深度学习框架性能评测（Caffe、CNTK、MXNet、TensorFlow、Torch）

最新推荐文章于 2023-03-09 17:39:17 发布
最新推荐文章于 2023-03-09 17:39:17 发布
阅读量2.3k 收藏
点赞数
分类专栏： DeepLearning
From: http://dlbench.comp.hkbu.edu.hk/
The source code and experimental data of

Benchmarking State-of-the-Art Deep Learning Software Tools

(Version 7, 17 Feb 2017)

Declaration: The runtime performance of each software tool depends not only on the hardware platform, but also on the third-party libraries and the network configuration files. Our results only reflect the performance of the tested networks with the associated configuration files and the specified third-party libraries on our testing machines, which are not necessarily the best performance that can be achieved by the software tool.
Tested Neural Networks
Tested Hardware
Computational UnitCoresMemoryOSCUDA
Intel CPU i7-3820464 GBUbuntu 14.04-
Intel CPU E5-2630x216128 GBCentOS 7.2-
GTX 98020484 GBUbuntu 14.048.0
GTX 108025608 GBUbuntu 14.048.0
Telsa K80249612 GBCentOS 7.28.0
Tested Hardware For Data Parallelization
GPUsCPUMemoryPCIeOSCUDA
K80x2E5-2630v4128 GBPCIe 3.0CentOS 7.28.0
Source Code

Fork me on Github: https://github.com/hclhkbu/dlbench at commit ID: ceac0ef

Acknowledgements
Previous versions (obsoleted)
Results

Item in cell: batchTime(totalTime)

fcn5 on i7-3820
i7-3820: Desktop CPU, with 4 physical cores Network: fcn5
#ThreadsCaffeCNTKMXNETTensorFlowTorch
12475.615ms(652.981s)1844.614ms(458.313s)1913.749ms(450.409s)2646.683ms(658.836s)1670.048ms(422.817s)
21498.604ms(402.902s)969.866ms(241.767s)1071.735ms(252.738s)1912.837ms(469.690s)926.492ms(236.096s)
41148.857ms(313.858s)660.524ms(168.645s)719.155ms(179.260s)1157.003ms(285.295s)565.190ms(145.938s)
8--571.499ms(158.367s)702.088ms(167.439s)918.880ms(227.657s)611.271ms(157.450s)
fcn5 on E5-2630v3
E5-2630v3: Server CPU, with 8x2 physical cores Network: fcn5
#ThreadsCaffeCNTKMXNETTensorFlowTorch
12281.508ms(615.700s)1592.251ms(405.787s)1608.602ms(378.418s)3409.853ms(857.297s)1378.898ms(350.399s)
21748.154ms(489.594s)857.496ms(222.066s)1064.738ms(252.231s)2540.506ms(630.329s)914.813ms(236.167s)
41403.408ms(393.114s)501.030ms(137.932s)731.204ms(174.141s)1296.658ms(325.172s)661.956ms(171.529s)
81210.973ms(342.969s)322.606ms(105.185s)534.293ms(127.368s)661.391ms(167.726s)440.350ms(116.364s)
161127.263ms(321.311s)251.620ms(89.187s)451.452ms(107.704s)360.799ms(93.728s)402.409ms(107.097s)
321126.975ms(342.757s)280.208ms(96.027s)446.716ms(106.671s)325.280ms(85.651s)365.679ms(97.311s)
fcn5 on GTX980
GTX 980, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: fcn5
BatchSizeCaffeCNTKMXNETTensorFlowTorch
34211.643ms(89.084s)11.650ms(121.637s)15.256ms(109.056s)18.669ms(488.123s)11.874ms(91.716s)
