機器學習進階筆記之七 | MXnet初體驗

引言

前段時間，『機器學習進階筆記』系列一直關注TensorFlow系統的技術實踐（想看TensorFlow技術實踐的同學可直接拉到文章底部看相關閱讀推薦），幫助大家從零開始，由淺入深，走上機器學習的進階之路。雖然之前都在誇TensorFlow的好，但其劣勢也很明顯——對計算力要求太高，雖然使用方便，但是顯存佔用太高，計算也不夠快，做公司項目還好，自己玩一些好玩的東西時太費時間了。

簡而言之，窮！

今天新開一篇，給大家介紹另一個優秀而強大的深度學習框架——MXnet，現在MXnet資源相對少一點，基於MXnet的有意思的開源項目也相對少一點，不過沒關係，都不是問題，他的優點是足夠靈活，速度足夠快，擴展新的功能比較容易，還有就是造MXnet都是一群說得上名字的大牛，能和大牛們玩一樣的東西，想想都很興奮有沒有！

那我們開始吧：）

前言

如何找到自己實用的丹爐，是一個深度修真之人至關重要的，丹爐的好壞直接關係到煉丹的成功與否，道途千載，尋一合適丹爐也不妨這千古悠悠的修真（正）之路。

為什麼學mxnet？ 熟悉本人博客的都知道，前段時間一直在關注TensorFlow也安利了很多次TFlearn，為什麼這次突然會寫MXnet的東西呢？原因是沒錢呀，TensorFlow計算力要求太高，雖然使用方便，但是顯存佔用太高，計算也不夠快，做公司項目還好，自己玩一些好玩的東西時太費時間，不過現在MXnet資源相對少一點，基於MXnet的有意思的開源項目也相對少一點，不過沒關係，都不是問題，另外一點就是造MXnet都是一群說得上名字的大牛，能和大牛們玩一樣的東西，想想都很興奮。

MXnet的文檔一直被一些愛好者噴，確實文檔比較少，不過考慮到開發者都是業餘時間造輪子（不，造丹爐！），很那像其他的框架有那麼熟悉的文檔，不過還好，在cv這塊還是比較容易下手的。 這裡有我從最近開始接觸MXnet（其實很早就聽說一直沒有用過），學習的一些代碼還有筆記mxnet 101,沒有特別細緻研究，只是了解怎麼用在CV上，完整的做一個項目。

新的丹方—inception-resnet-v2

每一付新的丹方，無不是深度前輩們多年經驗的結晶，丹方，很多時候在同樣煉丹材料表現天差地別，也成為傳奇前輩們的一個個標誌。

一看到這個名字就知道和resnet和inception（googlenet 即是inception-v1）逃脫不了干係，就是一個比較複雜的網路結構，具體多複雜？！玩過tflearn的去看看我寫的代碼，run下 然後從tensorboard的graph打開看看，（之前一個被merge的版本後來發現沒有batch normalization）改了的提了PR但是在寫博客的時候還沒有被mergeadd inception-resnet-v2 in branch inception-resnet-v2 #450。總之就是」丹方」特別複雜，具體去結合Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning，了解過resnet和googlenet的網路結構的小夥伴應該很容易弄明白，以下tflearn的代碼參考tf.slim下inception-resnet-v2。 基本的代碼結構：

