0.Pre-trained Model------resnet_v2_50
（1）简介
resnet v1和v2总结如下，首先给出resnet v2的paper里面kaiming大神给出的不同的结构对比：
图a为resnet v1的结构，图e为resnet v2的结构。（weight为conv层），左分支为identity分支，右分支为residual分支。
图的结构有两个特点：1、conv-bn-relu；2、最后的relu在addition后面。
（2）代码
具体代码实现见Tensorflow官方代码的模型库中：
https://github.com/tensorflow/models/tree/master/research/slim/nets
（3）加载方式
from tensorflow.contrib.slim.nets import resnet_v2（4）Pre-trained Model检查点文件下载
官方下载地址：https://github.com/tensorflow/models/tree/master/research/slim#Pretrained
1.自己的手势数据
（1）简介
输入的是类似上图所示的手势图片数据，总共有6个类。所给的数据已经加工过，是‘.h5’格式的数据。有1080张图片，120张测试数据。每一张图片是一个64x64的RGB图片。具体的数据格式为：
number of training examples = 1080
number of test examples = 120
X_train shape: (1080, 64, 64, 3)
Y_train shape: (1080, 6)
X_test shape: (120, 64, 64, 3)
Y_test shape: (120, 6)
x train max, 0.956; x train min, 0.015
x test max, 0.94; x test min, 0.011（2）下载地址
链接：https://pan.baidu.com/s/1iA004kLU1gocvA-gaiwSWw
提取码：sqj3
下载后如下格式：
（3）使用方法
包括训练时需要使用到的数据加载函数，具体代码如下，划分为utils.py文件：
import os
import numpy as np
import tensorflow as tf
import h5py
import math
def load_dataset(path):
train_dataset = h5py.File(path+'/train_signs.h5', "r")
train_set_x_orig = np.array(train_dataset["train_set_x"][:]) # your train set features
train_set_y_orig = np.array(train_dataset["train_set_y"][:]) # your train set labels
test_dataset = h5py.File(path+'/test_signs.h5', "r")
test_set_x_orig = np.array(test_dataset["test_set_x"][:]) # your test set features
test_set_y_orig = np.array(test_dataset["test_set_y"][:]) # your test set labels
classes = np.array(test_dataset["list_classes"][:]) # the list of classes
train_set_y_orig = train_set_y_orig.reshape((1, train_set_y_orig.shape[0]))
test_set_y_orig = test_set_y_orig.reshape((1, test_set_y_orig.shape[0]))
return train_set_x_orig, train_set_y_orig, test_set_x_orig, test_set_y_orig, classes
def process_orig_datasets(datasets):
"""
normalize x_train and convert y_train to one hot.
:param datasets: X_train_orig, Y_train_orig, X_test_orig, Y_test_orig, classes
:return: X_train, Y_train, X_test, Y_test
"""
X_train_orig, Y_train_orig, X_test_orig, Y_test_orig, classes = datasets
# Normalize image vectors
X_train = X_train_orig / 255.
X_test = X_test_orig / 255.
# Convert training and test labels to one hot matrices
Y_train = convert_to_one_hot(Y_train_orig, 6).T
Y_test = convert_to_one_hot(Y_test_orig, 6).T
return X_train, Y_train, X_test, Y_test
def random_mini_batches(X, Y, mini_batch_size=64, seed=None):
"""
Creates a list of random minibatches from (X, Y)
Arguments:
X -- input data, of shape (input size, number of examples) (m, Hi, Wi, Ci)
Y -- true "label" vector (containing 0 if cat, 1 if non-cat), of shape (1, number of examples) (m, n_y)
mini_batch_size - size of the mini-batches, integer
seed -- this is only for the purpose of grading, so that you're "random minibatches are the same as ours.
