pytorch 可视化feature map

admin管理员组

文章数量:1650768

之前做的一些项目中涉及到feature map 可视化的问题，一个层中feature map的数量往往就是当前层out_channels的值，我们可以通过以下代码可视化自己网络中某层的feature map，个人感觉可视化feature map对调参还是很有用的。不多说了，直接看代码：

import torch
from torch.autograd import Variable
import torch.nn as nn
import pickle

from sys import path
path.append('/residual model path')
import residual_model
from residual_model import Residual_Model

model = Residual_Model()
model.load_state_dict(torch.load('./model.pkl'))



class myNet(nn.Module):
    def __init__(self,pretrained_model,layers):
        super(myNet,self).__init__()
        self1 = nn.Sequential(*list(pretrained_model.children())[:layers[0]])
        self2 = nn.Sequential(*list(pretrained_model.children())[:layers[1]])
        self3 = nn.Sequential(*list(pretrained_model.children())[:layers[2]])

    def forward(self,x):
        out1 = self1(x)
        out2 = self(out1)
        out3 = self(out2)
        return out1,out2,out3

def get_features(pretrained_model, x, layers = [3, 4, 9]):  ## get_features 其实很简单
'''
1.首先import model 
2.将weights load 进model
3.熟悉model的每一层的位置，提前知道要输出feature map的网络层是处于网络的那一层
4.直接将test_x输入网络，*list(model.chidren())是用来提取网络的每一层的结构的。net1 = nn.Sequential(*list(pretrained_model.children())[:layers[0]]) ,就是第三层前的所有层。

'''
    net1 = nn.Sequential(*list(pretrained_model.children())[:layers[0]])  
#   print net1  
    out1 = net1(x)  

    net2 = nn.Sequential(*list(pretrained_model.children())[layers[0]:layers[1]])  
#   print net2  
    out2 = net2(out1)  

    #net3 = nn.Sequential(*list(pretrained_model.children())[layers[1]:layers[2]])  
    #out3 = net3(out2)  

    return out1, out2
with open('test.pickle','rb') as f:
    data = pickle.load(f)
x = data['test_mains'][0]
x = Variable(torch.from_numpy(x)).view(1,1,128,1)  ## test_x必须为Varibable
#x = Variable(torch.randn(1,1,128,1))
if torch.cuda.is_available():
    x = x.cuda()  # 如果模型的训练是用cuda加速的话，输入的变量也必须是cuda加速的，两个必须是对应的，网络的参数weight都是用cuda加速的，不然会报错
    model = model.cuda()
output1,output2 = get_features(model,x)## model是训练好的model,前面已经import 进来了Residual model
print('output1.shape:',output1.shape)
print('output2.shape:',output2.shape)
#print('output3.shape:',output3.shape)
output_1 = torch.squeeze(output2,dim = 0)
output_1_arr = output_1.data.cpu().numpy() # 得到的cuda加速的输出不能直接转变成numpy格式的，当时根据报错的信息首先将变量转换为cpu的，然后转换为numpy的格式
output_1_arr = output_1_arr.reshape([output_1_arr.shape[0],output_1_arr.shape[1]])

本文标签： Pytorchfeaturemap