使用pytorch实现可视化中间层的结果

摘要

一直比较想知道图片经过卷积之后中间层的结果，于是使用pytorch写了一个脚本查看，先看效果

这是原图，随便从网上下载的一张大概224*224大小的图片，如下

网络介绍

我们使用的VGG16，包含RULE层总共有30层可以可视化的结果，我们把这30层分别保存在30个文件夹中，每个文件中根据特征的大小保存了64~128张图片

结果如下：

原图大小为224224，经过第一层后大小为64224*224，下面是第一层可视化的结果，总共有64张这样的图片：

下面看看第六层的结果

这层的输出大小是 1128112*112，总共有128张这样的图片

下面是完整的代码

import cv2
import numpy as np
import torch
from torch.autograd import Variable
from torchvision import models

#创建30个文件夹
def mkdir(path): # 判断是否存在指定文件夹，不存在则创建
 # 引入模块
 import os

# 去除首位空格
 path = path.strip()
 # 去除尾部 \ 符号
 path = path.rstrip("\\")

# 判断路径是否存在
 # 存在   True
 # 不存在  False
 isExists = os.path.exists(path)

# 判断结果
 if not isExists:
   # 如果不存在则创建目录
   # 创建目录操作函数
   os.makedirs(path)
   return True
 else:

return False

def preprocess_image(cv2im, resize_im=True):
 """
   Processes image for CNNs

Args:
   PIL_img (PIL_img): Image to process
   resize_im (bool): Resize to 224 or not
 returns:
   im_as_var (Pytorch variable): Variable that contains processed float tensor
 """
 # mean and std list for channels (Imagenet)
 mean = [0.485, 0.456, 0.406]
 std = [0.229, 0.224, 0.225]
 # Resize image
 if resize_im:
   cv2im = cv2.resize(cv2im, (224, 224))
 im_as_arr = np.float32(cv2im)
 im_as_arr = np.ascontiguousarray(im_as_arr[..., ::-1])
 im_as_arr = im_as_arr.transpose(2, 0, 1) # Convert array to D,W,H
 # Normalize the channels
 for channel, _ in enumerate(im_as_arr):
   im_as_arr[channel] /= 255
   im_as_arr[channel] -= mean[channel]
   im_as_arr[channel] /= std[channel]
 # Convert to float tensor
 im_as_ten = torch.from_numpy(im_as_arr).float()
 # Add one more channel to the beginning. Tensor shape = 1,3,224,224
 im_as_ten.unsqueeze_(0)
 # Convert to Pytorch variable
 im_as_var = Variable(im_as_ten, requires_grad=True)
 return im_as_var

class FeatureVisualization():
 def __init__(self,img_path,selected_layer):
   self.img_path=img_path
   self.selected_layer=selected_layer
   self.pretrained_model = models.vgg16(pretrained=True).features
   #print( self.pretrained_model)
 def process_image(self):
   img=cv2.imread(self.img_path)
   img=preprocess_image(img)
   return img

def get_feature(self):
   # input = Variable(torch.randn(1, 3, 224, 224))
   input=self.process_image()
   print("input shape",input.shape)
   x=input
   for index,layer in enumerate(self.pretrained_model):
     #print(index)
     #print(layer)
     x=layer(x)
     if (index == self.selected_layer):
       return x

def get_single_feature(self):
   features=self.get_feature()
   print("features.shape",features.shape)
   feature=features[:,0,:,:]
   print(feature.shape)
   feature=feature.view(feature.shape[1],feature.shape[2])
   print(feature.shape)
   return features

def save_feature_to_img(self):
   #to numpy
   features=self.get_single_feature()
   for i in range(features.shape[1]):
     feature = features[:, i, :, :]
     feature = feature.view(feature.shape[1], feature.shape[2])
     feature = feature.data.numpy()
     # use sigmod to [0,1]
     feature = 1.0 / (1 + np.exp(-1 * feature))
     # to [0,255]
     feature = np.round(feature * 255)
     print(feature[0])
     mkdir('./feature/' + str(self.selected_layer))
     cv2.imwrite('./feature/'+ str( self.selected_layer)+'/' +str(i)+'.jpg', feature)
if __name__=='__main__':
 # get class
 for k in range(30):
   myClass=FeatureVisualization('/home/lqy/examples/TRP.PNG',k)
   print (myClass.pretrained_model)
   myClass.save_feature_to_img()

来源：https://blog.csdn.net/xz1308579340/article/details/85622579