用tensorlayer导入Slim模型迁移学习
上一篇博客【用tensorflow迁移学习猫狗分类】笔者讲到用tensorlayer的【VGG16模型】迁移学习图像分类，那麽问题来了，tensorlayer没提供的模型怎么办呢？别担心，tensorlayer提供了tensorflow中的【slim模型】导入功能，代码例子在tutorial_inceptionV3_tfslim。
那么什么是slim？slim到底有什么用？
slim是一个使构建，训练，评估神经网络变得简单的库。它可以消除原生tensorflow里面很多重复的模板性的代码，让代码更紧凑，更具备可读性。另外slim提供了很多计算机视觉方面的著名模型（VGG, AlexNet等），我们不仅可以直接使用，甚至能以各种方式进行扩展。（笔者注：总之功能跟tensorlayer差不多嘛）更多介绍可以看这篇文章：【Tensorflow】辅助工具篇——tensorflow slim(TF-Slim)介绍】
要进行迁移学习，首先需要slim模型代码以及预训练好的权重参数，这些谷歌都有提供下载，可以看到主页下面有各个模型以及在imagenet训练集下的参数地址。
列表还列出了各个模型的top1、top5的正确率，模型很多了。
好了我们下载Inception-ResNet-v2以及inception_resnet_v2_2016_08_30.tar.gz，py文件和解压出来的.ckpt文件放到项目根目录下面。至于为什么不用tensorlayer例子提供的Inception V3？因为Inception-ResNet-v2正确率高啊。（哈哈真正原因最后来讲）。
我们依旧进行猫狗分类，按照教程导入模型修改num_classes再导入训练数据，直接训练是会报错的，因为最后的Logits层几个参数在恢复时维度不匹配。
最后几个参数是不能恢复了，笔者也没有找到选择性恢复.ckpt参数的tensorflow方法。怎么办呢？幸好群里面有位朋友提供了一个方法，参见【Tensorflow 迁移学习】：
主要思想是：先把所有.ckpt参数恢复成npz格式，再选择恢复npz中的参数，恢复npz中的参数就跟前一篇博客操作一样的了。
所以整个过程分两步走：
1.将参数恢复然后保存为npz格式：
下面是具体代码：
import os import time from recordutil import * import numpy as np # from tensorflow.contrib.slim.python.slim.nets.resnet_v2 import resnet_v2_152 # from tensorflow.contrib.slim.python.slim.nets.vgg import vgg_16 import skimage import skimage.io import skimage.transform import tensorflow as tf from tensorlayer.layers import * # from scipy.misc import imread, imresize # from tensorflow.contrib.slim.python.slim.nets.alexnet import alexnet_v2 from inception_resnet_v2 import (inception_resnet_v2_arg_scope, inception_resnet_v2) from scipy.misc import imread, imresize from tensorflow.python.ops import variables import tensorlayer as tl slim = tf.contrib.slim try: from data.imagenet_classes import * except Exception as e: raise Exception( "{} / download the file from: https://github.com/zsdonghao/tensorlayer/tree/master/example/data".format(e)) n_epoch = 200 learning_rate = 0.0001 print_freq = 2 batch_size = 32 ## InceptionV3 / All TF-Slim nets can be merged into TensorLayer x = tf.placeholder(tf.float32, shape=[None, 299, 299, 3]) # 输出 y_ = tf.placeholder(tf.int32, shape=[None, ], name=\'y_\') net_in = tl.layers.InputLayer(x, name=\'input_layer\') with slim.arg_scope(inception_resnet_v2_arg_scope()): network = tl.layers.SlimNetsLayer( prev_layer=net_in, slim_layer=inception_resnet_v2, slim_args={ \'num_classes\': 1001, \'is_training\': True, }, name=\'InceptionResnetV2\' # <-- the name should be the same with the ckpt model ) # network = fc_layers(net_cnn) sess = tf.InteractiveSession() network.print_params(False) # network.print_layers() saver = tf.train.Saver() # 加载预训练的参数 # tl.files.assign_params(sess, npz, network) tl.layers.initialize_global_variables(sess) saver.restore(sess, "inception_resnet_v2.ckpt") print("Model Restored") all_params = sess.run(network.all_params) np.savez(\'inception_resnet_v2.npz\', params=all_params) sess.close()
执行成功之后，我们得到模型所有的908个参数。
2.部分恢复npz参数然后训练模型：
首先我们修改模型最后一层参数，由于进行的是2分类学习，所以做如下修改：
with slim.arg_scope(inception_resnet_v2_arg_scope()): network = tl.layers.SlimNetsLayer( prev_layer=net_in, slim_layer=inception_resnet_v2, slim_args={ \'num_classes\': 2, \'is_training\': True, }, name=\'InceptionResnetV2\' # <-- the name should be the same with the ckpt model )
num_classes改为2，is_training为True。
接着定义输入输出以及损失函数：
sess = tf.InteractiveSession() # saver = tf.train.Saver() y = network.outputs y_op = tf.argmax(tf.nn.softmax(y), 1) cost = tl.cost.cross_entropy(y, y_, name=\'cost\') correct_prediction = tf.equal(tf.cast(tf.argmax(y, 1), tf.float32), tf.cast(y_, tf.float32)) acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
下面是定义训练参数，我们只训练最后一层的参数，打印参数出来我们看到：
[TL] param 900: InceptionResnetV2/AuxLogits/Conv2d_2a_5x5/weights:0 (5, 5, 128, 768) float32_ref [TL] param 901: InceptionResnetV2/AuxLogits/Conv2d_2a_5x5/BatchNorm/beta:0 (768,) float32_ref [TL] param 902: InceptionResnetV2/AuxLogits/Conv2d_2a_5x5/BatchNorm/moving_mean:0 (768,) float32_ref [TL] param 903: InceptionResnetV2/AuxLogits/Conv2d_2a_5x5/BatchNorm/moving_variance:0 (768,) float32_ref [TL] param 904: InceptionResnetV2/AuxLogits/Logits/weights:0 (768, 2) float32_ref [TL] param 905: InceptionResnetV2/AuxLogits/Logits/biases:0 (2,) float32_ref [TL] param 906: InceptionResnetV2/Logits/Logits/weights:0 (1536, 2) float32_ref [TL] param 907: InceptionResnetV2/Logits/Logits/biases:0 (2,) float32_ref [TL] num of params: 56940900