51214.469ms(76.241s)14.276ms(105.928s)18.012ms(86.568s)17.570ms(89.108s)14.441ms(75.455s)
102425.449ms(68.553s)24.593ms(96.078s)28.996ms(70.072s)33.120ms(703.220s)24.963ms(65.848s)
204846.320ms(64.350s)45.323ms(94.793s)49.996ms(60.625s)61.083ms(414.290s)45.160ms(60.471s)
409689.306ms(63.177s)88.566ms(91.015s)90.875ms(56.701s)108.289ms(71.178s)85.051ms(58.942s)
fcn5 on GTX1080
GTX 1080, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: fcn5
BatchSizeCaffeCNTKMXNETTensorFlowTorch
3428.015ms(64.207s)7.943ms(90.240s)10.676ms(77.109s)11.000ms(626.976s)8.114ms(63.420s)
5129.872ms(54.295s)9.930ms(81.874s)12.810ms(62.343s)14.124ms(70.703s)9.838ms(52.520s)
102417.142ms(48.717s)16.932ms(74.423s)19.294ms(47.371s)19.948ms(51.193s)16.185ms(43.982s)
204831.670ms(47.159s)30.223ms(71.022s)33.159ms(40.966s)35.375ms(46.479s)28.773ms(39.511s)
409659.106ms(46.399s)58.517ms(72.789s)59.402ms(36.959s)66.267ms(51.311s)54.813ms(43.070s)
alexnet on i7-3820
i7-3820: Desktop CPU, with 4 physical cores Network: alexnet
#ThreadsCaffeCNTKMXNETTensorFlowTorch
13558.355ms(367.519s)9955.847ms(1004.176s)13400.545ms(1310.940s)4534.767ms(446.024s)5351.893ms(563.500s)
22587.100ms(266.949s)7262.622ms(723.156s)12304.886ms(1202.435s)3225.414ms(317.591s)3865.668ms(404.202s)
42156.781ms(224.637s)5518.582ms(548.624s)12277.678ms(1175.750s)1911.034ms(188.739s)3161.600ms(329.157s)
82963.310ms(305.985s)6014.520ms(597.872s)11949.900ms(1165.736s)1564.713ms(154.796s)3259.090ms(338.562s)
alexnet on GTX980
GTX 980, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: alexnet
BatchSizeCaffeCNTKMXNETTensorFlowTorch
864.301ms(108.774s)4.633ms(110.504s)6.953ms(163.015s)41.799ms(975.474s)7.005ms(177.634s)
1285.949ms(102.314s)6.898ms(110.231s)9.510ms(150.454s)29.309ms(460.757s)9.660ms(164.976s)
25611.145ms(96.113s)12.513ms(100.532s)18.074ms(142.486s)55.175ms(8931.529s)18.198ms(155.026s)
51221.383ms(92.916s)23.758ms(95.601s)33.721ms(132.835s)108.949ms(2482.215s)35.067ms(149.470s)
102441.261ms(89.427s)45.320ms(91.320s)59.848ms(118.155s)227.139ms(447.296s)68.511ms(147.083s)
204881.751ms(90.583s)88.789ms(91.103s)136.352ms(134.591s)457.436ms(459.560s)135.917ms(145.148s)
alexnet on E5-2630v3
E5-2630v3: Server CPU, with 8x2 physical cores Network: alexnet
#ThreadsCaffeCNTKMXNETTensorFlowTorch
14269.687ms(441.625s)9380.878ms(943.303s)17994.332ms(1757.913s)6124.044ms(602.513s)6554.429ms(692.208s)
23513.881ms(368.452s)6078.235ms(612.386s)17128.202ms(1673.296s)4228.606ms(416.670s)5288.282ms(555.088s)
43380.717ms(351.972s)4983.776ms(500.374s)16764.375ms(1637.772s)2200.187ms(217.881s)4364.674ms(457.293s)
83364.235ms(349.352s)4764.816ms(477.110s)16470.548ms(1609.071s)1396.016ms(139.160s)3939.802ms(411.040s)
164138.676ms(429.127s)6256.418ms(626.829s)17471.109ms(1706.749s)1035.552ms(103.621s)4157.214ms(430.323s)
