2D(size=(2,2))(x)
autoencoder = keras.Model(input_img, decoded)
autoencoder.summary()
adam = keras.optimizers.Adam(lr=0.01)
autoencoder.compile(optimizer=adam, loss='mean_squared_error')
num_epochs = 20
autoencoder.fit_generator(generator=my_training_batch_generator,
epochs=num_epochs,
verbose=1,
validation_data=my_validation_batch_generator
# use_multiprocessing=True,
# workers=6
)
请注意，我已经删除了steps_per_epoch和validation_tsteps参数，因为自定义数据生成器继承了keras.utils.Sequence不需要它们，可以直接从数据中推断它们。