从param 904开始训练就行了，参数恢复到param 903
下面是训练函数以及恢复部分参数，加载样本数据：
# 定义 optimizer train_params = network.all_params[904:] print(\'训练参数:\', train_params) # # 加载预训练的参数 # tl.files.assign_params(sess, params, network) train_op = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost, var_list=train_params) img, label = read_and_decode("D:\\001-Python\\train299.tfrecords") # 使用shuffle_batch可以随机打乱输入 X_train, y_train = tf.train.shuffle_batch([img, label], batch_size=batch_size, capacity=200, min_after_dequeue=100) tl.layers.initialize_global_variables(sess) params = tl.files.load_npz(\'\', \'inception_resnet_v2.npz\') params = params[0:904] print(\'当前参数大小：\', len(params)) tl.files.assign_params(sess, params=params, network=network)
下面依旧是训练模型的代码，跟上一篇一样：
# # 训练模型 coord = tf.train.Coordinator() threads = tf.train.start_queue_runners(sess=sess, coord=coord) step = 0 filelist = getfilelist() for epoch in range(n_epoch): start_time = time.time() val, l = sess.run([X_train, y_train])#next_data(filelist, batch_size) # for X_train_a, y_train_a in tl.iterate.minibatches(val, l, batch_size, shuffle=True): sess.run(train_op, feed_dict={x: X_train_a, y_: y_train_a}) if epoch + 1 == 1 or (epoch + 1) % print_freq == 0: print("Epoch %d of %d took %fs" % (epoch + 1, n_epoch, time.time() - start_time)) train_loss, train_acc, n_batch = 0, 0, 0 for X_train_a, y_train_a in tl.iterate.minibatches(val, l, batch_size, shuffle=True): err, ac = sess.run([cost, acc], feed_dict={x: X_train_a, y_: y_train_a}) train_loss += err train_acc += ac n_batch += 1 print(" train loss: %f" % (train_loss / n_batch)) print(" train acc: %f" % (train_acc / n_batch)) # tl.files.save_npz(network.all_params, name=\'model_vgg_16_2.npz\', sess=sess) coord.request_stop() coord.join(threads)
batchsize为20训练200代，部分结果如下：
Epoch 156 of 200 took 12.568609s train loss: 0.382517 train acc: 0.950000 Epoch 158 of 200 took 12.457161s train loss: 0.382509 train acc: 0.850000 Epoch 160 of 200 took 12.385407s train loss: 0.320393 train acc: 1.000000 Epoch 162 of 200 took 12.489218s train loss: 0.480686 train acc: 0.700000 Epoch 164 of 200 took 12.388841s train loss: 0.329189 train acc: 0.850000 Epoch 166 of 200 took 12.446472s train loss: 0.379127 train acc: 0.900000 Epoch 168 of 200 took 12.888571s train loss: 0.365938 train acc: 0.900000 Epoch 170 of 200 took 12.850605s train loss: 0.353434 train acc: 0.850000 Epoch 172 of 200 took 12.855129s train loss: 0.315443 train acc: 0.950000 Epoch 174 of 200 took 12.906666s train loss: 0.460817 train acc: 0.750000 Epoch 176 of 200 took 12.830738s train loss: 0.421025 train acc: 0.900000 Epoch 178 of 200 took 12.852572s train loss: 0.418784 train acc: 0.800000 Epoch 180 of 200 took 12.951322s train loss: 0.316057 train acc: 0.950000 Epoch 182 of 200 took 12.866213s train loss: 0.363328 train acc: 0.900000 Epoch 184 of 200 took 13.012520s train loss: 0.379462 train acc: 0.850000 Epoch 186 of 200 took 12.934583s train loss: 0.472857 train acc: 0.750000 Epoch 188 of 200 took 13.038168s train loss: 0.236005 train acc: 1.000000 Epoch 190 of 200 took 13.056378s train loss: 0.266042 train acc: 0.950000 Epoch 192 of 200 took 13.016137s train loss: 0.255430 train acc: 0.950000 Epoch 194 of 200 took 13.013147s train loss: 0.422342 train acc: 0.900000 Epoch 196 of 200 took 12.980659s train loss: 0.353984 train acc: 0.900000 Epoch 198 of 200 took 13.033676s train loss: 0.320018 train acc: 0.950000 Epoch 200 of 200 took 12.945982s train loss: 0.288049 train acc: 0.950000
好了，迁移学习Inception-ResNet-v2结束。
作者说SlimNetsLayer是能导入任何Slim Model的。笔者已经验证过导入Inception-ResNet-v2和VGG16成功，Inception V3导入后训练了两三天，正确率一直在10到70之间波动（跟笔者的心情一样不稳定），笔者一直找不出原因，心累，希望哪位朋友再去验证一下Inception V3咯。
版权声明：本文为zengfanlin原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。