324930.483ms(506.125s)6199.163ms(624.076s)17769.615ms(1735.967s)971.442ms(97.468s)4164.945ms(430.470s)
alexnet on K80
Tesla K80, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: alexnet
BatchSizeCaffeCNTKMXNETTensorFlowTorch
8613.423ms(334.833s)11.673ms(277.589s)16.527ms(387.186s)37.628ms(880.267s)14.076ms(353.056s)
12819.223ms(322.909s)15.543ms(248.793s)19.243ms(303.740s)54.020ms(848.850s)19.665ms(331.301s)
25636.184ms(306.014s)28.520ms(229.237s)37.166ms(292.492s)101.878ms(802.392s)36.921ms(310.800s)
51269.806ms(295.963s)56.674ms(227.783s)67.068ms(264.928s)199.405ms(785.269s)71.766ms(302.353s)
1024137.285ms(291.470s)107.579ms(216.317s)122.024ms(241.275s)384.566ms(757.357s)141.434ms(297.841s)
2048273.111ms(295.838s)207.505ms(213.135s)298.259ms(294.479s)777.247ms(780.365s)280.591ms(295.291s)
fcn5 on K80
Tesla K80, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: fcn5
BatchSizeCaffeCNTKMXNETTensorFlowTorch
34224.762ms(195.255s)25.001ms(239.932s)29.759ms(212.975s)28.400ms(211.781s)24.140ms(184.335s)
51232.088ms(172.273s)31.246ms(212.595s)37.983ms(182.191s)37.476ms(188.980s)30.304ms(156.039s)
102455.074ms(150.900s)52.691ms(190.474s)60.425ms(145.885s)62.848ms(160.573s)50.761ms(131.640s)
204898.418ms(139.026s)94.671ms(178.316s)106.143ms(128.699s)112.004ms(146.461s)89.630ms(116.487s)
4096238.965ms(168.775s)181.415ms(173.852s)183.543ms(111.039s)208.436ms(139.719s)165.457ms(108.179s)
alexnet on GTX1080
GTX 1080, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: alexnet
BatchSizeCaffeCNTKMXNETTensorFlowTorch
862.837ms(74.349s)3.127ms(75.097s)3.927ms(92.893s)25.090ms(586.886s)4.547ms(117.526s)
1284.023ms(71.412s)4.431ms(71.848s)5.292ms(84.541s)39.396ms(618.744s)6.232ms(108.012s)
2567.478ms(67.259s)8.133ms(66.424s)9.156ms(73.170s)88.525ms(696.555s)11.546ms(100.223s)
51214.249ms(64.636s)15.634ms(63.634s)16.969ms(67.941s)172.147ms(677.095s)21.853ms(96.673s)
102427.127ms(62.395s)30.627ms(62.529s)32.305ms(65.005s)317.155ms(623.930s)42.838ms(94.915s)
204853.525ms(63.120s)60.331ms(62.980s)62.734ms(63.415s)436.263ms(17423.743s)84.020ms(96.381s)
resnet on E5-2630v3
E5-2630v3: Server CPU, with 8x2 physical cores Network: resnet
#ThreadsCaffeCNTKMXNETTensorFlowTorch
17513.092ms(6657.780s)--44877.981ms(35061.857s)--8690.756ms(7463.500s)
26118.700ms(5520.624s)--43774.981ms(34200.196s)--4965.160ms(4241.255s)
46231.563ms(5556.999s)--42299.283ms(33047.148s)--3039.966ms(2577.879s)
86689.152ms(5962.167s)--42964.982ms(33567.228s)--2559.605ms(2129.883s)
167312.890ms(6511.768s)--43853.915ms(34261.666s)--2574.915ms(2116.325s)
329302.413ms(8110.684s)--44367.438ms(34662.891s)--2811.217ms(2295.419s)
resnet on i7-3820
i7-3820: Desktop CPU, with 4 physical cores Network: resnet
