tf.contrib.seq2seq.sequence_loss example:seqence loss 实例代码

tf.contrib.seq2seq.sequence_loss example:seqence loss 实例代码

[code lang=text]
#!/usr/bin/env python
# -*- coding: utf-8 -*-

import tensorflow as tf

import numpy as np

params=np.random.normal(loc=0.0,scale=1.0,size=[10,10])

encoder_inputs=tf.placeholder(dtype=tf.int32,shape=[10,10])
decoder_inputs=tf.placeholder(dtype=tf.int32,shape=[10,10])

logits=tf.placeholder(dtype=tf.float32,shape=[10,10,10])
targets=tf.placeholder(dtype=tf.int32,shape=[10,10])
weights=tf.placeholder(dtype=tf.float32,shape=[10,10])

train_encoder_inputs=np.ones(shape=[10,10],dtype=np.int32)
train_decoder_inputs=np.ones(shape=[10,10],dtype=np.int32)
train_weights=np.ones(shape=[10,10],dtype=np.float32)

num_encoder_symbols=10
num_decoder_symbols=10
embedding_size=10
cell=tf.nn.rnn_cell.BasicLSTMCell(10)

def seq2seq(encoder_inputs,decoder_inputs,cell,num_encoder_symbols,num_decoder_symbols,embedding_size):
encoder_inputs = tf.unstack(encoder_inputs, axis=0)
decoder_inputs = tf.unstack(decoder_inputs, axis=0)
results,states=tf.contrib.legacy_seq2seq.embedding_rnn_seq2seq(
encoder_inputs,
decoder_inputs,
cell,
num_encoder_symbols,
num_decoder_symbols,
embedding_size,
output_projection=None,
feed_previous=False,
dtype=None,
scope=None
)
return results

def get_loss(logits,targets,weights):
loss=tf.contrib.seq2seq.sequence_loss(
logits,
targets=targets,
weights=weights
)
return loss

results=seq2seq(encoder_inputs,decoder_inputs,cell,num_encoder_symbols,num_decoder_symbols,embedding_size)
logits=tf.stack(results,axis=0)
print(logits)
loss=get_loss(logits,targets,weights)

with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
results_value=sess.run(results,feed_dict={encoder_inputs:train_encoder_inputs,decoder_inputs:train_decoder_inputs})
print(type(results_value[0]))
print(len(results_value))
cost = sess.run(loss, feed_dict={encoder_inputs: train_encoder_inputs, targets: train_decoder_inputs,
weights:train_weights,decoder_inputs:train_decoder_inputs})
print(cost)

[/code]

Be the first to comment

Leave a Reply

Your email address will not be published.