Keras fit_generator producing exception: output of generator should be a tuple(x, y, sample_weight) or (x, y). Found: [[[[ 0.86666673

Question

I am trying to build an autoencoder for non MNIST, non Imagenet data. Using https://blog.keras.io/building-autoencoders-in-keras.html as my base. However, am getting the following error.

**Exception: output of generator should be a tuple (x, y, sample_weight) or (x, y). Found: [[[[ 0.86666673  0.86666673  0.86666673 ...,  0.62352943  0.627451
     0.63137257]
   [ 0.86666673  0.86666673  0.86666673 ...,  0.63137257  0.627451
     0.627451  ]
   [ 0.86666673  0.86666673  0.86666673 ...,  0.63137257  0.627451
     0.62352943]
   ...,**

Since this is an autoencoder, in my datagenerator, used class mode=None. My code is as follows.

from keras.layers import Input, Dense, Convolution2D, MaxPooling2D,   UpSampling2D,Activation, Dropout, Flatten
from keras.models import Model,Sequential
from keras.preprocessing.image import ImageDataGenerator 
import numpy as np
import os
import h5py


img_width=140 
img_height=140
train_data_dir=r'SitePhotos\train'
valid_data_dir=r'SitePhotos\validation'
input_img = Input(batch_shape=(32,3, img_width, img_width))

x = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(input_img)
x = MaxPooling2D((2, 2), border_mode='same')(x)
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x)
x = MaxPooling2D((2, 2), border_mode='same')(x)
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x)
encoded = MaxPooling2D((2, 2), border_mode='same')(x)

# at this point the representation is (8, 4, 4) i.e. 128-dimensional

x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(encoded)
x = UpSampling2D((2, 2))(x)
x = Convolution2D(8, 3, 3, activation='relu', border_mode='same')(x)
x = UpSampling2D((2, 2))(x)
x = Convolution2D(16, 3, 3, activation='relu')(x)
x = UpSampling2D((2, 2))(x)
decoded = Convolution2D(1, 3, 3, activation='sigmoid', border_mode='same')(x)

autoencoder = Model(input_img, decoded)
autoencoder.compile(optimizer='adadelta', loss='mse')



valid_datagen = ImageDataGenerator(rescale=1./255)
train_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
        train_data_dir,
        target_size=(img_width, img_height),
        batch_size=32,
        class_mode=None,
        shuffle=True)


valid_generator = valid_datagen.flow_from_directory(
        valid_data_dir,
        target_size=(img_width, img_height),
        batch_size=32,
        class_mode=None,
        shuffle=True)

autoencoder.fit_generator(train_generator,
                nb_epoch=50,                
                validation_data=valid_generator,
                samples_per_epoch=113,
                nb_val_samples=32
                )

Answer 1

The real solution lies in this Keras issue apparently by @skottapa.

https://github.com/fchollet/keras/issues/4260

rodgzilla provided an updated ImageDataGenerator that adds a class_mode='input' that solves the problem.

The nice thing is that you can backport the modification to older Keras versions. A slightly modified image module can be downloaded here:

https://gist.github.com/gsdefender/293db0987a800cf1b103b7777966f8af