# -*- coding: utf-8 -*-nn """ inception_resnet_v2.nn Applying inception_resnet_v2 to Oxfords 17 Category Flower Dataset classification task.nn References:n Inception-v4, Inception-ResNet and the Impact of Residual Connectionsn on Learningn Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi.nn Links:n [http://arxiv.org/abs/1602.07261](http://arxiv.org/abs/1602.07261)nn """nn from __future__ import division, print_function, absolute_importn import tflearnn from tflearn.layers.core import input_data, dropout, flatten, fully_connectedn from tflearn.layers.conv import conv_2d, max_pool_2d, avg_pool_2dn from tflearn.utils import repeatn from tflearn.layers.merge_ops import mergen from tflearn.data_utils import shuffle, to_categoricaln import tflearn.activations as activationsn import tflearn.datasets.oxflower17 as oxflower17n def block35(net, scale=1.0, activation=relu):n tower_conv = conv_2d(net, 32, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_1x1)n tower_conv1_0 = conv_2d(net, 32, 1, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_0a_1x1)n tower_conv1_1 = conv_2d(tower_conv1_0, 32, 3, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_0b_3x3)n tower_conv2_0 = conv_2d(net, 32, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_0a_1x1)n tower_conv2_1 = conv_2d(tower_conv2_0, 48,3, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_0b_3x3)n tower_conv2_2 = conv_2d(tower_conv2_1, 64,3, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_0c_3x3)n tower_mixed = merge([tower_conv, tower_conv1_1, tower_conv2_2], mode=concat, axis=3)n tower_out = conv_2d(tower_mixed, net.get_shape()[3], 1, normalizer_fn=batch_normalization, activation=None, name=Conv2d_1x1)n net += scale * tower_outn if activation:n if isinstance(activation, str):n net = activations.get(activation)(net)n elif hasattr(activation, __call__):n net = activation(net)n else:n raise ValueError("Invalid Activation.")n return netnn def block17(net, scale=1.0, activation=relu):n tower_conv = conv_2d(net, 192, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_1x1)n tower_conv_1_0 = conv_2d(net, 128, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_0a_1x1)n tower_conv_1_1 = conv_2d(tower_conv_1_0, 160,[1,7], normalizer_fn=batch_normalization, activation=relu,name=Conv2d_0b_1x7)n tower_conv_1_2 = conv_2d(tower_conv_1_1, 192, [7,1], normalizer_fn=batch_normalization, activation=relu,name=Conv2d_0c_7x1)n tower_mixed = merge([tower_conv,tower_conv_1_2], mode=concat, axis=3)n tower_out = conv_2d(tower_mixed, net.get_shape()[3], 1, normalizer_fn=batch_normalization, activation=None, name=Conv2d_1x1)n net += scale * tower_outn if activation:n if isinstance(activation, str):n net = activations.get(activation)(net)n elif hasattr(activation, __call__):n net = activation(net)n else:n raise ValueError("Invalid Activation.")n return netnnn def block8(net, scale=1.0, activation=relu):n """n """n tower_conv = conv_2d(net, 192, 1, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_1x1)n tower_conv1_0 = conv_2d(net, 192, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_0a_1x1)n tower_conv1_1 = conv_2d(tower_conv1_0, 224, [1,3], normalizer_fn=batch_normalization, name=Conv2d_0b_1x3)n tower_conv1_2 = conv_2d(tower_conv1_1, 256, [3,1], normalizer_fn=batch_normalization, name=Conv2d_0c_3x1)n tower_mixed = merge([tower_conv,tower_conv1_2], mode=concat, axis=3)n tower_out = conv_2d(tower_mixed, net.get_shape()[3], 1, normalizer_fn=batch_normalization, activation=None, name=Conv2d_1x1)n net += scale * tower_outn if activation:n if isinstance(activation, str):n net = activations.get(activation)(net)n elif hasattr(activation, __call__):n net = activation(net)n else:n raise ValueError("Invalid Activation.")n return netnn # Data loading and preprocessingn import tflearn.datasets.oxflower17 as oxflower17n