#ThreadsCaffeCNTKMXNETTensorFlowTorch
16741.489ms(5963.685s)--34408.543ms(26882.644s)--5757.811ms(4799.697s)
25450.932ms(4843.517s)--31254.625ms(500.901s)--3222.200ms(2678.413s)
44989.233ms(4428.678s)--30068.875ms(482.158s)--2368.000ms(1980.011s)
86691.455ms(5895.914s)--31388.000ms(503.106s)--2475.000ms(2032.297s)
resnet on GTX1080
GTX 1080, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: resnet
BatchSizeCaffeCNTKMXNETTensorFlowTorch
1141.475ms(7751.852s)23.521ms(4287.278s)17.570ms(3196.239s)55.816ms(21306.860s)21.109ms(4098.784s)
1646.547ms(6028.732s)28.421ms(3563.052s)18.810ms(2353.124s)58.662ms(7372.103s)22.714ms(3022.751s)
3255.255ms(3665.407s)44.852ms(2814.494s)26.668ms(1668.591s)65.831ms(4154.252s)31.167ms(2050.199s)
6477.494ms(2635.121s)76.339ms(2398.323s)42.650ms(1334.761s)90.679ms(2873.439s)54.352ms(1787.140s)
128115.779ms(2022.574s)137.871ms(2167.061s)73.652ms(1152.800s)152.262ms(2419.955s)101.467ms(1668.802s)
resnet on GTX980
GTX 980, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: resnet
BatchSizeCaffeCNTKMXNETTensorFlowTorch
1149.257ms(9226.630s)38.872ms(7080.567s)19.831ms(3607.282s)60.100ms(27164.810s)24.860ms(4786.322s)
1652.154ms(6789.402s)43.006ms(5387.785s)20.803ms(2603.821s)62.996ms(7914.140s)28.222ms(3717.527s)
3266.146ms(4405.589s)66.545ms(527.805s)34.500ms(2158.210s)79.380ms(36999.101s)44.989ms(2956.396s)
6499.718ms(3389.141s)113.609ms(3565.639s)56.076ms(1754.735s)129.794ms(4097.209s)79.692ms(2618.308s)
128--206.272ms(3238.111s)105.291ms(1647.051s)224.960ms(3554.434s)149.738ms(2461.040s)
lstm on GTX1080
GTX 1080, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: lstm
BatchSizeCNTKMXNETTensorFlowTorch
6417.372ms(1011.691s)41.531ms(809.654s)62.657ms(229.245s)200.448ms(562.313s)
12817.792ms(519.212s)79.286ms(771.149s)77.794ms(185.399s)403.671ms(564.105s)
25618.187ms(266.552s)155.215ms(753.402s)116.343ms(168.394s)830.178ms(581.703s)
51221.426ms(158.861s)315.074ms(764.354s)206.584ms(161.424s)1386.101ms(2668.249s)
102425.937ms(478.821s)612.110ms(733.646s)297.904ms(278.666s)--
lstm on GTX980
GTX 980, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: lstm
BatchSizeCNTKMXNETTensorFlowTorch
6417.469ms(1017.268s)46.329ms(180.992s)62.467ms(228.603s)199.492ms(557.714s)
12817.723ms(517.207s)90.114ms(176.293s)76.847ms(182.492s)402.582ms(562.994s)
25617.963ms(263.304s)182.706ms(181.382s)116.429ms(165.156s)830.133ms(579.981s)
51221.230ms(156.561s)360.533ms(176.667s)205.570ms(159.477s)--
102434.315ms(631.197s)698.546ms(835.722s)396.532ms(336.819s)--
lstm on i7-3820
i7-3820: Desktop CPU, with 4 physical cores Network: lstm
#ThreadsCNTKMXNETTensorFlowTorch
1185.640ms(2698.249s)--4662.373ms(2457.754s)6920.931ms(4066.496s)
