源鏈接：http://blog.csdn.net/u014595019/article/details/52759104

損失函數(shù)計(jì)算

# softmax_w , shape=[hidden_size, vocab_size], 用于將distributed表示的單詞轉(zhuǎn)化為one-hot表示 softmax_w = tf.get_variable("softmax_w", [size, vocab_size], dtype=data_type()) softmax_b = tf.get_variable("softmax_b", [vocab_size], dtype=data_type()) # [batch*numsteps, vocab_size] 從隱藏語義轉(zhuǎn)化成完全表示 logits = tf.matmul(output, softmax_w) + softmax_b# loss , shape=[batch*num_steps] # 帶權(quán)重的交叉熵計(jì)算 loss = tf.nn.seq2seq.sequence_loss_by_example([logits], # output [batch*numsteps, vocab_size][tf.reshape(self._targets, [-1])], # target, [batch_size, num_steps] 然后展開成一維【列表】[tf.ones([batch_size * num_steps], dtype=data_type())]) # weight self._cost = cost = tf.reduce_sum(loss) / batch_size # 計(jì)算得到平均每批batch的誤差 self._final_state = state self._lr = tf.Variable(0.0, trainable=False) # lr 指的是 learning_rate tvars = tf.trainable_variables() grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars),config.max_grad_norm) # 梯度下降優(yōu)化，指定學(xué)習(xí)速率 optimizer = tf.train.GradientDescentOptimizer(self._lr) # optimizer = tf.train.AdamOptimizer() # optimizer = tf.train.GradientDescentOptimizer(0.5) self._train_op = optimizer.apply_gradients(zip(grads, tvars)) # 將梯度應(yīng)用于變量 # self._train_op = optimizer.minimize(grads)

全部源代碼:

