十分钟搭建 Neural Style 服务
Neural style 是让机器模仿已有画作的绘画风格来把一张图片重新绘制的算法。原始论文参考【1】。
下面将介绍如何搭建基于 MxNet 的 neural style 服务，在阿里云 HPC （https://www.aliyun.com/product/hpc）上部署时间不超过十分钟。
获取MxNet源码：
# git clone https://github.com/dmlc/mxnet.git --recursive 正克隆到 'mxnet'... remote: Counting objects: 20971, done. remote: Compressing objects: 100% (10/10), done. remote: Total 20971 (delta 4), reused 2 (delta 2), pack-reused 20959 接收对象中: 100% (20971/20971), 5.67 MiB | 987.00 KiB/s, done. 处理 delta 中: 100% (12892/12892), done. 子模组 'dmlc-core' (https://github.com/dmlc/dmlc-core.git) 已为路径 'dmlc-core' 注册 子模组 'mshadow' (https://github.com/dmlc/mshadow.git) 已为路径 'mshadow' 注册 子模组 'ps-lite' (https://github.com/dmlc/ps-lite) 已为路径 'ps-lite' 注册 正克隆到 'dmlc-core'... remote: Counting objects: 3503, done. remote: Total 3503 (delta 0), reused 0 (delta 0), pack-reused 3503 接收对象中: 100% (3503/3503), 777.75 KiB | 144.00 KiB/s, done. 处理 delta 中: 100% (2075/2075), done. 子模组路径 'dmlc-core'：检出 '0fb74229bc635946667f7dfd1c17116b37d0d870' 正克隆到 'mshadow'... remote: Counting objects: 3566, done. remote: Total 3566 (delta 0), reused 0 (delta 0), pack-reused 3566 接收对象中: 100% (3566/3566), 1.17 MiB | 243.00 KiB/s, done. 处理 delta 中: 100% (2450/2450), done. 子模组路径 'mshadow'：检出 'f2df1886e43f114ae01a2384d77954acfede2aba' 正克隆到 'ps-lite'... remote: Counting objects: 1687, done. remote: Total 1687 (delta 0), reused 0 (delta 0), pack-reused 1687 接收对象中: 100% (1687/1687), 523.36 KiB | 148.00 KiB/s, done. 处理 delta 中: 100% (1066/1066), done. 子模组路径 'ps-lite'：检出 '7faaeb73bcb9d68b464186d3191494aa95243703'
修改编译选项
编辑 make/config.mk
(1) # whether use CUDA during compile USE_CUDA = 0 改为 # whether use CUDA during compile USE_CUDA = 1 (2) # add the path to CUDA library to link and compile flag # if you have already add them to environment variable, leave it as NONE # USE_CUDA_PATH = /usr/local/cuda USE_CUDA_PATH = NONE 改为 # add the path to CUDA library to link and compile flag # if you have already add them to environment variable, leave it as NONE USE_CUDA_PATH = /usr/local/cuda # USE_CUDA_PATH = NONE (3) # whether use CuDNN R3 library USE_CUDNN = 0 改为 # whether use CuDNN R3 library USE_CUDNN = 1 (4) # choose the version of blas you want to use # can be: mkl, blas, atlas, openblas # in default use atlas for linux while apple for osx UNAME_S := $(shell uname -s) ifeq ($(UNAME_S), Darwin) USE_BLAS = apple else USE_BLAS = atlas Endif 改为 # choose the version of blas you want to use # can be: mkl, blas, atlas, openblas # in default use atlas for linux while apple for osx UNAME_S := $(shell uname -s) ifeq ($(UNAME_S), Darwin) USE_BLAS = apple else USE_BLAS = openblas Endif 编辑 Makefile CFLAGS += -I$(ROOTDIR)/mshadow/ -I$(ROOTDIR)/dmlc-core/include -fPIC -Iinclude $(MSHADOW_CFLAGS) LDFLAGS = -pthread $(MSHADOW_LDFLAGS) $(DMLC_LDFLAGS) 改为 CFLAGS += -I/disk1/deeplearning/local_install/include -I$(ROOTDIR)/mshadow/ -I$(ROOTDIR)/dmlc-core/include -fPIC -Iinclude $(MSHADOW_CFLAGS) LDFLAGS = -L/disk1/deeplearning/local_install/lib -pthread $(MSHADOW_LDFLAGS) $(DMLC_LDFLAGS)
编译 MxNet ：
# export PKG_CONFIG_PATH=/disk1/deeplearning/local_install/lib/pkgconfig/ # make –j ……
安装：
# cd python/ # python setup.py install ……
准备 Neural Style 模型和数据：