X, Y = oxflower17.load_data(one_hot=True, resize_pics=(299, 299))nn num_classes = 17n dropout_keep_prob = 0.8nn network = input_data(shape=[None, 299, 299, 3])n conv1a_3_3 = conv_2d(network, 32, 3, strides=2, normalizer_fn=batch_normalization, padding=VALID,activation=relu,name=Conv2d_1a_3x3)n conv2a_3_3 = conv_2d(conv1a_3_3, 32, 3, normalizer_fn=batch_normalization, padding=VALID,activation=relu, name=Conv2d_2a_3x3)n conv2b_3_3 = conv_2d(conv2a_3_3, 64, 3, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_2b_3x3)n maxpool3a_3_3 = max_pool_2d(conv2b_3_3, 3, strides=2, padding=VALID, name=MaxPool_3a_3x3)n conv3b_1_1 = conv_2d(maxpool3a_3_3, 80, 1, normalizer_fn=batch_normalization, padding=VALID,activation=relu, name=Conv2d_3b_1x1)n conv4a_3_3 = conv_2d(conv3b_1_1, 192, 3, normalizer_fn=batch_normalization, padding=VALID,activation=relu, name=Conv2d_4a_3x3)n maxpool5a_3_3 = max_pool_2d(conv4a_3_3, 3, strides=2, padding=VALID, name=MaxPool_5a_3x3)nn tower_conv = conv_2d(maxpool5a_3_3, 96, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_5b_b0_1x1)nn tower_conv1_0 = conv_2d(maxpool5a_3_3, 48, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_5b_b1_0a_1x1)n tower_conv1_1 = conv_2d(tower_conv1_0, 64, 5, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_5b_b1_0b_5x5)nn tower_conv2_0 = conv_2d(maxpool5a_3_3, 64, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_5b_b2_0a_1x1)n tower_conv2_1 = conv_2d(tower_conv2_0, 96, 3, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_5b_b2_0b_3x3)n tower_conv2_2 = conv_2d(tower_conv2_1, 96, 3, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_5b_b2_0c_3x3)nn tower_pool3_0 = avg_pool_2d(maxpool5a_3_3, 3, strides=1, padding=same, name=AvgPool_5b_b3_0a_3x3)n tower_conv3_1 = conv_2d(tower_pool3_0, 64, 1, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_5b_b3_0b_1x1)nn tower_5b_out = merge([tower_conv, tower_conv1_1, tower_conv2_2, tower_conv3_1], mode=concat, axis=3)nn net = repeat(tower_5b_out, 10, block35, scale=0.17)tower_conv2_2 = conv_2d(tower_conv2_1, 96, 3, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_5b_b2_0c_3x3)nn tower_pool3_0 = avg_pool_2d(maxpool5a_3_3, 3, strides=1, padding=same, name=AvgPool_5b_b3_0a_3x3)n tower_conv3_1 = conv_2d(tower_pool3_0, 64, 1, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_5b_b3_0b_1x1)nn tower_5b_out = merge([tower_conv, tower_conv1_1, tower_conv2_2, tower_conv3_1], mode=concat, axis=3)nn net = repeat(tower_5b_out, 10, block35, scale=0.17)nn tower_conv = conv_2d(net, 384, 3, normalizer_fn=batch_normalization, strides=2,activation=relu, padding=VALID, name=Conv2d_6a_b0_0a_3x3)n tower_conv1_0 = conv_2d(net, 256, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_6a_b1_0a_1x1)n tower_conv1_1 = conv_2d(tower_conv1_0, 256, 3, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_6a_b1_0b_3x3)n tower_conv1_2 = conv_2d(tower_conv1_1, 384, 3, normalizer_fn=batch_normalization, strides=2, padding=VALID, activation=relu,name=Conv2d_6a_b1_0c_3x3)n tower_pool = max_pool_2d(net, 3, strides=2, padding=VALID,name=MaxPool_1a_3x3)n net = merge([tower_conv, tower_conv1_2, tower_pool], mode=concat, axis=3)n net = repeat(net, 20, block17, scale=0.1)nn tower_conv = conv_2d(net, 256, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_0a_1x1)n tower_conv0_1 = conv_2d(tower_conv, 384, 3, normalizer_fn=batch_normalization, strides=2, padding=VALID, activation=relu,name=Conv2d_0a_1x1)nn tower_conv1 = conv_2d(net, 256, 1, normalizer_fn=batch_normalization, padding=VALID, activation=relu,name=Conv2d_0a_1x1)n tower_conv1_1 = conv_2d(tower_conv1,288,3, normalizer_fn=batch_normalization, strides=2, padding=VALID,activation=relu, name=COnv2d_1a_3x3)nn tower_conv2 = conv_2d(net, 256,1, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_0a_1x1)n tower_conv2_1 = conv_2d(tower_conv2, 288,3, normalizer_fn=batch_normalization, name=Conv2d_0b_3x3,activation=relu)n tower_conv2_2 = conv_2d(tower_conv2_1, 320, 3, normalizer_fn=batch_normalization, strides=2, padding=VALID,activation=relu, name=Conv2d_1a_3x3)nn tower_pool = max_pool_2d(net, 3, strides=2, padding=VALID, name=MaxPool_1a_3x3)n net = merge([tower_conv0_1, tower_conv1_1,tower_conv2_2, tower_pool], mode=concat, axis=3)nn net = repeat(net, 9, block8, scale=0.2)n net = block8(net, activation=None)nn net = conv_2d(net, 1536, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_7b_1x1)n net = avg_pool_2d(net, net.get_shape().as_list()[1:3],strides=2, padding=VALID, name=AvgPool_1a_8x8)n net = flatten(net)n net = dropout(net, dropout_keep_prob)n loss = fully_connected(net, num_classes,activation=softmax)nnn network = tflearn.regression(loss, optimizer=RMSprop,n loss=categorical_crossentropy,n learning_rate=0.0001)n model = tflearn.DNN(network, checkpoint_path=inception_resnet_v2,n max_checkpoints=1, tensorboard_verbose=2, tensorboard_dir="./tflearn_logs/")n model.fit(X, Y, n_epoch=1000, validation_set=0.1, shuffle=True,n show_metric=True, batch_size=32, snapshot_step=2000,n snapshot_epoch=False, run_id=inception_resnet_v2_17flowers)nn tower_conv = conv_2d(net, 384, 3, normalizer_fn=batch_normalization, strides=2,activation=relu, padding=VALID, name=Conv2d_6a_b0_0a_3x3)n tower_conv1_0 = conv_2d(net, 256, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_6a_b1_0a_1x1)n tower_conv1_1 = conv_2d(tower_conv1_0, 256, 3, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_6a_b1_0b_3x3)n tower_conv1_2 = conv_2d(tower_conv1_1, 384, 3, normalizer_fn=batch_normalization, strides=2, padding=VALID, activation=relu,name=Conv2d_6a_b1_0c_3x3)n tower_pool = max_pool_2d(net, 3, strides=2, padding=VALID,name=MaxPool_1a_3x3)n net = merge([tower_conv, tower_conv1_2, tower_pool], mode=concat, axis=3)n net = repeat(net, 20, block17, scale=0.1)nn tower_conv = conv_2d(net, 256, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_0a_1x1)n tower_conv0_1 = conv_2d(tower_conv, 384, 3, normalizer_fn=batch_normalization, strides=2, padding=VALID, activation=relu,name=Conv2d_0a_1x1)nn tower_conv1 = conv_2d(net, 256, 1, normalizer_fn=batch_normalization, padding=VALID, activation=relu,name=Conv2d_0a_1x1)n tower_conv1_1 = conv_2d(tower_conv1,288,3, normalizer_fn=batch_normalization, strides=2, padding=VALID,activation=relu, name=COnv2d_1a_3x3)nn tower_conv2 = conv_2d(net, 256,1, normalizer_fn=batch_normalization, activation=relu,name=Conv2d_0a_1x1)n tower_conv2_1 = conv_2d(tower_conv2, 288,3, normalizer_fn=batch_normalization, name=Conv2d_0b_3x3,activation=relu)n tower_conv2_2 = conv_2d(tower_conv2_1, 320, 3, normalizer_fn=batch_normalization, strides=2, padding=VALID,activation=relu, name=Conv2d_1a_3x3)nn tower_pool = max_pool_2d(net, 3, strides=2, padding=VALID, name=MaxPool_1a_3x3)n net = merge([tower_conv0_1, tower_conv1_1,tower_conv2_2, tower_pool], mode=concat, axis=3)nn net = repeat(net, 9, block8, scale=0.2)n net = block8(net, activation=None)nn net = conv_2d(net, 1536, 1, normalizer_fn=batch_normalization, activation=relu, name=Conv2d_7b_1x1)n net = avg_pool_2d(net, net.get_shape().as_list()[1:3],strides=2, padding=VALID, name=AvgPool_1a_8x8)n net = flatten(net)n net = dropout(net, dropout_keep_prob)n loss = fully_connected(net, num_classes,activation=softmax)nnn network = tflearn.regression(loss, optimizer=RMSprop,n loss=categorical_crossentropy,n learning_rate=0.0001)n model = tflearn.DNN(network, checkpoint_path=inception_resnet_v2,n max_checkpoints=1, tensorboard_verbose=2, tensorboard_dir="./tflearn_logs/")n model.fit(X, Y, n_epoch=1000, validation_set=0.1, shuffle=True,n show_metric=True, batch_size=32, snapshot_step=2000,n snapshot_epoch=False, run_id=inception_resnet_v2_17flowers)n

