Deep compression code
https://github.com/songhan/SqueezeNet-Deep-Compression
import sys
import os
import numpy as np
import pickle help_ = '''
Usage:
decode.py <net.prototxt> <net.binary> <target.caffemodel>
Set variable CAFFE_ROOT as root of caffe before run this demo!
''' if len(sys.argv) != 4:
print help_
sys.exit()
else:
prototxt = sys.argv[1]
net_bin = sys.argv[2]
target = sys.argv[3] # os.system("cd $CAFFE_ROOT")
caffe_root = os.environ["CAFFE_ROOT"]
os.chdir(caffe_root)
print caffe_root
sys.path.insert(0, caffe_root + 'python')
import caffe caffe.set_mode_cpu()
net = caffe.Net(prototxt, caffe.TEST)
layers = filter(lambda x:'conv' in x or 'fc' in x or 'ip' in x, net.params.keys()) fin = open(net_bin, 'rb') def binary_to_net(weights, spm_stream, ind_stream, codebook, num_nz):
bits = np.log2(codebook.size)
if bits == 4:
slots = 2
elif bits == 8:
slots = 1
else:
print "Not impemented,", bits
sys.exit()
code = np.zeros(weights.size, np.uint8) # Recover from binary stream
spm = np.zeros(num_nz, np.uint8)
ind = np.zeros(num_nz, np.uint8)
if slots == 2:
spm[np.arange(0, num_nz, 2)] = spm_stream % (2**4)
spm[np.arange(1, num_nz, 2)] = spm_stream / (2**4)
else:
spm = spm_stream
ind[np.arange(0, num_nz, 2)] = ind_stream% (2**4)
ind[np.arange(1, num_nz, 2)] = ind_stream/ (2**4) # Recover the matrix
ind = np.cumsum(ind+1)-1
code[ind] = spm
data = np.reshape(codebook[code], weights.shape)
np.copyto(weights, data) nz_num = np.fromfile(fin, dtype = np.uint32, count = len(layers))
for idx, layer in enumerate(layers):
print "Reconstruct layer", layer
print "Total Non-zero number:", nz_num[idx]
#eg . Reconstruct layer conv1
#Total Non-zero number: 13902
if 'conv' in layer:
bits = 8 #卷积层使用8bit量化,全连接使用4bit
else:
bits = 4
codebook_size = 2 ** bits #所有码字的总数
codebook = np.fromfile(fin, dtype = np.float32, count = codebook_size)
bias = np.fromfile(fin, dtype = np.float32, count = net.params[layer][1].data.size)
np.copyto(net.params[layer][1].data, bias) #把fin里的值拷贝进去,原先net.params[layer][1].data全部都是0 spm_stream = np.fromfile(fin, dtype = np.uint8, count = (nz_num[idx]-1) / (8/bits) + 1)
ind_stream = np.fromfile(fin, dtype = np.uint8, count = (nz_num[idx]-1) / 2+1) binary_to_net(net.params[layer][0].data, spm_stream, ind_stream, codebook, nz_num[idx]) net.save(target)
Deep compression code的更多相关文章
- [综述]Deep Compression/Acceleration深度压缩/加速/量化
Survey Recent Advances in Efficient Computation of Deep Convolutional Neural Networks, [arxiv '18] A ...
- DEEP COMPRESSION小记
2016ICLR最佳论文 Deep Compression: Compression Deep Neural Networks With Pruning, Trained Quantization A ...
- Deep Compression Compressing Deep Neural Networks With Pruning, Trained QuantizationAnd Huffman Coding
转载请注明出处: http://www.cnblogs.com/sysuzyq/p/6200613.html by 少侠阿朱
- 论文翻译:2021_Towards model compression for deep learning based speech enhancement
论文地址:面向基于深度学习的语音增强模型压缩 论文代码:没开源,鼓励大家去向作者要呀,作者是中国人,在语音增强领域 深耕多年 引用格式:Tan K, Wang D L. Towards model c ...
- A Full Hardware Guide to Deep Learning
A Full Hardware Guide to Deep Learning Deep Learning is very computationally intensive, so you will ...
- 网络压缩论文集(network compression)
Convolutional Neural Networks ImageNet Models Architecture Design Activation Functions Visualization ...
- cs231n spring 2017 lecture15 Efficient Methods and Hardware for Deep Learning 听课笔记
1. 深度学习面临的问题: 1)模型越来越大,很难在移动端部署,也很难网络更新. 2)训练时间越来越长,限制了研究人员的产量. 3)耗能太多,硬件成本昂贵. 解决的方法:联合设计算法和硬件. 计算硬件 ...
- 深度学习网络压缩模型方法总结(model compression)
两派 1. 新的卷机计算方法 这种是直接提出新的卷机计算方式,从而减少参数,达到压缩模型的效果,例如SqueezedNet,mobileNet SqueezeNet: AlexNet-level ac ...
- (zhuan) Where can I start with Deep Learning?
Where can I start with Deep Learning? By Rotek Song, Deep Reinforcement Learning/Robotics/Computer V ...
随机推荐
- 【我所认知的BIOS】—> uEFI AHCI Driver(5) — 第一个protocol最终要開始安装了
[我所认知的BIOS]-> uEFI AHCI Driver(5) - 第一个protocol最终要開始安装了 LightSeed 4/28/2014 文章对EFI_DRIVER_BINDING ...
- git 忽略文件 .gitignore 以及规则
git提供了文件忽略系统,当对工作区某个目录或文件设置了忽略后,在执行status查看状态时,被忽略的文件即使存在也不会显示出来. 这样我就可以把那些不需要上传,不需要保留的文件或目录忽略掉(比如一些 ...
- DeJaVu update history
17.05.08 <-> Added Audi RB8 random code direct change -> Now can adapt VIN based keys or ke ...
- 3D单机游戏《天鹰教》源码发布(二)
3D单机游戏<天鹰教>源码发布 作者 作者: 游蓝海 博客: http://blog.csdn.net/you_lan_hai mail: you_lan_hai@foxmail. ...
- 2008 SCI 影响因子(Impact Factor)
2008 SCI 影响因子(Impact Factor) Excel download 期刊名缩写 影响因子 ISSN号 CA-CANCER J CLIN 74.575 0007-9235 NEW E ...
- 使用EF Model First创建edmx模型,数据库有数据的情况下,如何同时更新模型和数据库
使用"EF Model First",我们可以很直观地在edmx文件中创建数据模型,并根据模型生成数据库,整个过程快速而高效.可当数据库中有了一定的数据量,同时,可能需要对模型中字 ...
- Struts2 校验框架学习笔记
Struts2 校验框架 Struts2 和Struts1同样也提供了校验框架,但在Struts2 已经不再把校验框架做为一个插件,而是已经内置到了Struts2中,而且配置起来更为简单方便,功能也更 ...
- 解决sqoop 导入oracle表时 --split-by参数为日期类型时的报错:ORA-01861: literal does not match format string
报错栈: -- ::, INFO [main] org.apache.sqoop.mapreduce.db.DBRecordReader: Executing query: select " ...
- 四边形优化dp
理解: http://blog.renren.com/share/263498909/1064362501 http://www.cnblogs.com/ronaflx/archive/2011/03 ...
- data-stream-as-disjoint-intervals
https://leetcode.com/problems/data-stream-as-disjoint-intervals/ /** * Definition for an interval. * ...