遷移pytorch工程至matlab

好久沒寫博客了，險些以為自己找不到密碼了。

最近抽空參與了個小項目，很慚愧，只做了三件小事

1. 基於PyTorch訓練了一系列單圖像超分辨神經網絡

 

基於PyTorch訓練了一系列單圖像超分辨神經網絡，超分辨系數從2-10。
該部分的實現參考了pytorch官方repo中的SR例程，訓練程序包含於`./train`文件夾。該項目
基於高效子像素卷積層[1]進行空間分辨率提升操作，訓練速度極快。

[1] ["Shi W, Caballero J, Huszar F, et al. Real-Time Single Image and
    Video Super-Resolution Using an Efficient Sub-Pixel Convolutional
    Neural Network[J]. 2016:1874-1883.](https://arxiv.org/abs/1609.05158)

 

2. 把訓練好的模型權值轉存為MATLAB文件。

簡單粗暴，異常直接，只要把對應卷積層的權值全部提取出來就可以了。

提取的時候注意一點，要把pytorch中的Variable格式轉換為Tensor，再轉換為CPU模式，最終轉換為numpy數組。

這一系列過程合並起來就是：

Var.data.cpu().numpy()

具體實現如下：

from __future__ import print_function

import torch
import numpy as np
import scipy.io as sio

for i in [2, 3, 4, 5, 6, 7, 8, 9, 10]:

    model_name = 'model_upscale_{}_epoch_101.pth'.format(i)
    model = torch.load(model_name)
    print(model._modules)

    weight = dict()
    weight['conv1_w'] = model._modules['conv1']._parameters['weight'].data.cpu().numpy()
    weight['conv2_w'] = model._modules['conv2']._parameters['weight'].data.cpu().numpy()
    weight['conv3_w'] = model._modules['conv3']._parameters['weight'].data.cpu().numpy()
    weight['conv4_w'] = model._modules['conv4']._parameters['weight'].data.cpu().numpy()

    weight['conv1_b'] = model._modules['conv1']._parameters['bias'].data.cpu().numpy()
    weight['conv2_b'] = model._modules['conv2']._parameters['bias'].data.cpu().numpy()
    weight['conv3_b'] = model._modules['conv3']._parameters['bias'].data.cpu().numpy()
    weight['conv4_b'] = model._modules['conv4']._parameters['bias'].data.cpu().numpy()

    sio.savemat('model_upscale_{}.mat'.format(i), mdict=weight)

 

3. 把網絡的test過程移植到了MATLAB平台，並撰寫了測試代碼。

把卷積層和pixelshuffle層用matlab重寫了一下。

復現pixelshuffle層的時候遇到了一些麻煩，又回頭看了下pytorch里的測試代碼

`https://github.com/pytorch/pytorch/blob/master/test/test_nn.py `

# https://github.com/pytorch/pytorch/blob/master/test/test_nn.py
def _verify_pixel_shuffle(self, input, output, upscale_factor):
    for c in range(output.size(1)):
        for h in range(output.size(2)):
            for w in range(output.size(3)):
                height_idx = h // upscale_factor
                weight_idx = w // upscale_factor
                channel_idx = (upscale_factor * (h % upscale_factor)) + (w % upscale_factor) + \
                              (c * upscale_factor ** 2)
    self.assertEqual(output[:, c, h, w], input[:, channel_idx, height_idx, weight_idx])

理了理思路，改寫成MATLAB代碼：

function [ outputs ] = PixelShuffle( inputs, upscale_factor )
%    PixelShuffle :
%
%   input : N, upscale_factor ** 2, H, W
%   output : N, 1, H*upscale_factor, W*upscale_factor

[N, ~, H, W] = size(inputs);
H_out = H*upscale_factor;
W_out = W*upscale_factor;
outputs = zeros([N, 1, H_out, W_out]);
for i = 1:N
    for h = 1: H_out
        for w = 1:W_out
            height_idx = floor(h / upscale_factor+0.5);
            weight_idx = floor(w / upscale_factor+0.5);
            channel_idx = (upscale_factor * mod(h-1, upscale_factor)) + mod(w-1, upscale_factor)+1;
            outputs(i, 1, h, w) = inputs(i, channel_idx, height_idx, weight_idx);
        end
    end
end
end

4. 完整工程github鏈接。

https://github.com/JiJingYu/super-resolution-by-subpixel-convolution

模型權值已保存為matlab權值，直接在matlab中運行`demo.m`文件即可驗證