2120.482ms(1751.804s)--3384.970ms(1876.785s)3830.742ms(2671.429s)
490.242ms(1312.465s)--1935.037ms(1223.824s)2682.209ms(2157.523s)
8117.903ms(1714.261s)--1532.136ms(1050.076s)3127.210ms(2408.009s)
lstm on E5-2630v3
E5-2630v3: Server CPU, with 8x2 physical cores Network: lstm
#ThreadsCNTKMXNETTensorFlowTorch
1211.117ms(3070.970s)--6449.451ms(3139.567s)7470.979ms(4306.550s)
2139.350ms(2028.340s)--4351.027ms(2184.970s)4640.779ms(3049.491s)
4116.613ms(1698.041s)--2238.316ms(1253.531s)3580.448ms(2554.308s)
8114.257ms(1663.975s)--1183.465ms(787.290s)3260.005ms(2411.628s)
16114.058ms(1660.991s)--702.275ms(550.124s)5148.454ms(3305.953s)
32198.439ms(2886.724s)--598.232ms(514.259s)5851.234ms(3704.118s)
lstm on K80
Tesla K80, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: lstm
BatchSizeCNTKMXNETTensorFlowTorch
6442.367ms(2467.864s)82.350ms(317.818s)91.506ms(400.982s)227.185ms(771.529s)
12844.177ms(1289.657s)150.927ms(291.855s)137.823ms(356.113s)449.668ms(769.335s)
25644.917ms(658.572s)284.898ms(275.499s)223.547ms(333.269s)906.958ms(778.672s)
51251.490ms(191.936s)553.201ms(268.196s)392.045ms(322.059s)1916.194ms(824.828s)
102465.010ms(1198.673s)1088.193ms(1287.928s)767.132ms(666.042s)5032.770ms(4068.317s)
resnet on K80
Tesla K80, CUDA: 8.0 CUDNN: v5.1 CUDA_DRIVER: 367.48 Network: resnet
BatchSizeCaffeCNTKMXNETTensorFlowTorch
11101.072ms(18927.970s)95.944ms(17478.958s)50.842ms(9247.966s)98.223ms(17909.967s)46.895ms(9018.215s)
16108.262ms(14081.632s)113.751ms(14251.619s)60.336ms(7546.077s)114.249ms(14327.897s)57.973ms(7690.328s)
32142.706ms(9471.179s)167.179ms(10484.769s)90.385ms(5653.085s)175.975ms(11044.946s)89.459ms(5923.693s)
64221.093ms(7465.746s)333.863ms(10475.847s)154.470ms(4831.273s)276.773ms(8700.080s)161.570ms(5350.590s)
128378.136ms(6468.854s)561.990ms(8821.975s)269.626ms(4220.070s)522.959ms(8227.598s)300.598ms(4969.448s)
{$ block footer %}
凌风探梅
关注
  • 0
    点赞
  • 踩
  • 0
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
专栏目录
deep-base:Docker的深度学习基础映像（Tensorflow，Caffe，MXNet，Torch，Openface等）
03-26
深度学习基础图像当今的深度学习框架需要大量的工作来安装和运行。 该Docker容器将所有流行的深度学习框架捆绑到一个Docker实例中。 Ubuntu Linux是该容器的基本操作系统选择（这是CUDA的要求，所有DL框架都可以很好...
人工智能：PyTorch深度学习框架介绍
IT技术分享社区
11-17 1559
Debug调试方便：PyTorch 的设计思路是线性、直观且易于使用的，当你执行一行代码时，一步步去执行，不涉及异步调试的繁琐，所以当代码出现Bug 的时候，可以通过很方便定位代码的位置，避免因为bug错误的指向或者异步和不透明的引擎造成查询问题的麻烦。神经结构：PyTorch通过一种反向自动求导的技术，可以做到零延迟地任意改变神经网络的行为，避免因为构建神经网络完成后，如果需要调整神经网络结构，只能从头开始的麻烦，采用PyTorch大大节省了人力和时间成本。2、PyTorch常用的工具包。
深度学习框架caffe/CNTK/Tensorflow/Theano/Torch的对比
weixin_30526593的博客
04-10 174
在单GPU下，所有这些工具集都调用cuDNN,因此只要外层的计算或者内存分配差异不大其性能表现都差不多。 Caffe: 1）主流工业级深度学习工具，具有出色的卷积神经网络实现。在计算机视觉领域Caffe仍然是最流行的工具包，他有很多扩展，但是由于 一些遗留的架构问题，它对递归网络和语言建模的支持很差。此外，在caffe种图层需要使用C++定义，而网络则使用protobuf定义。 2）caf...