Returns:
mini_batches -- list of synchronous (mini_batch_X, mini_batch_Y)
"""
m = X.shape[0] # number of training examples
mini_batches = []
np.random.seed(seed)
# Step 1: Shuffle (X, Y)
permutation = list(np.random.permutation(m))
shuffled_X = X[permutation, :, :, :]
shuffled_Y = Y[permutation, :]
# Step 2: Partition (shuffled_X, shuffled_Y). Minus the end case.
num_complete_minibatches = math.floor(
m / mini_batch_size) # number of mini batches of size mini_batch_size in your partitionning
for k in range(0, num_complete_minibatches):
mini_batch_X = shuffled_X[k * mini_batch_size: k * mini_batch_size + mini_batch_size, :, :, :]
mini_batch_Y = shuffled_Y[k * mini_batch_size: k * mini_batch_size + mini_batch_size, :]
mini_batch = (mini_batch_X, mini_batch_Y)
mini_batches.append(mini_batch)
# Handling the end case (last mini-batch < mini_batch_size)
if m % mini_batch_size != 0:
mini_batch_X = shuffled_X[num_complete_minibatches * mini_batch_size: m, :, :, :]
mini_batch_Y = shuffled_Y[num_complete_minibatches * mini_batch_size: m, :]
mini_batch = (mini_batch_X, mini_batch_Y)
mini_batches.append(mini_batch)
return mini_batches
def convert_to_one_hot(Y, C):
Y = np.eye(C)[Y.reshape(-1)].T
return Y2、加载Pre-trained resnet_v2_50，恢复指定权重并继续重新训练(finetuning)
（1）训练代码
import tensorflow as tf
import time
from tensorflow.contrib.slim.nets import resnet_v2
from tensorflow.contrib.slim import nets
from tensorflow.python import pywrap_tensorflow
from utils import *
slim = tf.contrib.slim
tf.reset_default_graph()
data_dir = 'E:/Tfexercise/Resnet-Finetune/datasets' #自己手势数据的文件夹
model_save_path='E:/Tfexercise/Resnet-Finetune/save_model/' #微调后的模型保存地址
# 加载自己的的数据，函数见utils.py
orig_data = load_dataset(data_dir)
X_train, Y_train, X_test, Y_test = process_orig_datasets(orig_data)
mini_batches = random_mini_batches(X_train, Y_train, mini_batch_size=32, seed=None)
#pre-trained model保存地址
model_path="E:/Tfexercise/Resnet-Finetune/resnet_v2_50_2017_04_14/resnet_v2_50.ckpt"
images=tf.placeholder(tf.float32,(None,64,64,3),'imnput_images')
labels = tf.placeholder(tf.int64, [None, 6])
net,endpoins=resnet_v2.resnet_v2_50(images,is_training=True)
#自己添加的网络层，作用域命名为“finetune”
with tf.variable_scope('finetune'):
flatten = tf.layers.flatten(net)
x = tf.layers.dense(flatten, units=50, activation=tf.nn.relu)
keep_prob = tf.placeholder(tf.float32)
x = tf.nn.dropout(x, keep_prob)
logits = tf.layers.dense(x, units=6, activation=tf.nn.softmax)
# cross_entropy = tf.losses.sparse_softmax_cross_entropy(labels=y_, logits=y_conv)
y=tf.nn.softmax(logits)
cross_entropy = tf.losses.softmax_cross_entropy(onehot_labels=labels, logits=logits)
cross_entropy_cost = tf.reduce_mean(cross_entropy)
correct_predicion=tf.equal(tf.argmax(y,1),tf.argmax(labels,1))
accuracy=tf.reduce_mean(tf.cast(correct_predicion,'float'))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy_cost) #Resnet_v2_50和自己构建的模型层都训练
'''
Train 训练数据进行训练的过程
'''
saver = tf.train.Saver()
reader = pywrap_tensorflow.NewCheckpointReader(model_path)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
var_list=tf.trainable_variables()
print(var_list)
print(len(var_list))
if isinstance(var_list, dict):