# Copyright 2015 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # =============================================================================="""Example / benchmark for building a PTB LSTM model. Trains the model described in: (Zaremba, et. al.) Recurrent Neural Network Regularization http://arxiv.org/abs/1409.2329 There are 3 supported model configurations: =========================================== | config | epochs | train | valid | test =========================================== | small | 13 | 37.99 | 121.39 | 115.91 | medium | 39 | 48.45 | 86.16 | 82.07 | large | 55 | 37.87 | 82.62 | 78.29 The exact results may vary depending on the random initialization. The hyperparameters used in the model: - init_scale - the initial scale of the weights - learning_rate - the initial value of the learning rate - max_grad_norm - the maximum permissible norm of the gradient - num_layers - the number of LSTM layers - num_steps - the number of unrolled steps of LSTM - hidden_size - the number of LSTM units - max_epoch - the number of epochs trained with the initial learning rate - max_max_epoch - the total number of epochs for training - keep_prob - the probability of keeping weights in the dropout layer - lr_decay - the decay of the learning rate for each epoch after "max_epoch" - batch_size - the batch size The data required for this example is in the data/ dir of the PTB dataset from Tomas Mikolov's webpage: $ wget http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz $ tar xvf simple-examples.tgz To run: $ python ptb_word_lm.py --data_path=simple-examples/data/ """ from __future__ import absolute_import from __future__ import division from __future__ import print_functionimport timeimport numpy as np import tensorflow as tffrom tensorflow.models.rnn.ptb import readerflags = tf.flags logging = tf.loggingflags.DEFINE_string("model", "small","A type of model. Possible options are: small, medium, large.") flags.DEFINE_string("data_path", '/home/multiangle/download/simple-examples/data/', "data_path") flags.DEFINE_bool("use_fp16", False,"Train using 16-bit floats instead of 32bit floats")FLAGS = flags.FLAGSdef data_type():return tf.float16 if FLAGS.use_fp16 else tf.float32class PTBModel(object):"""The PTB model."""def __init__(self, is_training, config):""":param is_training: 是否要進(jìn)行訓(xùn)練.如果is_training=False,則不會(huì)進(jìn)行參數(shù)的修正。"""self.batch_size = batch_size = config.batch_sizeself.num_steps = num_steps = config.num_stepssize = config.hidden_sizevocab_size = config.vocab_sizeself._input_data = tf.placeholder(tf.int32, [batch_size, num_steps]) # 輸入self._targets = tf.placeholder(tf.int32, [batch_size, num_steps]) # 預(yù)期輸出，兩者都是index序列，長度為num_step# Slightly better results can be obtained with forget gate biases# initialized to 1 but the hyperparameters of the model would need to be# different than reported in the paper.lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(size, forget_bias=0.0, state_is_tuple=True)if is_training and config.keep_prob < 1: # 在外面包裹一層dropoutlstm_cell = tf.nn.rnn_cell.DropoutWrapper(lstm_cell, output_keep_prob=config.keep_prob)cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cell] * config.num_layers, state_is_tuple=True) # 多層lstm cell 堆疊起來self._initial_state = cell.zero_state(batch_size, data_type()) # 參數(shù)初始化,rnn_cell.RNNCell.zero_statewith tf.device("/cpu:0"):embedding = tf.get_variable("embedding", [vocab_size, size], dtype=data_type()) # vocab size * hidden size, 將單詞轉(zhuǎn)成embedding描述# 將輸入seq用embedding表示, shape=[batch, steps, hidden_size]inputs = tf.nn.embedding_lookup(embedding, self._input_data)if is_training and config.keep_prob < 1:inputs = tf.nn.dropout(inputs, config.keep_prob)# Simplified version of tensorflow.models.rnn.rnn.py's rnn().# This builds an unrolled LSTM for tutorial purposes only.# In general, use the rnn() or state_saving_rnn() from rnn.py.## The alternative version of the code below is:## inputs = [tf.squeeze(input_, [1])# for input_ in tf.split(1, num_steps, inputs)]# outputs, state = tf.nn.rnn(cell, inputs, initial_state=self._initial_state)outputs = []state = self._initial_state # state 表示 各個(gè)batch中的狀態(tài)with tf.variable_scope("RNN"):for time_step in range(num_steps):if time_step > 0: tf.get_variable_scope().reuse_variables()# cell_out: [batch, hidden_size](cell_output, state) = cell(inputs[:, time_step, :], state)outputs.append(cell_output) # output: shape[num_steps][batch,hidden_size]# 把之前的list展開，成[batch, hidden_size*num_steps],然后 reshape, 成[batch*numsteps, hidden_size]output = tf.reshape(tf.concat(1, outputs), [-1, size])# softmax_w , shape=[hidden_size, vocab_size], 用于將distributed表示的單詞轉(zhuǎn)化為one-hot表示softmax_w = tf.get_variable("softmax_w", [size, vocab_size], dtype=data_type())softmax_b = tf.get_variable("softmax_b", [vocab_size], dtype=data_type())# [batch*numsteps, vocab_size] 從隱藏語義轉(zhuǎn)化成完全表示logits = tf.matmul(output, softmax_w) + softmax_b# loss , shape=[batch*num_steps]# 帶權(quán)重的交叉熵計(jì)算loss = tf.nn.seq2seq.sequence_loss_by_example([logits], # output [batch*numsteps, vocab_size][tf.reshape(self._targets, [-1])], # target, [batch_size, num_steps] 然后展開成一維【列表】[tf.ones([batch_size * num_steps], dtype=data_type())]) # weightself._cost = cost = tf.reduce_sum(loss) / batch_size # 計(jì)算得到平均每批batch的誤差self._final_state = stateif not is_training: # 如果沒有訓(xùn)練，則不需要更新state的值。