# cd example/neural-style/ # ls download.sh find_mxnet.py model_vgg19.py README.md run.py # ./download.sh --2016-03-28 16:57:38-- https://github.com/dmlc/web-data/raw/master/mxnet/neural-style/model/vgg19.params 位置：https://raw.githubusercontent.com/dmlc/web-data/master/mxnet/neural-style/model/vgg19.params [跟随至新的 URL] --2016-03-28 16:57:54-- https://raw.githubusercontent.com/dmlc/web-data/master/mxnet/neural-style/model/vgg19.params 长度：80099200 (76M) [application/octet-stream] 正在保存至: “vgg19.params” 100%[===================================================================================================================================================>] 80,099,200 13.1MB/s 用时 6.2s 2016-03-28 16:58:11 (12.2 MB/s) - 已保存 “vgg19.params” [80099200/80099200]) --2016-03-28 16:58:11-- https://github.com/dmlc/web-data/raw/master/mxnet/neural-style/input/IMG_4343.jpg 位置：https://raw.githubusercontent.com/dmlc/web-data/master/mxnet/neural-style/input/IMG_4343.jpg [跟随至新的 URL] --2016-03-28 16:58:13-- https://raw.githubusercontent.com/dmlc/web-data/master/mxnet/neural-style/input/IMG_4343.jpg 长度：230594 (225K) [image/jpeg] 正在保存至: “IMG_4343.jpg” 100%[===================================================================================================================================================>] 230,594 902KB/s 用时 0.2s 2016-03-28 16:58:14 (902 KB/s) - 已保存 “IMG_4343.jpg” [230594/230594]) --2016-03-28 16:58:14-- https://github.com/dmlc/web-data/raw/master/mxnet/neural-style/input/starry_night.jpg 位置：https://raw.githubusercontent.com/dmlc/web-data/master/mxnet/neural-style/input/starry_night.jpg [跟随至新的 URL] --2016-03-28 16:58:15-- https://raw.githubusercontent.com/dmlc/web-data/master/mxnet/neural-style/input/starry_night.jpg 长度：315236 (308K) [image/jpeg] 正在保存至: “starry_night.jpg” 100%[===================================================================================================================================================>] 315,236 1.25MB/s 用时 0.2s 2016-03-28 16:58:16 (1.25 MB/s) - 已保存 “starry_night.jpg” [315236/315236])
运行 Neural Style 例程：
# python run.py INFO:root:load the content image, size = (1000, 1500) INFO:root:resize the content image to (400, 600) INFO:root:start training arguments Namespace(content_image='input/IMG_4343.jpg', content_weight=10, gpu=0, lr=0.1, max_long_edge=600, max_num_epochs=1000, model='vgg19', output='output/out.jpg', remove_noise=0.2, save_epochs=50, stop_eps=0.005, style_image='input/starry_night.jpg', style_weight=1) INFO:root:epoch 0, relative change 0.991449 INFO:root:epoch 1, relative change 0.639205 INFO:root:epoch 2, relative change 0.478858 INFO:root:epoch 3, relative change 0.385878 INFO:root:epoch 4, relative change 0.321652 INFO:root:epoch 5, relative change 0.274459 INFO:root:epoch 6, relative change 0.234445 INFO:root:epoch 7, relative change 0.185482 INFO:root:epoch 8, relative change 0.149244 INFO:root:epoch 9, relative change 0.127406 INFO:root:epoch 10, relative change 0.113780 INFO:root:Update[11]: Change learning rate to 9.00000e-02 INFO:root:epoch 11, relative change 0.101375 INFO:root:epoch 12, relative change 0.090326 INFO:root:epoch 13, relative change 0.084095 INFO:root:epoch 14, relative change 0.083168 INFO:root:epoch 15, relative change 0.085777 INFO:root:epoch 16, relative change 0.089530 INFO:root:epoch 17, relative change 0.092034 INFO:root:epoch 18, relative change 0.090668 INFO:root:epoch 19, relative change 0.086283 INFO:root:epoch 20, relative change 0.081812 INFO:root:Update[21]: Change learning rate to 8.10000e-02 INFO:root:epoch 21, relative change 0.076982 INFO:root:epoch 22, relative change 0.071818 INFO:root:epoch 23, relative change 0.066535 INFO:root:epoch 24, relative change 0.061628 INFO:root:epoch 25, relative change 0.057347 INFO:root:epoch 26, relative change 0.054294 INFO:root:epoch 27, relative change 0.053362 INFO:root:epoch 28, relative change 0.054526 INFO:root:epoch 29, relative change 0.056334 INFO:root:epoch 30, relative change 0.057255 INFO:root:Update[31]: Change learning rate to 7.29000e-02 INFO:root:epoch 31, relative change 0.056267 INFO:root:epoch 32, relative change 0.054968 INFO:root:epoch 33, relative change 