var_dict = var_list
print('true')
else:
var_dict = {var.op.name: var for var in var_list}
print(len(var_dict))
# print(var_dict)
print('False')
available_vars = {}
for var in var_dict:
if reader.has_tensor(var):
print("nice")
available_vars[var] = var_dict[var]
print(available_vars[var],var_dict[var])
var_list = available_vars
if var_list:
saver_restore = tf.train.Saver(var_list)
saver = tf.train.Saver(tf.global_variables())
saver_restore.restore(sess, model_path)
for i in range(10000):
X_mini_batch, Y_mini_batch = mini_batches[np.random.randint(0, len(mini_batches))]
train_step.run(feed_dict={images: X_mini_batch, labels: Y_mini_batch, keep_prob: 0.5})
if i % 20 == 0:
train_cost = sess.run(accuracy, feed_dict={images: X_mini_batch,
labels: Y_mini_batch, keep_prob: 1.0})
print('step %d, training cost %g' % (i, train_cost))
if (i + 1) % 100 == 0:
saver.save(sess, model_save_path, global_step=i + 1)
print('save mode to {}'.format(model_save_path))
'''
Evaluate 测试数据用于测试的过程
'''
# saver = tf.train.Saver(tf.global_variables())
# reader = pywrap_tensorflow.NewCheckpointReader(tf.train.latest_checkpoint(model_save_path))
# with tf.Session() as sess2:
# sess2.run(tf.global_variables_initializer())
# var_list=tf.trainable_variables()
# for var_i in var_list:
# if 'weights' in var_i.op.name:
# print(var_i.name)
# print(len(var_list))
# if isinstance(var_list, dict):
# var_dict = var_list
# print('true')
# else:
# var_dict = {var.op.name: var for var in var_list}
# print(len(var_dict))
# print('False')
# available_vars = {}
# for var in var_dict:
# if reader.has_tensor(var):
# print("nice")
# available_vars[var] = var_dict[var]
# print(available_vars)
#
# saver.restore(sess2, tf.train.latest_checkpoint(model_save_path))
# accu = sess2.run(accuracy, feed_dict={images: X_test, labels: Y_test, keep_prob: 1.0})
# print('accuracy %g' % (accu))
部分结果显示：
（2）测试代码
只需将上面代码的“Train”模块代码进行注释掉，将“Evaluate”模块代码显示就可以。
最终测试结果：
3、加载Pre-trained resnet_v2_50，恢复指定权重并使optimizer只训练自己添加的模型层（指定层、变量）
（1）训练代码
import tensorflow as tf
import time
from tensorflow.contrib.slim.nets import resnet_v2
from tensorflow.contrib.slim import nets
from tensorflow.python import pywrap_tensorflow
from utils import *
slim = tf.contrib.slim
tf.reset_default_graph()
data_dir = 'E:/Tfexercise/Resnet-Finetune/datasets' #自己手势数据的文件夹
model_save_path='E:/Tfexercise/Resnet-Finetune/save_model/' #微调后的模型保存地址
# 加载自己的的数据，函数见utils.py
orig_data = load_dataset(data_dir)
X_train, Y_train, X_test, Y_test = process_orig_datasets(orig_data)
mini_batches = random_mini_batches(X_train, Y_train, mini_batch_size=32, seed=None)
#pre-trained model保存地址
model_path="E:/Tfexercise/Resnet-Finetune/resnet_v2_50_2017_04_14/resnet_v2_50.ckpt"
images=tf.placeholder(tf.float32,(None,64,64,3),'imnput_images')
labels = tf.placeholder(tf.int64, [None, 6])
net,endpoins=resnet_v2.resnet_v2_50(images,is_training=True)
#自己添加的网络层，作用域命名为“finetune”
with tf.variable_scope('finetune'):
flatten = tf.layers.flatten(net)
x = tf.layers.dense(flatten, units=50, activation=tf.nn.relu)
keep_prob = tf.placeholder(tf.float32)
x = tf.nn.dropout(x, keep_prob)
logits = tf.layers.dense(x, units=6, activation=tf.nn.softmax)
