【【深度学习基础】Gated CNN 门控CNN 代码篇】转载自 https://blog.csdn.net/liuchonge/article/details/70254660
import numpy as np
import tensorflow as tfclass GatedCNN(object):def __init__(self, conf):
tf.reset_default_graph()
#定义输入X和Y的占位符
self.X = tf.placeholder(shape=[conf.batch_size, conf.context_size-1], dtype=tf.int32, name="X")
self.y = tf.placeholder(shape=[conf.batch_size, conf.context_size-1], dtype=tf.int32, name="y")#词嵌入层，将单词索引转化为词向量，shape=[conf.batch_size, conf.context_size-1, embed_size, 1]
embed = self.create_embeddings(self.X, conf)
h, res_input = embed, embed#堆叠num_layers个Gated CNN层
for i in range(conf.num_layers):
fanin_depth = h.get_shape()[-1]
# 最后一层卷积，filter数量为1
filter_size = conf.filter_size if i < conf.num_layers-1 else 1
shape = (conf.filter_h, conf.filter_w, fanin_depth, filter_size)with tf.variable_scope("layer_%d"%i):
#计算两个卷积w，v
conv_w = self.conv_op(h, shape, "linear")
conv_v = self.conv_op(h, shape, "gated")
#计算门限输出h, h[bz, max_q, emb, filter_size]
h = conv_w * tf.sigmoid(conv_v)
#将每5层Gated CNN组合成一个block。
if i % conf.block_size == 0:
h += res_input
res_input = h
#将模型输出h和y reshape
h = tf.reshape(h, (-1, conf.embedding_size))
y_shape = self.y.get_shape().as_list()
self.y = tf.reshape(self.y, (y_shape[0] * y_shape[1], 1))softmax_w = tf.get_variable("softmax_w", [conf.vocab_size, conf.embedding_size], tf.float32,
tf.random_normal_initializer(0.0, 0.1))
softmax_b = tf.get_variable("softmax_b", [conf.vocab_size], tf.float32, tf.constant_initializer(1.0))#Preferance: NCE Loss, heirarchial softmax, adaptive softmax
self.loss = tf.reduce_mean(tf.nn.nce_loss(softmax_w, softmax_b, h, self.y, conf.num_sampled, conf.vocab_size))
#训练模型，使用MomentumOptimizer优化器
trainer = tf.train.MomentumOptimizer(conf.learning_rate, conf.momentum)
gradients = trainer.compute_gradients(self.loss)
#将梯度进行clip截断
clipped_gradients = [(tf.clip_by_value(_[0], -conf.grad_clip, conf.grad_clip), _[1]) for _ in gradients]
self.optimizer = trainer.apply_gradients(clipped_gradients)
self.perplexity = tf.exp(self.loss)#将loss和perplexity进行记录，方便在tensorboard中观察模型训练效果
self.create_summaries()def create_embeddings(self, X, conf):
#这里使用的是随机初始化词向量
embeddings = tf.get_variable("embeds",(conf.vocab_size, conf.embedding_size), tf.float32, tf.random_uniform_initializer(-1.0,1.0))
embed = tf.nn.embedding_lookup(embeddings, X)
mask_layer = np.ones((conf.batch_size, conf.context_size-1, conf.embedding_size))
mask_layer[:,0:conf.filter_h/2,:] = 0
embed *= mask_layerembed_shape = embed.get_shape().as_list()
embed = tf.reshape(embed, (embed_shape[0], embed_shape[1], embed_shape[2], 1))
return embeddef conv_op(self, fan_in, shape, name):
W = tf.get_variable("%s_W"%name, shape, tf.float32, tf.random_normal_initializer(0.0, 0.1))
b = tf.get_variable("%s_b"%name, shape[-1], tf.float32, tf.constant_initializer(1.0))
#使用‘SAME’模式进行卷积
return tf.add(tf.nn.conv2d(fan_in, W, strides=[1,1,1,1], padding='SAME'), b)def create_summaries(self):
tf.summary.scalar("loss", self.loss)
tf.summary.scalar("perplexity", self.perplexity)
self.merged_summary_op = tf.summary.merge_all()
推荐阅读
- paddle|动手从头实现LSTM
- 人工智能|干货！人体姿态估计与运动预测
- 推荐系统论文进阶|CTR预估 论文精读(十一)--Deep Interest Evolution Network(DIEN)
- Python专栏|数据分析的常规流程
- 读书笔记|《白话大数据和机器学习》学习笔记1
- Pytorch学习|sklearn-SVM 模型保存、交叉验证与网格搜索
- Python机器学习基础与进阶|Python机器学习--集成学习算法--XGBoost算法
- 深度学习|深度学习笔记总结
- 机器学习|机器学习Sklearn学习总结
- 机器学习|线性回归原理与python实现