资料收集：深度学习框架比较
自然语言处理大菜鸟--HsingWang
02-28 1292
1. https://en.wikipedia.org/wiki/Comparison_of_deep_learning_software 2. https://deeplearning4j.org/compare-dl4j-torch7-pylearn 3. https://www.microway.com/hpc-tech-tips/deep-learning-framew
机器学习入门（二）：工具与框架的选择
weixin_34274029的博客
01-11 276
专栏系列文章 机器学习：新手入门概览（一）： juejin.im/post/5a5620… [译]什么是蒙特卡洛树搜索: juejin.im/post/59f16e… 论深度学习的局限性 juejin.im/post/597835… 所需工具 根据具体场景的需要（科研/工程），以及团队规模大小，开发人员能力等问题，选择合适的语言及框架和辅助工具来进行开发是十分必要的。这不但降低了开发难度，同时...
快速入门深度学习1（用时1h）——MXNet版本
最新发布
定期分享我的发现和想法，感谢你的陪伴和支持
03-09 6260
问题：判断图片是否有猫简化：① 假设所有图像的高和宽都是同样的400像素大小，一个像素由红绿蓝三个值构成，那么一张图像就由近50万个数值表示。② 哪些数值隐藏着我们需要的信息呢？是所有数值的平均数，还是四个角的数值，抑或是图像中的某一个特别的点？③ 事实上，要想解读图像中的内容，需要寻找仅仅在结合成千上万的数值时才会出现的特征，如边缘、质地、形状、眼睛、鼻子等，最终才能判断图像中是否有猫。
树莓派深度学习系统镜像，包含opencv、pytorch、tensorflow、pytorch_lightning、tflite
02-20
镜像包含opencv 4.5.1、pytorch 1.6.0、torchvision 0.7.0、pytorch_lightning 1.5.0、tensorflow 2.4.0、tflite 2.1.0等众多深度学习相关Python依赖，几乎涵盖了所有主流深度学习支持，全部安装在系统裸环境中，...
TensorFlow深度学习，一篇文章就够了
01-27
与Caffe、Theano、Torch、MXNet等框架相比，TensorFlow在Github上Fork数和Star数都是最多的，而且在图形分类、音频处理、推荐系统和自然语言处理等场景下都有丰富的应用。最近流行的Keras框架底层默认使用TensorFlow...
deeplearning-cfn：使用CloudFormation（CFN），MXNet和TensorFlow在AWS上进行分布式深度学习
01-29
AWS CloudFormation深度学习模板使用 （提供MXNet，TensorFlow，Caffe，Theano，Torch和CNTK框架）启动实例和执行分布式深度学习所需的其他AWS资源的集群。 使用此模板，我们将继续致力于使 。 AWS CloudFormation...
gpu 加速矩阵 深度学习_GPU如何加速深度学习
weixin_26630173的博客
09-03 1052
gpu 加速矩阵 深度学习Neural networks and deep learning are not recent methods. In fact, they are quite old. Perceptrons, the first neural networks, were created in 1958 by Frank Rosenblatt. Even the invention...