想要run下的可以去使用下tflearn，注意更改conv_2d裡面的內容，我這裡在本身conv_2d上加了個normalizer_fn，來使用batch_normalization。

MXnet 煉丹

不同的丹爐，即使是相同的丹方，煉丹的方式都不僅相同。

在打算用MXnet實現inception-resnet-v2之前，除了mxnet-101裡面的代碼，基本沒有寫過mxnet，但是沒關係，不怕，有很多其他大神寫的丹方，這裡具體參考了symbol_inception-bn.py。首先，為了減少代碼條數，參考創建一個ConvFactory，但是和inception-bn不同的是，inception-resnet-v2要考慮是否要激活函數的版本。所以inception-resnet-v2的ConvFactory如下：

def ConvFactory(data, num_filter, kernel, stride=(1,1), pad=(0, 0), act_type="relu", mirror_attr={},with_act=True):n conv = mx.symbol.Convolution(data=data, num_filter=num_filter, kernel=kernel, stride=stride, pad=pad)n bn = mx.symbol.BatchNorm(data=conv)n if with_act:n act = mx.symbol.Activation(data = bn, act_type=act_type, attr=mirror_attr)n return actn else:n return bnn

然後就簡單了，按照網路一路往下寫：

def get_symbol(num_classes=1000，input_data_shape=(64,3,299,299)):n data = mx.symbol.Variable(name=data)n conv1a_3_3 = ConvFactory(data=data, num_filter=32, kernel=(3,3), stride=(2, 2))n conv2a_3_3 = ConvFactory(conv1a_3_3, 32, (3,3))n conv2b_3_3 = ConvFactory(conv2a_3_3, 64, (3,3), pad=(1,1))n maxpool3a_3_3 = mx.symbol.Pooling(data=conv2b_3_3, kernel=(3, 3), stride=(2, 2), pool_type=max)n conv3b_1_1 = ConvFactory(maxpool3a_3_3, 80 ,(1,1))n conv4a_3_3 = ConvFactory(conv3b_1_1, 192, (3,3))n maxpool5a_3_3 = mx.symbol.Pooling(data=conv4a_3_3, kernel=(3,3), stride=(2,2), pool_type=max)nn tower_conv = ConvFactory(maxpool5a_3_3, 96, (1,1))n tower_conv1_0 = ConvFactory(maxpool5a_3_3, 48, (1,1))n tower_conv1_1 = ConvFactory(tower_conv1_0, 64, (5,5), pad=(2,2))nn tower_conv2_0 = ConvFactory(maxpool5a_3_3, 64, (1,1))n tower_conv2_1 = ConvFactory(tower_conv2_0, 96, (3,3), pad=(1,1))n tower_conv2_2 = ConvFactory(tower_conv2_1, 96, (3,3), pad=(1,1))nn tower_pool3_0 = mx.symbol.Pooling(data=maxpool5a_3_3, kernel=(3,3), stride=(1,1),pad=(1,1), pool_type=avg)n tower_conv3_1 = ConvFactory(tower_pool3_0, 64, (1,1))n tower_5b_out = mx.symbol.Concat(*[tower_conv, tower_conv1_1, tower_conv2_2, tower_conv3_1])n

然後就不對了，要重複條用一個block35的結構，repeat函數很容易實現，給定調用次數，調用函數，參數， 多次調用就好了：

def repeat(inputs, repetitions, layer, *args, **kwargs):n outputs = inputsn for i in range(repetitions):n outputs = layer(outputs, *args, **kwargs)n return outputsn

這裡很簡單，但是block35就有問題啦，這個子結構的目的要輸出與輸入同樣大小的channel數，之前因為在tensorflow下寫的，很容易拿到一個Variable的shape，但是在MXnet上就很麻煩，這裡不知道怎麼做，提了個issue How can i get the shape with the net?，然後就去查api，發現有個infer_shape，mxnet客服部小夥伴也讓我用這個去做， 試了試，挺管用能夠拿到shape，但是必須給入一個4d的tensor的shape，比如(64,3,299,299)，他會在graph運行時infer到對應symbol的shape，然後就這麼寫了:

def block35(net, input_data_shape, scale=1.0, with_act=True, act_type=relu, mirror_attr={}):n assert len(input_data_shape) == 4, input_data_shape should be len of 4, your n input_data_shape is len of %d%len(input_data_shape)n _, out_shape,_ = net.infer_shape(data=input_data_shape)n tower_conv = ConvFactory(net, 32, (1,1))n tower_conv1_0 = ConvFactory(net, 32, (1,1))n tower_conv1_1 = ConvFactory(tower_conv1_0, 32, (3, 3), pad=(1,1))n tower_conv2_0 = ConvFactory(net, 32, (1,1))n tower_conv2_1 = ConvFactory(tower_conv2_0, 48, (3, 3), pad=(1,1))n tower_conv2_2 = ConvFactory(tower_conv2_1, 64, (3, 3), pad=(1,1))n tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_1, tower_conv2_2])n tower_out = ConvFactory(tower_mixed, out_shape[0][1], (1,1), with_act=False)nn net += scale * tower_outn if with_act:n act = mx.symbol.Activation(data = net, act_type=act_type, attr=mirror_attr)n return actn else:n return netn