returnself._lr = tf.Variable(0.0, trainable=False)tvars = tf.trainable_variables()# clip_by_global_norm: 梯度衰減，具體算法為t_list[i] * clip_norm / max(global_norm, clip_norm)# 這里gradients求導(dǎo)，ys和xs都是張量# 返回一個(gè)長為len(xs)的張量，其中的每個(gè)元素都是\grad{\frac{dy}{dx}}# clip_by_global_norm 用于控制梯度膨脹,前兩個(gè)參數(shù)t_list, global_norm, 則# t_list[i] * clip_norm / max(global_norm, clip_norm)# 其中 global_norm = sqrt(sum([l2norm(t)**2 for t in t_list]))grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars),config.max_grad_norm)# 梯度下降優(yōu)化，指定學(xué)習(xí)速率optimizer = tf.train.GradientDescentOptimizer(self._lr)# optimizer = tf.train.AdamOptimizer()# optimizer = tf.train.GradientDescentOptimizer(0.5)self._train_op = optimizer.apply_gradients(zip(grads, tvars)) # 將梯度應(yīng)用于變量self._new_lr = tf.placeholder(tf.float32, shape=[], name="new_learning_rate") # 用于外部向graph輸入新的 lr值self._lr_update = tf.assign(self._lr, self._new_lr) # 使用new_lr來更新lr的值def assign_lr(self, session, lr_value):# 使用 session 來調(diào)用 lr_update 操作session.run(self._lr_update, feed_dict={self._new_lr: lr_value})@propertydef input_data(self):return self._input_data@propertydef targets(self):return self._targets@propertydef initial_state(self):return self._initial_state@propertydef cost(self):return self._cost@propertydef final_state(self):return self._final_state@propertydef lr(self):return self._lr@propertydef train_op(self):return self._train_opclass SmallConfig(object):"""Small config."""init_scale = 0.1 #learning_rate = 1.0 # 學(xué)習(xí)速率max_grad_norm = 5 # 用于控制梯度膨脹，num_layers = 2 # lstm層數(shù)num_steps = 20 # 單個(gè)數(shù)據(jù)中，序列的長度。hidden_size = 200 # 隱藏層規(guī)模max_epoch = 4 # epoch<max_epoch時(shí)，lr_decay值=1,epoch>max_epoch時(shí),lr_decay逐漸減小max_max_epoch = 13 # 指的是整個(gè)文本循環(huán)13遍。keep_prob = 1.0lr_decay = 0.5 # 學(xué)習(xí)速率衰減batch_size = 20 # 每批數(shù)據(jù)的規(guī)模，每批有20個(gè)。vocab_size = 10000 # 詞典規(guī)模，總共10K個(gè)詞class MediumConfig(object):"""Medium config."""init_scale = 0.05learning_rate = 1.0max_grad_norm = 5num_layers = 2num_steps = 35hidden_size = 650max_epoch = 6max_max_epoch = 39keep_prob = 0.5lr_decay = 0.8batch_size = 20vocab_size = 10000class LargeConfig(object):"""Large config."""init_scale = 0.04learning_rate = 1.0max_grad_norm = 10num_layers = 2num_steps = 35hidden_size = 1500max_epoch = 14max_max_epoch = 55keep_prob = 0.35lr_decay = 1 / 1.15batch_size = 20vocab_size = 10000class TestConfig(object):"""Tiny config, for testing."""init_scale = 0.1learning_rate = 1.0max_grad_norm = 1num_layers = 1num_steps = 2hidden_size = 2max_epoch = 1max_max_epoch = 1keep_prob = 1.0lr_decay = 0.5batch_size = 20vocab_size = 10000def run_epoch(session, model, data, eval_op, verbose=False):"""Runs the model on the given data."""# epoch_size 表示批次總數(shù)。也就是說，需要向session喂這么多次數(shù)據(jù)epoch_size = ((len(data) // model.batch_size) - 1) // model.num_steps # // 表示整數(shù)除法start_time = time.time()costs = 0.0iters = 0state = session.run(model.initial_state)for step, (x, y) in enumerate(reader.ptb_iterator(data, model.batch_size,model.num_steps)):fetches = [model.cost, model.final_state, eval_op] # 要進(jìn)行的操作，注意訓(xùn)練時(shí)和其他時(shí)候eval_op的區(qū)別feed_dict = {} # 設(shè)定input和target的值feed_dict[model.input_data] = xfeed_dict[model.targets] = yfor i, (c, h) in enumerate(model.initial_state):feed_dict[c] = state[i].c # 這部分有什么用？看不懂feed_dict[h] = state[i].hcost, state, _ = session.run(fetches, feed_dict) # 運(yùn)行session,獲得cost和statecosts += cost # 將 cost 累積iters += model.num_stepsif verbose and step % (epoch_size // 10) == 10: # 也就是每個(gè)epoch要輸出10個(gè)perplexity值print("%.3f perplexity: %.3f speed: %.0f wps" %(step * 1.0 / epoch_size, np.exp(costs / iters),iters * model.batch_size / (time.time() - start_time)))return np.exp(costs / iters)def get_config():if FLAGS.model == "small":return SmallConfig()elif FLAGS.model == "medium":return MediumConfig()elif FLAGS.model == "large":return LargeConfig()elif FLAGS.model == "test":return TestConfig()else:raise ValueError("Invalid model: %s", FLAGS.model)# def main(_): if __name__=='__main__':if not FLAGS.data_path:raise ValueError("Must set --data_path to PTB data directory")print(FLAGS.data_path)raw_data = reader.ptb_raw_data(FLAGS.data_path) # 獲取原始數(shù)據(jù)train_data, valid_data, test_data, _ = raw_dataconfig = get_config()eval_config = get_config()eval_config.batch_size = 1eval_config.num_steps = 1with tf.Graph().as_default(), tf.Session() as session:initializer = tf.random_uniform_initializer(-config.init_scale, # 定義如何對(duì)參數(shù)變量初始化config.init_scale)with tf.variable_scope("model", reuse=None,initializer=initializer):m = PTBModel(is_training=True, config=config) # 訓(xùn)練模型， is_trainable=Truewith tf.variable_scope("model", reuse=True,initializer=initializer):mvalid = PTBModel(is_training=False, config=config) # 交叉檢驗(yàn)和測(cè)試模型，is_trainable=Falsemtest = PTBModel(is_training=False, config=eval_config)summary_writer = tf.train.SummaryWriter('/tmp/lstm_logs',session.graph)tf.initialize_all_variables().run() # 對(duì)參數(shù)變量初始化for i in range(config.max_max_epoch): # 所有文本要重復(fù)多次進(jìn)入模型訓(xùn)練# learning rate 衰減# 在 遍數(shù)小于max epoch時(shí)， lr_decay = 1 ; > max_epoch時(shí)， lr_decay = 0.5^(i-max_epoch)lr_decay = config.lr_decay ** max(i - config.max_epoch, 0.0)m.assign_lr(session, config.learning_rate * lr_decay) # 設(shè)置learning rateprint("Epoch: %d Learning rate: %.3f" % (i + 1, session.run(m.lr)))train_perplexity = run_epoch(session, m, train_data, m.train_op,verbose=True) # 訓(xùn)練困惑度print("Epoch: %d Train Perplexity: %.3f" % (i + 1, train_perplexity))valid_perplexity = run_epoch(session, mvalid, valid_data, tf.no_op()) # 檢驗(yàn)困惑度print("Epoch: %d Valid Perplexity: %.3f" % (i + 1, valid_perplexity))test_perplexity = run_epoch(session, mtest, test_data, tf.no_op()) # 測(cè)試?yán)Щ蠖?/span>print("Test Perplexity: %.3f" % test_perplexity)# if __name__ == "__main__": # tf.app.run()

總結(jié)

以上是生活随笔為你收集整理的gradient clipping的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。