conv2d_transpose vs conv2d

conv2d.py

# coding: utf-8

'''
conv2d: weight(kernel_size,kernel_size,in_channel,out_channel)
conv2d_transpose: weight(kernel_size,kernel_size,out_channel,in_channel)

'''


import numpy as np
import tensorflow as tf


import matplotlib
 
import matplotlib.pyplot as plt
tf.reset_default_graph()

def without_strides():
    img_size=5
    in_channel=4
    out_channel=2
    kernel_size=2
    
    
    x = np.random.randn(1,img_size,img_size,in_channel).astype(np.float32)
    
    x_pad = np.pad(x,((0,0),(1,1),(1,1),(0,0)),constant_values=0.0)
    
    w = np.random.randn(kernel_size,kernel_size,out_channel,in_channel).astype(np.float32)
    
    xx = tf.convert_to_tensor(x)
    xx_pad = tf.convert_to_tensor(x_pad)
    
    
    
    w_flip = np.transpose(w,(0,1,3,2))
    w_flip_stack = []
    for i in range(in_channel):
        w_flip_stack.append(np.flip(w_flip[:,:,i:i+1,:],(0,1)))
    
    w_flip  = np.concatenate(w_flip_stack,axis=2)
    
    print("w.shape",w.shape,"w_flip.shape",w_flip.shape)
    
    
    
    y = tf.layers.conv2d(xx_pad,filters=out_channel,kernel_size=kernel_size,kernel_initializer=tf.constant_initializer(w_flip),use_bias=False,padding='valid')
    yy = tf.layers.conv2d_transpose(xx,filters=out_channel,kernel_size=kernel_size,kernel_initializer=tf.constant_initializer(w),use_bias=False,padding='valid')
    
    
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    
    print(y.shape,yy.shape)
    
    y_,yy_ = sess.run([y,yy])
    # print("y",y_)
    # print("yy",yy_)
    # print("diff",y_-yy_)
    
    print(np.allclose(y_,yy_))


def with_strides():
    img_size=5
    in_channel=4
    out_channel=2
    kernel_size=2
    strides = 2 # don't change.
    
    x = np.random.randn(1,img_size,img_size,in_channel).astype(np.float32)
    
    x_pad = np.insert(x, range(img_size+1), 0, axis=1)
    x_pad = np.insert(x_pad, range(img_size+1), 0, axis=2)
    
    w = np.random.randn(kernel_size,kernel_size,out_channel,in_channel).astype(np.float32)
    
    xx = tf.convert_to_tensor(x)
    xx_pad = tf.convert_to_tensor(x_pad)
    
    
    
    w_flip = np.transpose(w,(0,1,3,2))
    w_flip_stack = []
    for i in range(in_channel):
        w_flip_stack.append(np.flip(w_flip[:,:,i:i+1,:],(0,1)))
    
    w_flip  = np.concatenate(w_flip_stack,axis=2)
    
    print("w.shape",w.shape,"w_flip.shape",w_flip.shape)
    
    
    
    y = tf.layers.conv2d(xx_pad,filters=out_channel,kernel_size=kernel_size,kernel_initializer=tf.constant_initializer(w_flip),use_bias=False,padding='valid')
    yy = tf.layers.conv2d_transpose(xx,filters=out_channel,kernel_size=kernel_size,strides=strides,kernel_initializer=tf.constant_initializer(w),use_bias=False,padding='valid')
    
    
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    
    print(y.shape,yy.shape)
    
    y_,yy_ = sess.run([y,yy])
    # print("y",y_)
    # print("yy",yy_)
    # print("diff",y_-yy_)
    
    print(np.allclose(y_,yy_))

https://stackoverflow.com/questions/39373230/what-does-tensorflows-conv2d-transpose-operation-do

tensor upsampling using zeros in between values of a tensor

https://stackoverflow.com/questions/45837419/alternative-to-np-insert-in-tensorflow

img_size=5
a = np.random.randn(2,img_size,img_size,3)
b=np.insert(a, range(img_size+1), 0, axis=1)
b=np.insert(b, range(img_size+1), 0, axis=2)
print(a)
print(b)