0.054624 INFO:root:epoch 34, relative change 0.053384 INFO:root:epoch 35, relative change 0.051284 INFO:root:epoch 36, relative change 0.048232 INFO:root:epoch 37, relative change 0.046385 INFO:root:epoch 38, relative change 0.044870 INFO:root:epoch 39, relative change 0.043447 INFO:root:epoch 40, relative change 0.042659 INFO:root:Update[41]: Change learning rate to 6.56100e-02 INFO:root:epoch 41, relative change 0.039969 INFO:root:epoch 42, relative change 0.038084 INFO:root:epoch 43, relative change 0.036276 INFO:root:epoch 44, relative change 0.034864 INFO:root:epoch 45, relative change 0.034026 INFO:root:epoch 46, relative change 0.034046 INFO:root:epoch 47, relative change 0.033529 INFO:root:epoch 48, relative change 0.032085 INFO:root:epoch 49, relative change 0.031398 INFO:root:save output to output/tmp_50.jpg /disk1/deeplearning/anaconda2/lib/python2.7/site-packages/skimage/util/dtype.py:111: UserWarning: Possible precision loss when converting from float64 to uint8 "%s to %s" % (dtypeobj_in, dtypeobj)) INFO:root:epoch 50, relative change 0.031347 INFO:root:Update[51]: Change learning rate to 5.90490e-02 INFO:root:epoch 51, relative change 0.031458 INFO:root:epoch 52, relative change 0.030541 INFO:root:epoch 53, relative change 0.028714 INFO:root:epoch 54, relative change 0.028467 INFO:root:epoch 55, relative change 0.027554 INFO:root:epoch 56, relative change 0.025826 INFO:root:epoch 57, relative change 0.026127 INFO:root:epoch 58, relative change 0.024360 INFO:root:epoch 59, relative change 0.024311 INFO:root:epoch 60, relative change 0.022046 INFO:root:Update[61]: Change learning rate to 5.31441e-02 INFO:root:epoch 61, relative change 0.022097 INFO:root:epoch 62, relative change 0.019917 INFO:root:epoch 63, relative change 0.020184 INFO:root:epoch 64, relative change 0.017810 INFO:root:epoch 65, relative change 0.017987 INFO:root:epoch 66, relative change 0.016383 INFO:root:epoch 67, relative change 0.017593 INFO:root:epoch 68, relative change 0.015544 INFO:root:epoch 69, relative change 0.017172 INFO:root:epoch 70, relative change 0.015113 INFO:root:Update[71]: Change learning rate to 4.78297e-02 INFO:root:epoch 71, relative change 0.015867 INFO:root:epoch 72, relative change 0.014007 INFO:root:epoch 73, relative change 0.014800 INFO:root:epoch 74, relative change 0.013104 INFO:root:epoch 75, relative change 0.014838 INFO:root:epoch 76, relative change 0.012413 INFO:root:epoch 77, relative change 0.013954 INFO:root:epoch 78, relative change 0.012503 INFO:root:epoch 79, relative change 0.014131 INFO:root:epoch 80, relative change 0.011837 INFO:root:Update[81]: Change learning rate to 4.30467e-02 INFO:root:epoch 81, relative change 0.012995 INFO:root:epoch 82, relative change 0.011472 INFO:root:epoch 83, relative change 0.012635 INFO:root:epoch 84, relative change 0.010827 INFO:root:epoch 85, relative change 0.012640 INFO:root:epoch 86, relative change 0.010638 INFO:root:epoch 87, relative change 0.012243 INFO:root:epoch 88, relative change 0.010504 INFO:root:epoch 89, relative change 0.011789 INFO:root:epoch 90, relative change 0.010193 INFO:root:Update[91]: Change learning rate to 3.87420e-02 INFO:root:epoch 91, relative change 0.011762 INFO:root:epoch 92, relative change 0.009751 INFO:root:epoch 93, relative change 0.010758 INFO:root:epoch 94, relative change 0.009456 INFO:root:epoch 95, relative change 0.011349 INFO:root:epoch 96, relative change 0.009183 INFO:root:epoch 97, relative change 0.010511 INFO:root:epoch 98, relative change 0.009244 INFO:root:epoch 99, relative change 0.010658 INFO:root:save output to output/tmp_100.jpg INFO:root:epoch 100, relative change 0.008773 INFO:root:Update[101]: Change