# cross_entropy = tf.losses.sparse_softmax_cross_entropy(labels=y_, logits=y_conv)
y=tf.nn.softmax(logits)
cross_entropy = tf.losses.softmax_cross_entropy(onehot_labels=labels, logits=logits)
cross_entropy_cost = tf.reduce_mean(cross_entropy)
correct_predicion=tf.equal(tf.argmax(y,1),tf.argmax(labels,1))
accuracy=tf.reduce_mean(tf.cast(correct_predicion,'float'))
'''
变化部分
'''
loss_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='finetune') ##关键！获取指定scope下的变量
print(loss_vars)
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy_cost,var_list=loss_vars)
'''
Train 训练数据进行训练的过程
'''
saver = tf.train.Saver()
reader = pywrap_tensorflow.NewCheckpointReader(model_path)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
var_list=tf.trainable_variables()
print(var_list)
print(len(var_list))
if isinstance(var_list, dict):
var_dict = var_list
print('true')
else:
var_dict = {var.op.name: var for var in var_list}
print(len(var_dict))
# print(var_dict)
print('False')
available_vars = {}
for var in var_dict:
if reader.has_tensor(var):
print("nice")
available_vars[var] = var_dict[var]
print(available_vars[var],var_dict[var])
var_list = available_vars
if var_list:
saver_restore = tf.train.Saver(var_list)
saver = tf.train.Saver(tf.global_variables())
saver_restore.restore(sess, model_path)
for i in range(10000):
X_mini_batch, Y_mini_batch = mini_batches[np.random.randint(0, len(mini_batches))]
train_step.run(feed_dict={images: X_mini_batch, labels: Y_mini_batch, keep_prob: 0.5})
if i % 20 == 0:
train_cost = sess.run(accuracy, feed_dict={images: X_mini_batch,
labels: Y_mini_batch, keep_prob: 1.0})
print('step %d, training cost %g' % (i, train_cost))
if (i + 1) % 100 == 0:
saver.save(sess, model_save_path, global_step=i + 1)
print('save mode to {}'.format(model_save_path))
'''
Evaluate 测试数据用于测试的过程
'''
# saver = tf.train.Saver(tf.global_variables())
# reader = pywrap_tensorflow.NewCheckpointReader(tf.train.latest_checkpoint(model_save_path))
# with tf.Session() as sess2:
# sess2.run(tf.global_variables_initializer())
# var_list=tf.trainable_variables()
# for var_i in var_list:
# if 'weights' in var_i.op.name:
# print(var_i.name)
# print(len(var_list))
# if isinstance(var_list, dict):
# var_dict = var_list
# print('true')
# else:
# var_dict = {var.op.name: var for var in var_list}
# print(len(var_dict))
# print('False')
# available_vars = {}
# for var in var_dict:
# if reader.has_tensor(var):
# print("nice")
# available_vars[var] = var_dict[var]
# print(available_vars)
#
# saver.restore(sess2, tf.train.latest_checkpoint(model_save_path))
# accu = sess2.run(accuracy, feed_dict={images: X_test, labels: Y_test, keep_prob: 1.0})
# print('accuracy %g' % (accu))
显然变化部分仅仅是：
'''
变化部分
'''
loss_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='finetune') ##关键！获取指定scope下的变量
print(loss_vars)
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy_cost,var_list=loss_vars)最终训练结果显示：
（2）测试代码
只需将上面代码的“Train”模块代码进行注释掉，将“Evaluate”模块代码显示就可以。
最终测试结果：
4、对比
对比第2、3节的微调方法，可以看出，在本数据集上，在epoch相同的情况下，加载Pre-trained Resnet_v2_50模型后，进行包括Resnet在内的全部模型变量重新训练所达到的效果更佳。
1285
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
