pytorch与tensorflow的卷积差别

无论是tensorflow还是pytorch的卷积操作在一般情况的padding都是正常的及两边上下补，只有当两边补除不尽的时候才会有不同，
个人测试：

如果pading=1,一般为两边都补充0即：
torch：padding=1 == tf：paddings=[[0, 0], [1, 1], [1, 1], [0, 0]]

如果padding=1 步长为2除不尽时，他就会只在上左进行填充：
torch：padding=1 == tf：padding=[[0, 0], [1, 0], [1, 0], [0, 0]]

如果padding=1 步长为2除不尽时，
torch：padding=1 == tf：padding=[[0, 0], [2, 1], [2,1], [0, 0]]

下面有完整代码，大家可以试一下，如果我的结论片面，请大家及时评论

在这里插入代码片
# 测试tensorflow 与torch 卷积的不同
import torch
import torch.nn.functional as F
import torchvision.transforms as transforms
import tensorflow as tf
import numpy as np
from tensorflow.keras import initializers
 
 
def model_th(x, kernel_init):
    print(x.shape)
    print(kernel_init.shape)
    x = F.conv2d(x, weight=kernel_init, stride=2, padding=1)
    print(x.shape)
    return x
 
 
def model_tf(k_init):
    kernel_init = initializers.Constant(k_init)
 
    x0 = tf.keras.layers.Input(shape=[6, 6, 1])

    x = tf.pad(x0, paddings=[[0, 0], [1, 0], [1, 0], [0, 0]])
    x = tf.keras.layers.Conv2D(filters=32, kernel_size=3, strides=2, padding='valid', use_bias=False, kernel_initializer=kernel_init)(x)
 
    out = tf.keras.models.Model(inputs=x0, outputs=x)
 
    return out


def model_tf_nopad(k_init):
    kernel_init = initializers.Constant(k_init)

    x0 = tf.keras.layers.Input(shape=[6, 6, 1])

    #x = tf.pad(x0, paddings=[[0, 0], [1, 0], [1, 0], [0, 0]])
    x = tf.keras.layers.Conv2D(filters=32, kernel_size=3, strides=1, padding='same', use_bias=False,
                               kernel_initializer=kernel_init)(x0)

    out = tf.keras.models.Model(inputs=x0, outputs=x)

    return out
if __name__ == '__main__':
    img = np.random.uniform(0, 255, [6, 6, 1])
 
    kernel = np.random.uniform(-5, 5, [32, 1, 3, 3])
 
    tf_kernel = kernel.transpose([2, 3, 1, 0])
    th_kernel = torch.from_numpy(kernel).float()
 
    totensor = transforms.ToTensor()
    th_img = totensor(img).unsqueeze(0).float()
    tf_img = tf.expand_dims(img, axis=0)
    tf_img = tf.cast(tf_img, tf.float32)
 
    th_out = model_th(th_img, th_kernel)
    tf_out = model_tf(tf_kernel)(tf_img)
    tf_out_nopad = model_tf_nopad(tf_kernel)(tf_img)
    print('--------------out-------------------')

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        tf_out = sess.run(tf_out)
        tf_out_nopad = sess.run(tf_out_nopad)
        print('tf_out', tf_out.shape, tf_out)
        print('th_out.permute([0, 2, 3, 1])', th_out.permute([0, 2, 3, 1]).shape, th_out.permute([0, 2, 3, 1]))
        #print('valu',tf_out - th_out.permute([0, 2, 3, 1]).numpy())
        print('tf_out_nopad',tf_out_nopad.shape, tf_out_nopad)