直接稀疏

对一些model可以直接做稀疏压缩也不掉点

相应的代码为

'''
author :dezan zhao
contact:1183489276
date   :2018-02-23
func   :prune caffemodel file directly
'''
#-*-coding:utf-8-*-
import pdb
import caffe
import re
import numpy as np
if __name__ == '__main__':
#1.读取：读取caffemodel
	MODEL_FILE ='/data1/zhaodz/caffe-master/examples/mnist_cmp/lenet_train_test_compress_stage1_deploy.prototxt'
	PRETRAIN_FILE = '/data1/zhaodz/caffe-master/examples/mnist_cmp/models/lenet_finetune_stage1_iter_500.caffemodel'
	#PRETRAIN_FILE = '/data1/zhaodz/caffe-master/examples/mnist_cmp/models/lenet_iter_10000.caffemodel'
	net = caffe.Net(MODEL_FILE, PRETRAIN_FILE, caffe.TEST)
	sparse_ratio_vec = [0.33, 0.8, 0.9, 0.8] 
	num = 0
	for index in range(len(net.layers)):
		#if re.match('conv.',param_name) or re.match('fc.',param_name):
		if re.match('.*?Convolution',net.layers[index].type) or re.match('.*?InnerProduct',net.layers[index].type):
			param_name = net._layer_names[index] #层的名字
			print param_name
			weight = net.params[param_name][0].data
			shape = weight.shape
			
			#2.处理：使权重按照近似公式进行运算
			#weight.shape = (-1, 1)
			weight = weight.flatten()
			count = len(weight)
			sort_weight = np.zeros(count)
			#pdb.set_trace()
			for i in range(count):
				sort_weight[i] = abs(weight[i])
			sort_weight.sort()
			ratio = sparse_ratio_vec[num];
			index = int(count*ratio)
			mask_data = np.zeros(count)
			muweight = np.zeros(count)
			thr = sort_weight[index - 1]
			#pdb.set_trace()
			for i in range(count):
				if weight[i] > thr or weight[i] < -thr:
					mask_data[i] = 1
				else:
					mask_data[i] = 0
				muweight[i] = weight[i] * mask_data[i]
			#pdb.set_trace()
			muweight = muweight.reshape(shape)
			net.params[param_name][0].data[...] = muweight
			num = num + 1
	net.save('/data1/zhaodz/caffe-master/examples/mnist_cmp/mnist_cmp.caffemodel')

depwise的直接稀疏为

#-*-coding:utf-8-*-
import pdb
import caffe
import re
import numpy as np
if __name__ == '__main__':
#1.读取：读取caffemodel
	MODEL_FILE ='deploy.prototxt'
	PRETRAIN_FILE = './models1/mobilenet_iter_1000000.caffemodel'
	#PRETRAIN_FILE = '/data1/zhaodz/caffe-master/examples/mnist_cmp/models/lenet_iter_10000.caffemodel'
	net = caffe.Net(MODEL_FILE, PRETRAIN_FILE, caffe.TEST)
	sparse_ratio_vec = [0.33, 0.8, 0.9, 0.8] 
	num = 0
	for index in range(len(net.layers)):
		print net.layers[index].type
		#if re.match('conv.',param_name) or re.match('fc.',param_name):
		if net.layers[index].type == 'Convolution' or net.layers[index].type=='InnerProduct':
			param_name = net._layer_names[index] #层的名字
			print param_name
			weight = net.params[param_name][0].data
			shape = weight.shape
			#pdb.set_trace()
			#2.处理：使权重按照近似公式进行运算
			#weight.shape = (-1, 1)
			weight = weight.flatten()
			count = len(weight)
			sort_weight = np.zeros(count)
			#pdb.set_trace()
			for i in range(count):
				sort_weight[i] = abs(weight[i])
			sort_weight.sort()
			#ratio = sparse_ratio_vec[num];
			ratio = 0.5
			index_tmp = int(count*ratio)
			mask_data = np.zeros(count)
			muweight = np.zeros(count)
			thr = sort_weight[index_tmp - 1]
			#pdb.set_trace()
			for i in range(count):
				if weight[i] > thr or weight[i] < -thr:
					mask_data[i] = 1
				else:
					mask_data[i] = 0
				muweight[i] = weight[i] * mask_data[i]
			#pdb.set_trace()
			muweight = muweight.reshape(shape)
			net.params[param_name][0].data[...] = muweight
			num = num + 1
		elif net.layers[index].type == 'ConvolutionDepthwise' or net.layers[index].type == 'InnerProductDepthwise':
			param_name = net._layer_names[index] #层的名字
			print param_name
			weight = net.params[param_name][0].data
			for i in range(len(weight)):
				shape = weight[i][0].shape
				weight_k = weight[i][0].flatten()
				count = len(weight_k)
				#pdb.set_trace()
				sort_weight = np.zeros(count)
				for j in range(count):
					sort_weight[j] = abs(weight_k[j])
				sort_weight.sort()
				ratio = 0.3
				index_tmp = int(count*ratio)
				mask_data = np.zeros(count)
				muweight = np.zeros(count)
				thr = sort_weight[index_tmp - 1]
				#pdb.set_trace()
				for j in range(count):
					if weight_k[j] > thr or weight_k[j] < -thr:
						mask_data[j] = 1
					else:
						mask_data[j] = 0
					muweight[j] = weight_k[j] * mask_data[j]
				muweight = muweight.reshape(shape)
				net.params[param_name][0].data[i][0] = muweight
	net.save('cmp.caffemodel')