大家是不是感到很彆扭，我也覺得很彆扭，但是我一直是個『不拘小節』的工程師，對這塊不斤斤計較，所以，寫完這塊之後也覺得就成了接下來就是block17， block8， 這裡都很簡單的很類似，就不提了。

然後就接下來一段，很快就完成了：

net = repeat(tower_5b_out, 10, block35, scale=0.17, input_num_channels=320)n tower_conv = ConvFactory(net, 384, (3,3),stride=(2,2))n tower_conv1_0 = ConvFactory(net, 256, (1,1))n tower_conv1_1 = ConvFactory(tower_conv1_0, 256, (3,3), pad=(1,1))n tower_conv1_2 = ConvFactory(tower_conv1_1, 384, (3,3),stride=(2,2))n tower_pool = mx.symbol.Pooling(net, kernel=(3,3), stride=(2,2), pool_type=max)n net = mx.symbol.Concat(*[tower_conv, tower_conv1_2, tower_pool])n net = repeat(net, 20, block17, scale=0.1, input_num_channels=1088)n tower_conv = ConvFactory(net, 256, (1,1))n tower_conv0_1 = ConvFactory(tower_conv, 384, (3,3), stride=(2,2))n tower_conv1 = ConvFactory(net, 256, (1,1))n tower_conv1_1 = ConvFactory(tower_conv1, 288, (3,3), stride=(2,2))n tower_conv2 = ConvFactory(net, 256, (1,1))n tower_conv2_1 = ConvFactory(tower_conv2, 288, (3,3), pad=(1,1))n tower_conv2_2 = ConvFactory(tower_conv2_1, 320, (3,3), stride=(2,2))n tower_pool = mx.symbol.Pooling(net, kernel=(3,3), stride=(2,2), pool_type=max)n net = mx.symbol.Concat(*[tower_conv0_1, tower_conv1_1, tower_conv2_2, tower_pool])nn net = repeat(net, 9, block8, scale=0.2, input_num_channels=2080)n net = block8(net, with_act=False, input_num_channel=2080)nn net = ConvFactory(net, 1536, (1,1))n net = mx.symbol.Pooling(net, kernel=(1,1), global_pool=True, stride=(2,2), pool_type=avg)n net = mx.symbol.Flatten(net)n net = mx.symbol.Dropout(data=net,p= 0.8)n net = mx.symbol.FullyConnected(data=net,num_hidden=num_classes)n softmax = mx.symbol.SoftmaxOutput(data=net, name=softmax)n

感覺很開心，寫完了，這麼簡單，大家先忽略先忽悠所有的pad值，因為找不到沒有pad的版本，所以大家先忽略下。然後就是寫樣例測試呀，又是17flowers這個數據集，參考mxnet-101中如何把dataset轉換為binary， 首先寫個py來get到所有的圖像list，index還有他的label_index，這個很快就解決了。

具體參考我這裡的mxnet-101 然後就是拿數據開始run啦，Ready？ Go！

咦,車開不起來，不對，都是些什麼鬼？ infer_shape 有問題？ 沒事，查api tensorflow中padding是」valid」和」same」，mxnet中沒有， 沒有…，要自己計算，什麼鬼？沒有valid，same，我不會呀！！！

寫了這麼久，就不寫了？不行，找下怎麼搞定，看了tensorflow的文檔，翻了資料，same就是保證input與output保持一致，valid就無所謂，不需要設置pad，所以當tensorflow中有same的時候，就需要在mxnet中設置對應的pad值，kernel為3的時候pad=1， kernel=5，pad=2。這裡改來改去，列印出每一層網路後的shape，前後花了我大概6個小時，終於讓我一步一步debug出來了，但是不對，在repeat 10次block35後，怎麼和tf.slim的inception-resnet-v2的注釋的shape不同？

