CNN with multiple conv3d in keras - tensorflow

Question

I am trying to create a CNN model in Keras with multiple conv3d to work on cifar10 dataset. But facing the following issue:
ValueError: ('The specified size contains a dimension with value <=
0', (-8000, 256))
Below is my code that I am trying to execute.
from __future__ import print_function
import keras
from keras.datasets import cifar10
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Conv3D, MaxPooling3D
from keras.optimizers import SGD
import os
from keras import backend as K
batch_size = 128
num_classes = 10
epochs = 20
learning_rate = 0.01
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
img_rows = x_train.shape[1]
img_cols = x_train.shape[2]
colors = x_train.shape[3]
if K.image_data_format() == 'channels_first':
x_train = x_train.reshape(x_train.shape[0], 1,colors, img_rows, img_cols)
x_test = x_test.reshape(x_test.shape[0], 1,colors, img_rows, img_cols)
input_shape = (1, colors, img_rows, img_cols)
else:
x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, colors, 1)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, colors, 1)
input_shape = (img_rows, img_cols, colors, 1)
# Convert class vectors to binary class matrices.
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
model = Sequential()
model.add(Conv3D(32, kernel_size=(3, 3, 3),activation='relu',input_shape=input_shape))
model.add(Conv3D(32, kernel_size=(3, 3, 3),activation='relu'))
model.add(MaxPooling3D(pool_size=(2, 2, 1)))
model.add(Dropout(0.25))
model.add(Conv3D(64, kernel_size=(3, 3, 3),activation='relu'))
model.add(Conv3D(64, kernel_size=(3, 3, 3),activation='relu'))
model.add(MaxPooling3D(pool_size=(2, 2, 1)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(256, activation='relu'))
model.add(Dense(num_classes, activation='softmax'))
sgd=SGD(lr=learning_rate)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=sgd,
metrics=['accuracy'])
history = model.fit(x_train, y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(x_test, y_test))
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
I have tried with single conv3d and it worked but the accuracy was very low. Code snippet as below:
model = Sequential()
model.add(Conv3D(32, kernel_size=(3, 3, 3),activation='relu',input_shape=input_shape))
model.add(MaxPooling3D(pool_size=(2, 2, 1)))
model.add(Flatten())
model.add(Dense(256, activation='relu'))
model.add(Dense(num_classes, activation='softmax'))

Answer 1

Problem
The problem is with the color channel: it equals 3 initially and you're applying the convolution with size 3 and padding='valid'. After the very first Conv3D, the output tensor is:
(None, 30, 30, 1, 32)
... and no more convolutions can be applied to that dimension. The trivial example that you provide is working simply because there's just one convolutional layer.
Solution
One option for you is to set padding='same', so that the tensor shape is preserved:
(None, 32, 32, 3, 32)
However, to me the convolution over colors doesn't add a lot of value, so I'd go with this model:
model = Sequential()
model.add(Conv3D(32, kernel_size=(3, 3, 1), activation='relu', input_shape=input_shape))
model.add(Conv3D(32, kernel_size=(3, 3, 1), activation='relu'))
model.add(MaxPooling3D(pool_size=(2, 2, 1)))
model.add(Dropout(0.25))
model.add(Conv3D(64, kernel_size=(3, 3, 1), activation='relu'))
model.add(Conv3D(64, kernel_size=(3, 3, 1), activation='relu'))
model.add(MaxPooling3D(pool_size=(2, 2, 1)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(256, activation='relu'))
model.add(Dense(10, activation='softmax'))

Answer 2

In practice,in convolution layer the dimension is preserved and in pooling layer you can down sample.
The problem is you are losing dimensionality here. So you can set padding same or use 3X3 filters with one channel instead of using 3 channels.