-
Notifications
You must be signed in to change notification settings - Fork 12
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
feiyuxiao
committed
Dec 15, 2018
1 parent
3f7b835
commit 5d0d83f
Showing
14 changed files
with
584 additions
and
582 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,9 @@ | ||
| # ReadMe | ||
| 该模块为整个回归系统的回归器。分为4个部分，分别是：数据处理、Resnet搭建、网络训练、模型预测 | ||
| 模块的主要功能为： | ||
|
|
||
| 1. 从本地目录获取所有清洗好的数据，将数据按照Resnet的要求加载到内存。 | ||
| 2. 对数据自动分为训练数据和预测数据，其中训练数据喂入Resnet进行模型训练。 | ||
| 3. 对训练好的模型进行预测结果的输出。 | ||
| 4. 支持断点续训，训练过程中模型会每10轮保存一次，可以自由跳转到已经保存好的模型进行继续训练或者模型输出 | ||
| 5. 支持迁移学习。即利用已经训练好的模型训练新的场景，提高训练的效率和降低对样本数量的需求。 |
342 changes: 171 additions & 171 deletions
342
Code/ResNet/ResnetTHU.py → Code/Processor/ResNet/ResnetTHU.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -1,171 +1,171 @@ | ||
| ''' | ||
| 该模块为前向传播网路，定义网络结构 | ||
| ''' | ||
| import tensorflow as tf | ||
|
|
||
| CONV_SIZE = 5#卷积核大小 | ||
|
|
||
| # 各层通道数 | ||
| CONV1_KERNEL_NUM = 64 | ||
| CONV2_KERNEL_NUM = 64 | ||
| CONV3_KERNEL_NUM = 64 | ||
| CONV4_KERNEL_NUM = 64 | ||
| CONV5_KERNEL_NUM = 64 | ||
| CONV6_KERNEL_NUM = 64 | ||
| CONV7_KERNEL_NUM = 64 | ||
| CONV8_KERNEL_NUM = 64 | ||
| CONV9_KERNEL_NUM = 64 | ||
| CONV10_KERNEL_NUM = 64 | ||
| CONV11_KERNEL_NUM = 64 | ||
| CONV12_KERNEL_NUM = 1 | ||
|
|
||
| # 生成卷积核，权重w | ||
| def get_weight(shape, regularizer=None): | ||
| '''生成卷积核，卷积核形状shape,权重w由正态分布产生''' | ||
| w = tf.Variable(tf.truncated_normal(shape, stddev=0.01)) | ||
| if regularizer != None: | ||
| tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(regularizer)(w)) | ||
| return w | ||
|
|
||
| # 生成偏置项 | ||
| def get_bias(shape): | ||
| '''初始化0偏置项b''' | ||
| b = tf.Variable(tf.zeros(shape)) | ||
| return b | ||
|
|
||
| def conv2d(x,w): | ||
| ''' | ||
| 定义卷积操作 | ||
| x:输入图片 | ||
| w:卷积核 | ||
| strides:移动步长，左右上下移动步长为1 | ||
| padding:0填充操作，为了卷积操作后图片维度与原来相等 | ||
| ''' | ||
| return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME') | ||
|
|
||
|
|
||
| # 定义前向传播网络，按照文章的ResNet | ||
| def forward(x, channel, regularizer=None): | ||
|
|
||
| # 第1个卷积层 | ||
| conv1_w = get_weight([CONV_SIZE, CONV_SIZE, channel, CONV1_KERNEL_NUM], regularizer) | ||
| conv1_b = get_bias([CONV1_KERNEL_NUM]) | ||
| conv1 = conv2d(x, conv1_w) | ||
| conv1_op = tf.nn.bias_add(conv1, conv1_b) | ||
