Keras dimension mismatch with ImageDataGenerator

问题

I am attempting to 'flow' my data into a neural network with Keras. I am using the .flow_from_directory method and the process is giving me fits. I am using the basic example from the keras documentation (I am using tensorflow):

ROWS = 64
COLS = 64
CHANNELS = 3

from keras.preprocessing.image import ImageDataGenerator
train_datagen = ImageDataGenerator(
    rescale=1./255)

test_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
    'train',
    target_size=(64, 64),
    batch_size=32,
    class_mode='binary')

validation_generator = test_datagen.flow_from_directory(
    '../tutorial/l1/kaggle_solutions/dogs_vs_cats/valid',
    target_size=(64, 64),
    batch_size=1,
    class_mode='binary')
from keras.models import Sequential
from keras.layers import Convolution2D
from keras.layers import Dense, Activation, Flatten, Dropout, MaxPooling2D
from keras.regularizers import l2

model = Sequential()
model.add(Convolution2D(4, 4, 4, border_mode='same', input_shape=(64, 64,3), activation='relu'))
from keras.utils.np_utils import to_categorical
from keras.optimizers import SGD, RMSprop

model.compile(loss='binary_crossentropy', optimizer=RMSprop(lr=1e-4), metrics=['accuracy'])


model.fit_generator(
    train_generator,  
    samples_per_epoch=2500,
    nb_epoch=20,
    validation_data=validation_generator,
    nb_val_samples=3100)

Running this i get the following error:

Exception: Error when checking model target: expected convolution2d_84 to have 4 dimensions, but got array with shape (32, 1)

I have been tinkering around for a long time and found the following--switching the 'model.add' to grayscale input model.add(Convolution2D(4, 4, 4, border_mode='same', input_shape=(64, 64,3), activation='relu')) gives me the following error (as expected--but appears to confirm my original input was correct):

Error when checking model input: expected convolution2d_input_49 to have shape (None, 64, 64, 1) but got array with shape (32, 64, 64, 3)

So I am passing (in the original) a 4-d array of 32,64,64,3 with the original, but I am getting the error that I THINK means Expected (1,64,64,3) and got (32,64,64,3)

As I am sending data in batches of 32. Curiously enough if I set the batch to zero (to give a 0,64,64,3 input) I get:

Exception: Error when checking model target: expected convolution2d_87 to have 4 dimensions, but got array with shape (0, 1)

Based on the documentation, I cannot figure out the proper way to flow the data into the model--i cannot pass the batch size to the model when using fit_generator, and it appears that the batch_size (num of samples) is the problem.

Any help would be greatly appreciated.

回答1:

There is no problem with your ImageDataGenerator. As stated in the error message there is a mismatch between the shape of your model output and the shape of its targets. You use class_mode = 'binary', so expected output of your model is a single value, but instead it yields output of shape (batch_size, 64, 64, 4) since you have one convolutional layer and nothing else in your model.

Try something like this:

model.add(Convolution2D(4, 4, 4, border_mode='same', input_shape=(64, 64,3), activation='relu'))
model.add(Flatten())
model.add(Dense(1))
model.add(Activation('sigmoid'))

来源：https://stackoverflow.com/questions/41946894/keras-dimension-mismatch-with-imagedatagenerator