我了個擦，當時已經好像快凌晨4點了，本以為run起來了，怎麼就解釋不通呢？不會tensorflow的注釋有問題吧？我了個擦，老美真是數學有點問題，提了個issue，很快就有人fix然後commit了 may be an error in slim.nets.inception_resnet_v2 #634，不過貌似到現在還沒有被merge。

一切ok，開始run了，用17flowers，很快可以收斂，沒有更多的資源來測試更大的數據集，就直接提交了，雖然代碼很爛，但怎麼著也是一步一步寫出來的，可是，始終確實是有點問題，後來經過github的好心人指點肯定也是一個大牛，告訴我Pooling 有個global_pool來做全局的池化，我了個擦，這麼好的東西，tensorflow上可沒有，所以tensorflow上用的是通過get_shape拿到對應的tensor的width和height來做pooling，我也二筆的在mxne它裡面這樣用，所以需要input_shape_shape來infer到所在layer的shape，來做全局池化，有了這個，我還infer_shape個什麼鬼，blockxx裡面也不需要了，channel數可以直接手工計算，傳一個channel數就好了，get_symbol也可以保持和原來一樣不需要傳什麼input_data_shape啦！！！

感謝zhreshold的提示，一切都ok，更改了，但是後面mxnet的大神在重構一些代碼，還沒有merge，不過沒有關係，等他們ok了 我再把inception-resnet-v2整理下，再提pr（教練，我想當mxnet contributor）。

def block35(net, input_num_channels, scale=1.0, with_act=True, act_type=relu, mirror_attr={}):n tower_conv = ConvFactory(net, 32, (1,1))n tower_conv1_0 = ConvFactory(net, 32, (1,1))n tower_conv1_1 = ConvFactory(tower_conv1_0, 32, (3, 3), pad=(1,1))n tower_conv2_0 = ConvFactory(net, 32, (1,1))n tower_conv2_1 = ConvFactory(tower_conv2_0, 48, (3, 3), pad=(1,1))n tower_conv2_2 = ConvFactory(tower_conv2_1, 64, (3, 3), pad=(1,1))n tower_mixed = mx.symbol.Concat(*[tower_conv, tower_conv1_1, tower_conv2_2])n tower_out = ConvFactory(tower_mixed, input_num_channels, (1,1), with_act=False)nn net += scale * tower_outn if with_act:n act = mx.symbol.Activation(data = net, act_type=act_type, attr=mirror_attr)n return actn else:n return netn

一直到這裡，inception-resnet-v2就寫出來了，但是只是測試了小數據集，後來在zhihu上偶遇李沐大神，果斷上去套近乎，最後拿到一個一台機器，就在測大一點的數據集，其實也不大，102flowers，之後會請沐神幫忙擴展一個大點的盤來放下ImageNet，測試一下性能，不過現在102flowers也還行，效果還不錯。

丹成

金丹品階高低，以丹紋記，不同煉丹材料丹紋不同，評判標準也不同，acc是最常用判斷金丹品階高低的手段。

將102flower按9：1分成訓練集和驗證集，設置300個epoch(數據集比較小，貌似設置多點epoch才能有比較好的性能，看有小夥伴用inception-bn在imagenet上只需要50個epoch)，網路inception-resnet-v2確實大，如此小的數據集上300 epoch大概也需要1天，不過對比tensorflow那是快多了。

編不下去了（predict）

這裡，會簡單地寫一個inference的例子，算作學習如果使用訓練好的model，注意還是最好使用python的opencv，因為mxnet官方是用的opencv，使用cv2這個庫，我在網上找的使用skimage的庫，做出來的始終有問題，應該是brg2rgb的問題，使用cv2的cv2.cvtColor(img, cv2.COLOR_BGR2RGB之後會成功:

