DL之CNN可视化：利用SimpleConvNet算法【3层,im2col优化】基于mnist数据集训练并对卷积层输出进行可视化

输出结果

设计思路

核心代码

def filter_show(filters, suptitle, nx=8, margin=3, scale=10,):

   FN, C, FH, FW = filters.shape  

   ny = int(np.ceil(FN / nx))      

   fig = plt.figure()               fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)

   for i in range(FN):          

       ax = fig.add_subplot(ny, nx, i+1, xticks=[], yticks=[])

       ax.imshow(filters[i, 0], cmap=plt.cm.gray_r, interpolation='nearest')

   plt.suptitle(suptitle)

   plt.show()

filter_show(network.params['W1'],suptitle='CNN(SimpleConvNet): Weight after random initialization on mnist')

network.load_params("params.pkl")          

filter_show(network.params['W1'],suptitle='CNN(SimpleConvNet): Weight after learning on mnist(params.pkl)')