Keras: [help wanted] Input shape / target size flow from directory

[vkessler]

Hi,
i´m wondering about the correct input_shape and/or target_size when using fit_generator and flow_from_directory. I´m running keras 2.0.8-tf and tensorflow 1.4 as the backend. My images have the following dimensions:

width: 725 Pixel
height: 180 Pixel
channels: 3

I define the Input_shape as: (width, height, channels) --> (725,180,3), as it is suggested when using tensorflow as the backend.

I use the ImageDataGenerator with the flow_from_directory method to get the training data. Here the target_size may be specified. The target_size can be either (width, height) or (height, width). The API is defining it as (height, width).

And now the Problem occurs. If the Input_shape is (width, height, channels) and the target_size is (heigth, width) - as it should be - a ValueError is raised, saying the shapes don´t match (Error when checking input: expected conv2d_1_input to have shape (None, 725, 180, 3) but got array with shape (50, 180, 725, 3)).

If i set the target_size to (width, height), the training is starting, but if i then plot an Image using the Generator, i see that width and height are exchanged and hence the Image is massively transformed.

How to solve this issue? In fchollets great example the target_size is also set the other way around as in the API.

Any help is appreciated! Thank you.

Here is my code example:

# Set the dimension of the images 
img_width = 725 
img_height = 180

# Define epochs and batch size
epochs = 1
batch_size = 50

# TensorFlow is the backend, so ordering of input_shape is as below
input_shape = (img_width, img_height, 3)

# Helper function to compile model
def compile_model(model):
    optimizer = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
    model.compile(loss='categorical_crossentropy',
                  optimizer=optimizer,
                  #metrics=['accuracy'])
                  metrics=['accuracy',metrics.categorical_accuracy])

# Define the model
def generate_model(model_path):

    # check if model exists, if exists then load model from saved state
    if Path(model_path).is_file():
        sys.stdout.write('Loading existing model\n\n')
        sys.stdout.flush() #Warum?
        model = load_model(model_path)
        compile_model(model)
        return model

    sys.stdout.write('Loading new model\n\n')
    sys.stdout.flush() #Warum?

    # Define new model
    model = Sequential()
    model.add(Conv2D(32, (4, 4), input_shape=input_shape))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))

    model.add(Conv2D(32, (4, 4)))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))

    model.add(Conv2D(32, (4, 4), input_shape=input_shape))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))

    model.add(Conv2D(64, (4, 4)))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))

    model.add(Flatten())
    model.add(Dense(64))
    model.add(Activation('relu'))
    model.add(Dropout(0.5))
    model.add(Dense(num_classes, activation='softmax'))

    compile_model(model)

    # Save model
    with open(model_path, 'w') as outfile:
        json.dump(model.to_json(), outfile)
        outfile.close()

    return model


model = generate_model(model_path)

# this is the augmentation configuration we will use for training
train_datagen = ImageDataGenerator(
    rescale=1. / 255,
    rotation_range=180.,
    #shear_range=0.2,
    #zoom_range=0.2,
    horizontal_flip=True)

# this is the augmentation configuration we will use for testing:
# only rescaling
test_datagen = ImageDataGenerator(rescale=1. / 255)

train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=(img_height, img_width), # not running this way
    batch_size=batch_size,
    class_mode='categorical')

validation_generator = test_datagen.flow_from_directory(
    validation_data_dir,
    target_size=(img_height, img_width), # not running this way
    batch_size=batch_size,
    class_mode='categorical')

hist = model.fit_generator(
    train_generator,
    steps_per_epoch=nb_train_samples // batch_size,
    epochs=epochs,
    validation_data=validation_generator,
    validation_steps=nb_validation_samples // batch_size)

model.save(model_path)
print('Saved trained model at %s ' % model_path)


x, y = train_generator.next()
classes_val = validation_generator.class_indices
print(classes_val)
for i in range(0,12):
    img = x[i]
    print(y[i])
    plt.imshow(img)
    plt.show()

[davinnovation]

@vkessler https://keras.io/layers/convolutional/. by keras doc, conv2d input shape should be (height, width) == (row, colum) sequence. I think example has a slight mistake of it.

_{Sent with GitHawk}

[vkessler]

Thank you very much for the clarification. I guess i didn´t pay attention enough to the convolutional API.