|
|
||
| # 第1个激活 | ||
| relu1 = tf.nn.relu(conv1_op) | ||
|
|
||
| # 第2个卷积层 | ||
| conv2_w = get_weight([CONV_SIZE, CONV_SIZE, CONV1_KERNEL_NUM, CONV2_KERNEL_NUM], regularizer) | ||
| conv2_b = get_bias([CONV2_KERNEL_NUM]) | ||
| conv2 = conv2d(relu1, conv2_w) | ||
| conv2_op = tf.nn.bias_add(conv2, conv2_b) | ||
|
|
||
| # 第2个激活 | ||
| relu2 = tf.nn.relu(conv2_op) | ||
|
|
||
| # 第3个卷积层 | ||
| conv3_w = get_weight([CONV_SIZE, CONV_SIZE, CONV2_KERNEL_NUM, CONV3_KERNEL_NUM], regularizer) | ||
| conv3_b = get_bias([CONV3_KERNEL_NUM]) | ||
| conv3 = conv2d(relu2, conv3_w) | ||
| conv3_op = tf.nn.bias_add(conv3,conv3_b) | ||
|
|
||
| # 第1个add层 | ||
| add1 = tf.add(conv3_op, conv1_op) | ||
|
|
||
| # 第3个激活层 | ||
| relu3 = tf.nn.relu(add1) | ||
|
|
||
| # 第4个卷积层 | ||
| conv4_w = get_weight([CONV_SIZE, CONV_SIZE, CONV3_KERNEL_NUM, CONV4_KERNEL_NUM], regularizer) | ||
| conv4_b = get_bias([CONV4_KERNEL_NUM]) | ||
| conv4 = conv2d(relu3, conv4_w) | ||
| conv4_op = tf.nn.bias_add(conv4, conv4_b) | ||
|
|
||
| # 第4个激活层 | ||
| relu4 = tf.nn.relu(conv4_op) | ||
|
|
||
| # 第5个卷积层 | ||
| conv5_w = get_weight([CONV_SIZE, CONV_SIZE, CONV4_KERNEL_NUM, CONV5_KERNEL_NUM], regularizer) | ||
| conv5_b = get_bias([CONV5_KERNEL_NUM]) | ||
| conv5 = conv2d(relu4, conv5_w) | ||
| conv5_op = tf.nn.bias_add(conv5, conv5_b) | ||
|
|
||
| # 第2个add层 | ||
| add2 = tf.add(conv5_op, add1) | ||
|
|
||
| # 第5个激活层 | ||
| relu5 = tf.nn.relu(add2) | ||
|
|
||
| # 第6个卷积层 | ||
| conv6_w = get_weight([CONV_SIZE, CONV_SIZE, CONV5_KERNEL_NUM, CONV6_KERNEL_NUM], regularizer) | ||
| conv6_b = get_bias([CONV6_KERNEL_NUM]) | ||
| conv6 = conv2d(relu5, conv6_w) | ||
| conv6_op = tf.nn.bias_add(conv6, conv6_b) | ||
|
|
||
| # 第6个激活层 | ||
| relu6 = tf.nn.relu(conv6_op) | ||
|
|
||
| # 第7个卷积层 | ||
| conv7_w = get_weight([CONV_SIZE, CONV_SIZE, CONV6_KERNEL_NUM, CONV7_KERNEL_NUM], regularizer) | ||
| conv7_b = get_bias([CONV7_KERNEL_NUM]) | ||
| conv7 = conv2d(relu6, conv7_w) | ||
| conv7_op = tf.nn.bias_add(conv7, conv7_b) | ||
|
|
||
| # 第3个add层 | ||
| add3 = tf.add(conv7_op, add2) | ||
|
|
||
| # 第7个激活层 | ||
| relu7 = tf.nn.relu(add3) | ||
|
|
||
| # 第8个卷积层 | ||
| conv8_w = get_weight([CONV_SIZE, CONV_SIZE, CONV7_KERNEL_NUM, CONV8_KERNEL_NUM], regularizer) | ||
| conv8_b = get_bias([CONV8_KERNEL_NUM]) | ||