MXNET深度学习官方指南
07-16
MXNET深度学习官方指南MXNET深度学习官方指南MXNET深度学习官方指南
第二期 Tensorflow 基准测试 《显卡就是开发板》
aggresss
11-03 2287
深度神经网络相关的blog里总会提示我们使用GPU训练网络是必须的，但没有一个数值的概念，我们真的不好评判GPU在训练网络方面到底比CPU快多少，这一期我们就利用上一期搭建的实验环境跑个小Demo实测一下GPU的性能。 Benchmarking State-of-the-Art Deep Learning Software Tools http://dlbench.comp.hkbu.edu.hk/
深度学习框架的比较（MXNet, Caffe, TensorFlow, Torch, Theano)
热门推荐
MyArrow的专栏
07-29 5万+
1. 比较表 2.详细描述
TensorFlow和Caffe、MXNet、Keras等其他深度学习框架的对比
xiangz_csdn的博客
02-25 4万+
[转] http://www.leiphone.com/news/201702/T5e31Y2ZpeG1ZtaN.html 雷锋网按：本文作者黄文坚，PPmoney 大数据算法总监，《 TensorFlow 实战》作者。本文节选自《 TensorFlow 实战》第二章。 在下周二（2月28日）雷锋网(公众号：雷锋网)硬创公开课栏目将会推出黄文坚老师的公开课《深度学习之经典卷积神
深度学习框架tensorflow,paddle,mxnet的实战性能对比测试
lujian45的博客
04-25 1万+
选用了业界主流的三种深度学习框架Paddle，TensorFlow和Mxnet，对它们分别作了研究和实战评估。用于测试的模型包括基于logistic回归模型和LSTM模型。
深度学习框架哪家强？MXNet称霸CNN、RNN和情感分析，TensorFlow仅擅长推断特征提取
AI科技大本营
12-01 1万+
深度学习框架哪家强：TensorFlow？Caffe？MXNet？Keras？PyTorch？对于这几大框架在运行各项深度任务时的性能差异如何，各位读者不免会有所好奇。 微软数据科学家Ilia Karmanov最新测试的结果显示，亚马逊MXNet在CNN、RNN与NLP情感分析任务上性能强劲，而TensorFlow仅擅长于特征提取。 测试详情更新在Ilia Ka
五大主流深度学习框架比较分析：MXNET是最好选择
isuker的博客
09-07 1万+
选自 RevolutionAnalytics http://www.aiweibang.com/yuedu/139947035.html 机器之心编译 作者：Anusua Trivedi 参与：YanChen、黄清纬、李亚洲 近日，微软数据科学家 Anusua Trivedi 发表系列文章讲解深度学习，这是此系列的第一篇，对主流深度学习框架进行
经验 | Pytorch还是Tensorflow？英伟达工程师帮你总结了
AI科技大本营
10-27 4623
翻译 | AI科技大本营（rgznai100） 参与 |reason_W 本文作者Dominic Monn，是NVIDIA机器学习工程师。在本文中，作者讲述了自己在深度学习过程中使用PyTorch和TensorFlow的一些实际体会，他从安装、使用、文档、社区和工具五个角度，比较了PyTorch和TensorFlow的优点和不足。 作者此前
深度学习框架Caffe/CNTK/Tensorflow/Theano/Torch的比较
ZCC的专栏
05-17 4857
1.网络和模型能力 Caffe Caffe可能是第一个主流的工业级深度学习工具，它开始于2013年底,具有出色的卷积神经网络实现。在计算机视觉领域Caffe依然是最流行的工具包，它有很多扩展，但是由于一些遗留的架构问题，它对递归网络和语言建模的支持很差。此外，在Caffe中图层需要使用C++定义，而网络则使用Protobuf定义。 CNTK 由深度学习热潮的发起演讲人创建,
Tensorflow入门：计算图模型与深度学习框架解析
接着，文章提到了几种不同的深度学习框架，如Torch、Caffe、Theano、CuDNN、Tensorflow、Mxnet和Cuda-convnet。这些框架各有特点，例如Caffe以其配置文件式的灵活性著称，而Tensorflow和Theano则属于编程式框架，更...
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值