caffe可视化各层feature map（附python脚本）

通过查看各层feature map可以更好地观察各层的输出，以及各层的效果。

代码：

# -*- coding: utf-8 -*-

import sys
sys.path.append("/data_1/SSD/caffe/python")
import numpy as np
import matplotlib.pyplot as plt
import os
import pickle
import cv2
import Image as img
from skimage import io,transform
caffe_root='/data_1/SSD/caffe'

sys.path.insert(0,caffe_root+'python')
import caffe

deployPrototxt =  '/data_1/SSD/caffe/models/VGGNet/VOC0712/SSD_300x300/deploy.prototxt'
modelFile = '/data_1/SSD/caffe/models/VGGNet/VOC0712/SSD_300x300/VGG_VOC0712_SSD_300x300_iter_18.caffemodel'
meanFile = ''

#网络初始化
def initilize():
    print 'initilize ... '
    sys.path.insert(0, caffe_root + 'python')
    caffe.set_mode_gpu()
    caffe.set_device(0)
    net = caffe.Net(deployPrototxt, modelFile,caffe.TEST)
    return net

# 此处将卷积图和进行显示，
def vis_square(data, padsize=1, padval=0):
    data -= data.min()
    data /= data.max()
    
    print data[0:1].shape

    #让合成图为方
    n = int(np.ceil(np.sqrt(data.shape[0])))
    padding = ((0, n ** 2 - data.shape[0]), (0, padsize), (0, padsize)) + ((0, 0),) * (data.ndim - 3)
    data = np.pad(data, padding, mode='constant', constant_values=(padval, padval))

    #合并卷积图到一个图像中    
    data = data.reshape((n, n) + data.shape[1:]).transpose((0, 2, 1, 3) + tuple(range(4, data.ndim + 1)))
    data = data.reshape((n * data.shape[1], n * data.shape[3]) + data.shape[4:])
    #print data.shape    
    plt.imshow(data)

#def getforward_net(image,net):
    ## input preprocessing: 'data' is the name of the input blob == net.inputs[0]
    #transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
    #transformer.set_transpose('data', (2,0,1))
    ##transformer.set_mean('data', np.load(caffe_root + meanFile ).mean(1).mean(1)) # mean pixel
    ##transformer.set_raw_scale('data', 255)  
    ##the reference model operates on images in [0,255] range instead of [0,1]
    ##transformer.set_channel_swap('data', (2,1,0))  
    ## the reference model has channels in BGR order instead of RGB
    ## set net to batch size of 50
    #net.blobs['data'].reshape(1,3,64,64)
    #net.blobs['data'].reshape(1,1,28,28)

    #net.blobs['data'].data[...] = transformer.preprocess('data',caffe.io.load_image(image,False) )    
    #out = net.forward()


def getforward_net(image,net):
    # input preprocessing: 'data' is the name of the input blob == net.inputs[0]
    transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
    transformer.set_transpose('data', (2,0,1))
    #transformer.set_mean('data', np.load(caffe_root + meanFile ).mean(1).mean(1)) # mean pixel
    transformer.set_raw_scale('data', 255)  
    # the reference model operates on images in [0,255] range instead of [0,1]
    transformer.set_channel_swap('data', (2,1,0))  
    # the reference model has channels in BGR order instead of RGB
    # set net to batch size of 50
    net.blobs['data'].reshape(1,3,300,300)

    net.blobs['data'].data[...] = transformer.preprocess('data', caffe.io.load_image(image))    
    out = net.forward()

#取出网络中的params和net.blobs的中的数据
def getNetDetails(image, net):

    image_data=plt.imread(image)

    plt.figure("Image")
    plt.imshow(image_data)
    
    getforward_net(image,net)    
    #网络提取conv1的卷积核
    """
    filters = net.params['conv1_1'][0].data
    with open('FirstLayerFilter.pickle','wb') as f:
       pickle.dump(filters,f)
    vis_square(filters.transpose(0, 2, 3, 1))
    """ 



    feat = net.blobs['data'].data[0, :3]    
    plt.figure("data")   
    vis_square(feat,padval=1) 

    feat = net.blobs['conv9_1'].data[0, :256]    
    plt.figure("conv9_1")   
    vis_square(feat,padval=1) 
    #vis_square(feat) 

    #feat = net.blobs['conv1_2'].data[0, :20]    
    #plt.figure("conv1_2")   
    #vis_square(feat,padval=1) 


    #feat = net.blobs['conv2_1'].data[0, :20]    
    #plt.figure("conv2_1")   
    #vis_square(feat,padval=1) 

    #feat = net.blobs['conv2_2'].data[0, :20]    
    #plt.figure("conv2_2")   
    #vis_square(feat,padval=1) 
   
    #feat = net.blobs['loc_conv2'].data[0, :32]    
    #plt.figure("loc_conv2")   
    #vis_square(feat,padval=1) 

    #feat = net.blobs['st_output'].data[0]    
    #plt.figure("st_output")   
    #vis_square(feat,padval=1) 