learning rate to 3.48678e-02 INFO:root:epoch 101, relative change 0.010101 INFO:root:epoch 102, relative change 0.008547 INFO:root:epoch 103, relative change 0.009856 INFO:root:epoch 104, relative change 0.008356 INFO:root:epoch 105, relative change 0.009952 INFO:root:epoch 106, relative change 0.008303 INFO:root:epoch 107, relative change 0.009713 INFO:root:epoch 108, relative change 0.008323 INFO:root:epoch 109, relative change 0.009423 ^^INFO:root:epoch 110, relative change 0.008171 INFO:root:Update[111]: Change learning rate to 3.13811e-02 INFO:root:epoch 111, relative change 0.009585 INFO:root:epoch 112, relative change 0.008034 INFO:root:epoch 113, relative change 0.008760 INFO:root:epoch 114, relative change 0.007746 INFO:root:epoch 115, relative change 0.009353 INFO:root:epoch 116, relative change 0.007418 INFO:root:epoch 117, relative change 0.008406 INFO:root:epoch 118, relative change 0.007384 INFO:root:epoch 119, relative change 0.008828 INFO:root:epoch 120, relative change 0.007084 INFO:root:Update[121]: Change learning rate to 2.82430e-02 INFO:root:epoch 121, relative change 0.008017 INFO:root:epoch 122, relative change 0.006935 INFO:root:epoch 123, relative change 0.008137 INFO:root:epoch 124, relative change 0.006811 INFO:root:epoch 125, relative change 0.008026 INFO:root:epoch 126, relative change 0.006733 INFO:root:epoch 127, relative change 0.007884 INFO:root:epoch 128, relative change 0.006674 INFO:root:epoch 129, relative change 0.007826 INFO:root:epoch 130, relative change 0.006627 INFO:root:Update[131]: Change learning rate to 2.54187e-02 INFO:root:epoch 131, relative change 0.007637 INFO:root:epoch 132, relative change 0.006630 INFO:root:epoch 133, relative change 0.007475 INFO:root:epoch 134, relative change 0.006713 INFO:root:epoch 135, relative change 0.007901 INFO:root:epoch 136, relative change 0.006373 INFO:root:epoch 137, relative change 0.007069 INFO:root:epoch 138, relative change 0.006252 INFO:root:epoch 139, relative change 0.007505 INFO:root:epoch 140, relative change 0.006020 INFO:root:Update[141]: Change learning rate to 2.28768e-02 INFO:root:epoch 141, relative change 0.006644 INFO:root:epoch 142, relative change 0.005896 INFO:root:epoch 143, relative change 0.006905 INFO:root:epoch 144, relative change 0.005786 INFO:root:epoch 145, relative change 0.006633 INFO:root:epoch 146, relative change 0.005680 INFO:root:epoch 147, relative change 0.006628 INFO:root:epoch 148, relative change 0.005598 INFO:root:epoch 149, relative change 0.006543 INFO:root:save output to output/tmp_150.jpg INFO:root:epoch 150, relative change 0.005588 INFO:root:Update[151]: Change learning rate to 2.05891e-02 INFO:root:epoch 151, relative change 0.006343 INFO:root:epoch 152, relative change 0.005581 INFO:root:epoch 153, relative change 0.006313 INFO:root:epoch 154, relative change 0.005680 INFO:root:epoch 155, relative change 0.006554 INFO:root:epoch 156, relative change 0.005409 INFO:root:epoch 157, relative change 0.005958 INFO:root:epoch 158, relative change 0.005302 INFO:root:epoch 159, relative change 0.006252 INFO:root:epoch 160, relative change 0.005130 INFO:root:Update[161]: Change learning rate to 1.85302e-02 INFO:root:epoch 161, relative change 0.005618 INFO:root:epoch 162, relative change 0.005035 INFO:root:epoch 163, relative change 0.005801 INFO:root:epoch 164, relative change 0.004944 INFO:root:eps < args.stop_eps, training finished INFO:root:save output to output/out.jpg
查看结果
输入图像为：
风格图像为：
输出图像为：
运行一次图片风格转换时间为1-2 min。
上述demo只能在命令行下调用，你也可以搭一个简易的web server，支持用户上传下载图片，例如图片风格转换服务。
立刻尝试在阿里云 HPC 上搭建你的服务！入口
更多新的 demo 请关注阿里云 HPC 官方论坛
参考文献：
【1】 A Neural Algorithm of Artistic Style， arXiv:1508.06576v2