Cannot use generator to feed auto-encoder

[harry.k...@gmail.com]

I've successfully trained a convolutional auto-encoder on a 'smaller' (~17.7k samples) dataset however I do not have enough memory to scale it up to a larger set (~1.2M samples)

Most sources have suggested using the fit_generator function, however I have not been able to successfully run it due to a dimensionality mismatch in the final layer of the network. The final layer is correctly outputting a (300, 300, 3) image, the same size as the input image, however the generator is yielding the full batch size (50, 300, 300, 30 and causing the error.

I am very new to generators so there's a high chance it is a trivial solution, but I have tried several alterations with no success

thanks in regards

CODE:

def BatchGenerator(tarFile,batch_size):

    tar = tarfile.open(FILE_PATH + '/' + tarFile, "r:gz")
    members=tar.getnames()

    while True:
        batch_labels = np.zeros((batch_size,1))
        batch_features = np.zeros((batch_size, 300, 300, 3))
        count=0
        while count<batch_size:
            member= random.choice(members)
            f = tar.extractfile(member)
            if f:
                content = f.read()
                if content:
                    try:
                        img = Image.open(f)
                        if img.mode == 'RGBA':
                            img = img.convert('RGB')
                        imgArray = np.array(img)
                        batch_features[count]=imgArray
                        count+=1
                        print(count)
                    except:
                        continue
        yield batch_features,batch_labels

a=BatchGenerator('1.tar.gz',batch_size)

model = Sequential()

model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same', input_shape=(300, 300, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size, border_mode='same'))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(3, 3), border_mode='same'))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size, border_mode='same'))


model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(UpSampling2D(pool_size))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(UpSampling2D((3, 3)))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(UpSampling2D(pool_size))
model.add(Convolution2D(3, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('sigmoid'))

model.summary()

checkpointer = ModelCheckpoint(filepath="./tmp/weights.hdf5", verbose=1, save_best_only=True)
logger = CSVLogger("./log.csv", separator=',', append=False)
model.compile(loss='mse', optimizer=RMSprop(), metrics=['accuracy'])

# history = model.fit(X_train, X_train, callbacks=[checkpointer, logger], batch_size=batch_size, nb_epoch=nb_epoch, verbose=1, validation_data=(X_train, X_train))
history = model.fit_generator(a, samples_per_epoch = 50, nb_epoch = nb_epoch, verbose=2, callbacks=[], validation_data=None, class_weight=None, nb_worker=1)

ERROR:
Traceback (most recent call last):
  File "/home/kre033/PycharmProjects/DeepLearning/RBM/icons/iconAutoencoder.py", line 241, in <module>
    history = model.fit_generator(a, samples_per_epoch = 50, nb_epoch = nb_epoch, verbose=2, callbacks=[], validation_data=None, class_weight=None, nb_worker=1)
  File "/usr/local/lib/python3.5/dist-packages/keras/models.py", line 907, in fit_generator
    pickle_safe=pickle_safe)
  File "/usr/local/lib/python3.5/dist-packages/keras/engine/training.py", line 1451, in fit_generator
    class_weight=class_weight)
  File "/usr/local/lib/python3.5/dist-packages/keras/engine/training.py", line 1220, in train_on_batch
    check_batch_dim=True)
  File "/usr/local/lib/python3.5/dist-packages/keras/engine/training.py", line 967, in _standardize_user_data
    exception_prefix='model target')
  File "/usr/local/lib/python3.5/dist-packages/keras/engine/training.py", line 100, in standardize_input_data
    str(array.shape))
Exception: Error when checking model target: expected activation_7 to have 4 dimensions, but got array with shape (50, 1)

[Daπid]

On 23 March 2017 at 02:30, <harry.k...@gmail.com> wrote:

model = Sequential()

model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same', input_shape=(300, 300, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size, border_mode='same'))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(3, 3), border_mode='same'))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size, border_mode='same'))


model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(UpSampling2D(pool_size))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(UpSampling2D((3, 3)))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('relu'))
model.add(UpSampling2D(pool_size))
model.add(Convolution2D(3, kernel_size[0], kernel_size[1], border_mode='same'))
model.add(Activation('sigmoid'))

model.summary()

This model returns an image the same size as the input: 300x300. But:

        batch_labels = np.zeros((batch_size,1))

You are feeding one value per image (like a binary classification).  If you are building an autoencoder, you need to replace that for the original image:

yield batch_features, batch_features

/David.