我需要一些关于张量流中分离 3D 卷积的帮助

Question

I am student and studying Deep Neural Network.

I saw one paper which is titled:

" 3D Depthwise Convolution: Reducing Model Parameters in 3D Vision Tasks "

I need to implement the 3D Separable Convolution introduced in this paper.

I want you to point out my sources.

separable_conv3d.py

def separable_conv3d(input, output_dim, depth_filter_channel=1, strides=1, padding='SAME', name=None):
    input_tensor_channel = input.get_shape().as_list()[-1]
    kernel1 = tf.Variable(tf.truncated_normal(shape=[batch_frame, 3, 3, input_tensor_channel, depth_filter_channel], stddev=0.1))
    feature_list = []
    for c in range(0, len(input_tensor_channel)):
        feature1 = conv3d(input.shape[:, :, c], weight=kernel1, strides=strides, padding=padding, name=name)
        feature_list.append(feature1)

    total_feature = tf.concat(feature_list, axis=-1)
    total_tensor_channel = total_feature.get_shape().as_list()[-1]
    kernel3 = tf.Variable(tf.truncated_normal(shape=[1, 1, 1, total_tensor_channel, output_dim]))
    pw1 = conv3d(input=total_feature, weight=kernel3, strides=strides, padding=padding, name=name)

return pw1

This is the image I referenced.

Any criticism is welcome. This is REAL.

Answer 1

I did not test your code but as I understood while selecting features in the for loop you use 3 dimensions and put c in to 3rd dimension, but as far as I know 3D convolution is 4D so it would be correct if you add additional dimension. I will be testing and editing your code soon.