import mxnet as mxn import loggingn import numpy as npn import cv2n import scipy.io as sionn logger = logging.getLogger()n logger.setLevel(logging.DEBUG)n num_round = 260n prefix = "102flowers"n model = mx.model.FeedForward.load(prefix, num_round, ctx=mx.cpu(), numpy_batch_size=1)n # synset = [l.strip() for l in open(Inception/synset.txt).readlines()]nnn def PreprocessImage(path, show_img=False):n # load imagen img = cv2.imread(path)n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)n mean_img = mx.nd.load(mean.bin).values()[0].asnumpy()n print img.shapen print mean_img.shapen img = cv2.resize(img,(299,299))n img = np.swapaxes(img, 0, 2)n img = np.swapaxes(img, 1, 2)n img = img -mean_imgn img = img[np.newaxis, :]n print img.shapenn return imgnn right = 0n sum = 0n with open(test.lst, r) as fread:n for line in fread.readlines()[:20]:n sum +=1n batch = ../day2/102flowers/ + line.split("t")[2].strip("n")n print batchn batch = PreprocessImage(batch, False)n prob = model.predict(batch)[0]n pred = np.argsort(prob)[::-1]n # # Get top1 labeln # top1 = synset[pred[0]]n top_1 = pred[0]n if top_1 == int(line.split("t")[1]):n print top1 rightn right += 1nn print top 1 accuracy: %f %(right/(1.0*sum))n

使用第260個epoch的模型weight，這裡因為手賤刪除了9：1時的test.lst，只能用7：3是的test.lst暫時做計算，最後accuracy應該會比較偏高，不過這不是重點。

總結（繼續編不下去了）

在這樣一次暢快淋漓的mxnet之旅後，總結一下遇到的幾個坑，與大家分享：

• 無法直接拿到tensor的shape信息，通過infer_shape，在設計代碼時走了很多；
• im2rec時，準備的train.lst, test.lst未shuffle，在102flowers上我都沒有發覺，在後面做鑒黃的training的時候發現開始training accuracy，分析可能是train.lst未shuffle的問題（以為在ImageRecordIter中有shuffle參數，就不需要），改了後沒有training accuracy從開始就為1的情況；
• pad值的問題，翻閱了很多資料才解決，文檔也沒有特別多相關的，對於我這種從tensorflow轉mxnet的小夥伴來說是個比較大的坑；
• predict的問題，找了mxnet github的源上的example，並不能成功，在找官網上的example發現使用的是cv2，並不是一些例子當中的skimage，考慮到mxnet在安裝時需要opencv，可能cv2和skimage在一些標準上有差異，就改用cv2的predict版本，還有讀入圖片之後要cv2.cvtColor(img, cv2.COLOR_BGR2RGB).

• 還是predict的問題，在mxnet中，構造ImageRecordIter時沒有指定mean.bin，但是並不是說計算的時候不會減到均值圖片在訓練，開始誤解為不需要減到均值圖片，後來發現一直不正確，考慮到train的時候會自己生成mean.bin，猜測可能是這裡的問題，通過mean_img = mx.nd.load(mean.bin).values()[0].asnumpy()讀入後，在原始圖片減去均值圖，結果ok；但整個流程相對於tf.slim的predict還是比較複雜的。

優點

• 速度快，速度快，速度快，具體指沒有做測量，但是相對於tensorflow至少兩到三倍；

• 佔用內存低， 同樣batch和模型，12g的顯存，tf會爆，但是mxnet只需要佔用7g多點；
• im2rec很方便，相對於tensorflow下tfrecord需要寫部分代碼，更容易入手，但是切記自己生成train.lst, test.lst的時候要shuffle；

• Pooling下的global_pool是個好東西，tensorflow沒有；

That is all！之後會在ImageNet Dataset做一下測試，感覺會更有意思。

坐等更新，期待！

相關閱讀推薦：

機器學習進階筆記之六 | 深入理解Fast Neural Style

機器學習進階筆記之五 | 深入理解VGGResidual Network

機器學習進階筆記之四 | 深入理解GoogLeNet

機器學習進階筆記之三 | 深入理解Alexnet

機器學習進階筆記之二 | 深入理解Neural Style

機器學習進階筆記之一 | TensorFlow安裝與入門

關於作者：

Burness（@段石石 ）， UCloud平台研發中心深度學習研發工程師，tflearn Contributor & tensorflow Contributor，做過電商推薦、精準化營銷相關演算法工作，專註於分散式深度學習框架、計算機視覺演算法研究，平時喜歡玩玩演算法，研究研究開源的項目，偶爾也會去一些數據比賽打打醬油，生活中是個極客，對新技術、新技能痴迷。

你可以在Github上找到他：http://hacker.duanshishi.com/

「UCloud機構號」將獨家分享雲計算領域的技術洞見、行業資訊以及一切你想知道的相關訊息。

歡迎提問&求關注 o(*////▽////*)q～

以上。