   # feat = net.blobs['conv3'].data[0]    
    #plt.figure("conv3")   
    #vis_square(feat,padval=1) 

    #print 'loc_ip1:\n', net.blobs['loc_ip1'].data
    #print 'theta:\n', net.blobs['theta'].data
    #print 'class:\n', net.blobs['class'].data
    #print 'conv3:\n', net.blobs['conv2'].data
    #print 'feat:\n', net.blobs['feat'].data

    #conv1_1的特征图
    # feat = net.blobs['conv1_1'].data[0, :64]    
    # plt.figure("conv1_1")   
    # vis_square(feat,padval=1) 

    # feat = net.blobs['conv1_2_D'].data[0, :34]    
    # plt.figure("conv1_2_D")   
    # vis_square(feat,padval=1) 

    """
    #conv1_1的特征图
    feat = net.blobs['fc8_1'].data[0, :512]    
    plt.figure("fc8_1")   
    vis_square(feat,padval=1)    
 
    #conv1_1的特征图
    feat = net.blobs['fc8_2'].data[0, :512]    
    plt.figure("fc8_2")   
    vis_square(feat,padval=1)   
    
    #conv1_1的特征图
    feat = net.blobs['fc8_3'].data[0, :512]    
    plt.figure("fc8_3")   
    vis_square(feat,padval=1)   
    
    #conv1_1的特征图
    feat = net.blobs['fc8_4'].data[0, :512]    
    plt.figure("fc8_4")   
    vis_square(feat,padval=1)   

    #conv1_1的特征图
    feat = net.blobs['fc8'].data[0, :512]    
    plt.figure("fc8")   
    vis_square(feat,padval=1)  
    """

    """  
    #conv1_1的特征图
    feat = net.blobs['conv5_2'].data[0, :512]    
    plt.figure("conv5_2")   
    vis_square(feat,padval=1)
    #conv1_1的特征图
    feat = net.blobs['conv5_1'].data[0, :512]    
    plt.figure("conv5_1")   
    vis_square(feat,padval=1)    
       
   
   
   
    #conv1_1的特征图
    feat = net.blobs['conv5_3'].data[0, :512]    
    plt.figure("conv5_3")   
    vis_square(feat,padval=1)    
    
    #conv1_1的特征图
    feat = net.blobs['conv5_2'].data[0, :512]    
    plt.figure("conv5_2")   
    vis_square(feat,padval=1)
    #conv1_1的特征图
    feat = net.blobs['conv5_1'].data[0, :512]    
    plt.figure("conv5_1")   
    vis_square(feat,padval=1)    
    
    #conv1_1的特征图
    feat = net.blobs['conv4_3'].data[0, :512]    
    plt.figure("conv4_3")   
    vis_square(feat,padval=1)    
    
    #conv1_1的特征图
    feat = net.blobs['conv4_2'].data[0, :512]    
    plt.figure("conv4_2")   
    vis_square(feat,padval=1)   
   
    #conv1_1的特征图
    feat = net.blobs['conv4_1'].data[0, :512]    
    plt.figure("conv4_1")   
    vis_square(feat,padval=1)
    
    #conv1_1的特征图
    feat = net.blobs['conv3_3'].data[0, :256]    
    plt.figure("conv3_3")   
    vis_square(feat,padval=1)
    
    #conv1_1的特征图
    feat = net.blobs['conv3_2'].data[0, :256]    
    plt.figure("conv3_2")   
    vis_square(feat,padval=1)
    
    #conv1_1的特征图
    feat = net.blobs['conv3_1'].data[0, :256]    
    plt.figure("conv3_1")   
    vis_square(feat,padval=1)
    
    #conv1_1的特征图
    feat = net.blobs['conv2_2'].data[0, :128]    
    plt.figure("conv2_2")   
    vis_square(feat,padval=1)
    
    #conv1_1的特征图
    feat = net.blobs['conv2_1'].data[0, :128]    
    plt.figure("conv2_1")   
    vis_square(feat,padval=1)

     #conv1_1的特征图
    feat = net.blobs['conv1_2'].data[0, :64]    
    plt.figure("conv1_2")   
    vis_square(feat,padval=1)

    #conv1_1的特征图    
    feat = net.blobs['conv1_1'].data[0, :64]    
    plt.figure("conv1_1")   
    vis_square(feat,padval=1)
    """
    plt.show()

if __name__ == "__main__":
    net = initilize()
    testimage = '/data_1/SSD/caffe/examples/images/cat.jpg'
    getNetDetails(testimage, net)

需要修改的地方： 

1、sys.path.append，修改为自己的python路径（本程序已SSD为例，则添加至SSD路径下python文件夹）;

2、caffe_root，修改为SSD根目录;

3、deployPrototxt;

4、modelFile;（模型路径）

5、net.blobs['data'].reshape(1,3,300,300)   ，网络输入图像尺寸（此处为SSD300*300）;

6、feat = net.blobs['conv9_1'].data[0, :256]   ， 输出特征图的名称（256为通道数）;

7、testimage;

ok，运行脚本即可查看了。

 

原图像（孤独无助可怜的猫）

RGB三通道图像

SSD-CONV2_1 

SSD-CONV9_1 

 

从特征图conv2_1看出语义信息还不够充分，到了conv9_1语义信息充分提取。