| conv8 = conv2d(relu7, conv8_w) | ||
| conv8_op = tf.nn.bias_add(conv8, conv8_b) | ||
|
|
||
| # 第8个激活层 | ||
| relu8 = tf.nn.relu(conv8_op) | ||
|
|
||
| # 第9个卷积层 | ||
| conv9_w = get_weight([CONV_SIZE, CONV_SIZE, CONV8_KERNEL_NUM, CONV9_KERNEL_NUM], regularizer) | ||
| conv9_b = get_bias([CONV9_KERNEL_NUM]) | ||
| conv9 = conv2d(relu8, conv9_w) | ||
| conv9_op = tf.nn.bias_add(conv9, conv9_b) | ||
|
|
||
| # 第4个add层 | ||
| add4 = tf.add(conv9_op, add3) | ||
|
|
||
| # 第9个激活层 | ||
| relu9 = tf.nn.relu(add4) | ||
|
|
||
| # 第10个卷积层 | ||
| conv10_w = get_weight([CONV_SIZE, CONV_SIZE, CONV9_KERNEL_NUM, CONV10_KERNEL_NUM], regularizer) | ||
| conv10_b = get_bias([CONV10_KERNEL_NUM]) | ||
| conv10 = conv2d(relu9, conv10_w) | ||
| conv10_op = tf.nn.bias_add(conv10, conv10_b) | ||
|
|
||
| # 第10个激活层 | ||
| relu10 = tf.nn.relu(conv10_op) | ||
|
|
||
| # 第11个卷积层 | ||
| conv11_w = get_weight([CONV_SIZE, CONV_SIZE, CONV10_KERNEL_NUM, CONV11_KERNEL_NUM], regularizer) | ||
| conv11_b = get_bias([CONV11_KERNEL_NUM]) | ||
| conv11 = conv2d(relu10, conv11_w) | ||
| conv11_op = tf.nn.bias_add(conv11, conv11_b) | ||
|
|
||
| # 第5个add层 | ||
| add5 = tf.add(conv11_op, add4) | ||
|
|
||
| # 第11个激活层 | ||
| relu11 = tf.nn.relu(add5) | ||
|
|
||
| # 第12个卷积层 | ||
| conv12_w = get_weight([CONV_SIZE, CONV_SIZE, CONV11_KERNEL_NUM, CONV12_KERNEL_NUM], regularizer) | ||
| conv12_b = get_bias([CONV12_KERNEL_NUM]) | ||
| conv12 = conv2d(relu11, conv12_w) | ||
| conv12_op = tf.nn.bias_add(conv12, conv12_b) | ||
|
|
||
| # 输出结果 | ||
| return conv12_op | ||
| ''' | ||
| 该模块为前向传播网路，定义网络结构 | ||
| ''' | ||
| import tensorflow as tf | ||
|
|
||
| CONV_SIZE = 5#卷积核大小 | ||
|
|
||
| # 各层通道数 | ||
| CONV1_KERNEL_NUM = 64 | ||
| CONV2_KERNEL_NUM = 64 | ||
| CONV3_KERNEL_NUM = 64 | ||
| CONV4_KERNEL_NUM = 64 | ||
| CONV5_KERNEL_NUM = 64 | ||
| CONV6_KERNEL_NUM = 64 | ||
| CONV7_KERNEL_NUM = 64 | ||
| CONV8_KERNEL_NUM = 64 | ||
| CONV9_KERNEL_NUM = 64 | ||
| CONV10_KERNEL_NUM = 64 | ||
| CONV11_KERNEL_NUM = 64 | ||
| CONV12_KERNEL_NUM = 1 | ||
|
|
||
| # 生成卷积核，权重w | ||
| def get_weight(shape, regularizer=None): | ||
| '''生成卷积核，卷积核形状shape,权重w由正态分布产生''' | ||
| w = tf.Variable(tf.truncated_normal(shape, stddev=0.01)) | ||
| if regularizer != None: | ||
| tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(regularizer)(w)) | ||
