Pytorch data input format requirements and conversion

Data format requirements

The 4 dimensions of input_patch are <batch size, image height, image width, image channel> respectively. In Pytorch, the input channel should be in the second dimension. That’s why the permutation is required.
After the permutation, the 4 dimensions of in_img will be <batch size, image channel, image height, image width>¹.

take numpy Format image into corresponding tensor Format

Before feeding into the deep learning network , The data format needs to be converted , Need to put N x H x W x C Of numpy Format image into corresponding tensor Format N x C x H x W²：

def toTensor(img):
    img = torch.from_numpy(img.transpose((0, 3, 1, 2)))
    return img.float().div(255).unsqueeze(0)

Format conversion of input data as required

Use .permute() Function to adjust the dimensions of data , And use .unsqueeze() Function to increase its dimension , The code and demonstration are as follows ：

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# @Time : 2022/7/12 12:26
# @FileName: unsqueeze.py
# @Software: PyCharm
import torch

x = torch.rand(8,128,192)

print(x.shape)     # torch.Size([8, 128, 192])

x1 = torch.unsqueeze(x, -1)
print(x1.shape)      # torch.Size([8, 128, 192, 1])

print(x1[0].shape)    # torch.Size([128, 192, 1])
print(x1[1].shape)    # torch.Size([128, 192, 1])
print(x1[-1].shape)   # torch.Size([128, 192, 1])

---

If my article helps you , Then I'll be happy too , How far can one go , It's about who to walk with .

---

Reference resources

---

1. Why permute (0,3,1,2) is required? #1︎

2. torch Various image format conversion in ︎