Autoencoder for color images in Keras backed by MXNet

Question

I need to train an autoencoder in Keras with the JPG images I took myself.

model = Sequential()

#1st convolution layer
model.add(Conv2D(16, (3, 3), padding='same', data_format='channels_first', input_shape=(3,224,224)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2,2), padding='same'))

#2nd convolution layer
model.add(Conv2D(2,(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2,2), padding='same'))

#-------------------------

#3rd convolution layer
model.add(Conv2D(2,(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(UpSampling2D((2, 2)))

#4rd convolution layer
model.add(Conv2D(16,(3, 3), padding='same'))
model.add(Activation('relu'))
model.add(UpSampling2D((2, 2)))

#-------------------------

model.add(Conv2D(1,(3, 3), padding='same'))
model.add(Activation('sigmoid'))

model.summary()

which generates a model as:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 16, 224, 224)      448       
_________________________________________________________________
activation_1 (Activation)    (None, 16, 224, 224)      0         
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 16, 112, 112)      0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 2, 112, 112)       290       
_________________________________________________________________
activation_2 (Activation)    (None, 2, 112, 112)       0         
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 2, 56, 56)         0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 2, 56, 56)         38        
_________________________________________________________________
activation_3 (Activation)    (None, 2, 56, 56)         0         
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 2, 112, 112)       0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 16, 112, 112)      304       
_________________________________________________________________
activation_4 (Activation)    (None, 16, 112, 112)      0         
_________________________________________________________________
up_sampling2d_2 (UpSampling2 (None, 16, 224, 224)      0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 1, 224, 224)       145       
_________________________________________________________________
activation_5 (Activation)    (None, 1, 224, 224)       0         
=================================================================
Total params: 1,225
Trainable params: 1,225
Non-trainable params: 0
_________________________________________________________________

I compile and train as:

model.compile(optimizer='adadelta', loss='binary_crossentropy')
batch_size = 16 #16

train_datagen = ImageDataGenerator(rescale=1./255)
train_generator = train_datagen.flow_from_directory(
    'cropped/',
    target_size=(224, 224),
    batch_size=batch_size,
    class_mode='categorical'
    )
test_datagen = ImageDataGenerator(rescale=1./255)
validation_generator = test_datagen.flow_from_directory(
    'cropped/',
    target_size=(224, 224),
    batch_size=batch_size,
    class_mode='categorical'
    )

model.fit_generator(
        train_generator,
        steps_per_epoch=1000,
        epochs=20,
        validation_data=validation_generator,
        validation_steps=1000)

I end up with the error message:

ValueError: Error when checking target: expected activation_5 to have 4 dimensions, but got array with shape (16, 2)

Should I use Conv2D instead?

Answer 1

It was not about the model, but about ImageDataGenerator. I found the answer here: https://stackoverflow.com/a/51673998/493080