| return w | ||
|
|
||
| # 生成偏置项 | ||
| def get_bias(shape): | ||
| '''初始化0偏置项b''' | ||
| b = tf.Variable(tf.zeros(shape)) | ||
| return b | ||
|
|
||
| def conv2d(x,w): | ||
| ''' | ||
| 定义卷积操作 | ||
| x:输入图片 | ||
| w:卷积核 | ||
| strides:移动步长，左右上下移动步长为1 | ||
| padding:0填充操作，为了卷积操作后图片维度与原来相等 | ||
| ''' | ||
| return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME') | ||
|
|
||
|
|
||
| # 定义前向传播网络，按照文章的ResNet | ||
| def forward(x, channel, regularizer=None): | ||
|
|
||
| # 第1个卷积层 | ||
| conv1_w = get_weight([CONV_SIZE, CONV_SIZE, channel, CONV1_KERNEL_NUM], regularizer) | ||
| conv1_b = get_bias([CONV1_KERNEL_NUM]) | ||
| conv1 = conv2d(x, conv1_w) | ||
| conv1_op = tf.nn.bias_add(conv1, conv1_b) | ||
|
|
||
| # 第1个激活 | ||
| relu1 = tf.nn.relu(conv1_op) | ||
|
|
||
| # 第2个卷积层 | ||
| conv2_w = get_weight([CONV_SIZE, CONV_SIZE, CONV1_KERNEL_NUM, CONV2_KERNEL_NUM], regularizer) | ||
| conv2_b = get_bias([CONV2_KERNEL_NUM]) | ||
| conv2 = conv2d(relu1, conv2_w) | ||
| conv2_op = tf.nn.bias_add(conv2, conv2_b) | ||
|
|
||
| # 第2个激活 | ||
| relu2 = tf.nn.relu(conv2_op) | ||
|
|
||
| # 第3个卷积层 | ||
| conv3_w = get_weight([CONV_SIZE, CONV_SIZE, CONV2_KERNEL_NUM, CONV3_KERNEL_NUM], regularizer) | ||
| conv3_b = get_bias([CONV3_KERNEL_NUM]) | ||
| conv3 = conv2d(relu2, conv3_w) | ||
| conv3_op = tf.nn.bias_add(conv3,conv3_b) | ||
|
|
||
| # 第1个add层 | ||
| add1 = tf.add(conv3_op, conv1_op) | ||
|
|
||
| # 第3个激活层 | ||
| relu3 = tf.nn.relu(add1) | ||
|
|
||
| # 第4个卷积层 | ||
| conv4_w = get_weight([CONV_SIZE, CONV_SIZE, CONV3_KERNEL_NUM, CONV4_KERNEL_NUM], regularizer) | ||
| conv4_b = get_bias([CONV4_KERNEL_NUM]) | ||
| conv4 = conv2d(relu3, conv4_w) | ||
| conv4_op = tf.nn.bias_add(conv4, conv4_b) | ||
|
|
||
| # 第4个激活层 | ||
| relu4 = tf.nn.relu(conv4_op) | ||
|
|
||
| # 第5个卷积层 | ||
| conv5_w = get_weight([CONV_SIZE, CONV_SIZE, CONV4_KERNEL_NUM, CONV5_KERNEL_NUM], regularizer) | ||
| conv5_b = get_bias([CONV5_KERNEL_NUM]) | ||
| conv5 = conv2d(relu4, conv5_w) | ||
| conv5_op = tf.nn.bias_add(conv5, conv5_b) | ||
|
|
||
| # 第2个add层 | ||
| add2 = tf.add(conv5_op, add1) | ||
|
|
||
| # 第5个激活层 | ||
| relu5 = tf.nn.relu(add2) | ||
|
|
||
| # 第6个卷积层 | ||
| conv6_w = get_weight([CONV_SIZE, CONV_SIZE, CONV5_KERNEL_NUM, CONV6_KERNEL_NUM], regularizer) | ||
| conv6_b = get_bias([CONV6_KERNEL_NUM]) | ||
| conv6 = conv2d(relu5, conv6_w) | ||
| conv6_op = tf.nn.bias_add(conv6, conv6_b) | ||
|
|
||
| # 第6个激活层 | ||
| relu6 = tf.nn.relu(conv6_op) | ||
|
|
||
| # 第7个卷积层 | ||
| conv7_w = get_weight([CONV_SIZE, CONV_SIZE, CONV6_KERNEL_NUM, CONV7_KERNEL_NUM], regularizer) | ||
| conv7_b = get_bias([CONV7_KERNEL_NUM]) | ||
| conv7 = conv2d(relu6, conv7_w) | ||
| conv7_op = tf.nn.bias_add(conv7, conv7_b) | ||
|
|
||
| # 第3个add层 | ||
| add3 = tf.add(conv7_op, add2) | ||
|
|
||
| # 第7个激活层 | ||
| relu7 = tf.nn.relu(add3) | ||
|
|
||
| # 第8个卷积层 | ||
| conv8_w = get_weight([CONV_SIZE, CONV_SIZE, CONV7_KERNEL_NUM, CONV8_KERNEL_NUM], regularizer) | ||
| conv8_b = get_bias([CONV8_KERNEL_NUM]) | ||
| conv8 = conv2d(relu7, conv8_w) | ||
| conv8_op = tf.nn.bias_add(conv8, conv8_b) | ||
|
|
||
| # 第8个激活层 | ||
| relu8 = tf.nn.relu(conv8_op) | ||
|
|
||
| # 第9个卷积层 | ||
| conv9_w = get_weight([CONV_SIZE, CONV_SIZE, CONV8_KERNEL_NUM, CONV9_KERNEL_NUM], regularizer) | ||
| conv9_b = get_bias([CONV9_KERNEL_NUM]) | ||
| conv9 = conv2d(relu8, conv9_w) | ||
| conv9_op = tf.nn.bias_add(conv9, conv9_b) | ||
|
|
||
| # 第4个add层 | ||
| add4 = tf.add(conv9_op, add3) | ||
|
|
||
| # 第9个激活层 | ||
| relu9 = tf.nn.relu(add4) | ||
|
|
||
| # 第10个卷积层 | ||
| conv10_w = get_weight([CONV_SIZE, CONV_SIZE, CONV9_KERNEL_NUM, CONV10_KERNEL_NUM], regularizer) | ||
| conv10_b = get_bias([CONV10_KERNEL_NUM]) | ||
| conv10 = conv2d(relu9, conv10_w) | ||
| conv10_op = tf.nn.bias_add(conv10, conv10_b) | ||
|
|
||
| # 第10个激活层 | ||
| relu10 = tf.nn.relu(conv10_op) | ||
|
|
||
| # 第11个卷积层 | ||
| conv11_w = get_weight([CONV_SIZE, CONV_SIZE, CONV10_KERNEL_NUM, CONV11_KERNEL_NUM], regularizer) | ||
| conv11_b = get_bias([CONV11_KERNEL_NUM]) | ||
| conv11 = conv2d(relu10, conv11_w) | ||
| conv11_op = tf.nn.bias_add(conv11, conv11_b) | ||
|
|
||
| # 第5个add层 | ||
| add5 = tf.add(conv11_op, add4) | ||
|
|
||
| # 第11个激活层 | ||
| relu11 = tf.nn.relu(add5) | ||
|
|
||
| # 第12个卷积层 | ||
| conv12_w = get_weight([CONV_SIZE, CONV_SIZE, CONV11_KERNEL_NUM, CONV12_KERNEL_NUM], regularizer) | ||
| conv12_b = get_bias([CONV12_KERNEL_NUM]) | ||
| conv12 = conv2d(relu11, conv12_w) | ||
| conv12_op = tf.nn.bias_add(conv12, conv12_b) | ||
|
|
||
| # 输出结果 | ||
| return conv12_op |
Oops, something went wrong.