一个古龙风格的闲聊机器人，实现主要基于tensorflow官网的seq2seq tutorials.zip

#coding=utf-8 import tensorflow as tf import logging class Seq2seq: def __init__(self, max_sequence_len, embedding_size, hidden_size, vocab_size, sent_sos_id, sent_eos_id, pretrained_embedding_mat=None): self.max_sequence_len = max_sequence_len self.embedding_size = embedding_size self.hidden_size = hidden_size self.vocab_size = vocab_size self.sent_sos_id = sent_sos_id self.sent_eos_id = sent_eos_id self.global_step = tf.Variable(0, name="global_step", trainable=False) self.encoder_inputs = tf.placeholder(tf.int32, [None, max_sequence_len], name="encoder_inputs") self.encoder_inputs_actual_lengths = tf.placeholder(tf.int32, [None], name="encoder_inputs_actual_lengths") self.decoder_outputs = tf.placeholder(tf.int32, [None, max_sequence_len], name="decoder_outputs") self.decoder_outputs_actual_lengths = tf.placeholder(tf.int32, [None], name="decoder_outputs_actual_lengths") self.batch_size = tf.placeholder(tf.int32,name="batch_size") #self.embedding_mat = tf.Variable(pretrained_embedding_mat, name="vocab_W") self.embedding_mat = tf.Variable(tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0), name="vocab_W") #batch_size = self.decoder_outputs.get_shape().as_list()[0] self.decoder_outputs_no_eos = tf.strided_slice(self.decoder_outputs, [0,0], [self.batch_size, -1], [1,1]) self.decoder_inputs = tf.concat([tf.fill([self.batch_size, 1], sent_sos_id), self.decoder_outputs_no_eos], 1) self.encoder_inputs_embeded = tf.nn.embedding_lookup(self.embedding_mat, self.encoder_inputs) self.decoder_inputs_embeded = tf.nn.embedding_lookup(self.embedding_mat, self.decoder_inputs) self.encoder_inputs_time_major = tf.transpose(self.encoder_inputs_embeded, perm=[1, 0, 2]) self.decoder_inputs_time_major = tf.transpose(self.decoder_inputs_embeded, perm=[1, 0, 2]) self.decoder_cell = tf.nn.rnn_cell.BasicLSTMCell(self.hidden_size) self.projection_layer = tf.layers.Dense(vocab_size, use_bias=False) logging.debug("encoder_inputs_time_major's shape: {}".format(self.encoder_inputs_time_major.get_shape().as_list())) encoder_cell = tf.nn.rnn_cell.BasicLSTMCell(self.hidden_size) self.encoder_outputs, self.encoder_state = tf.nn.dynamic_rnn( encoder_cell, self.encoder_inputs_time_major, sequence_length=self.encoder_inputs_actual_lengths, dtype=tf.float32, time_major=True) logging.debug("encoder_state's shape: {}".format(self.encoder_state.h.get_shape().as_list())) def decoding_layer_train(self): # TrainingHelper: A helper for use during training. Only reads inputs. # Returned sample_ids are the argmax of the RNN output logits. helper = tf.contrib.seq2seq.TrainingHelper( self.decoder_inputs_time_major, self.decoder_outputs_actual_lengths, time_major=True) decoder = tf.contrib.seq2seq.BasicDecoder( self.decoder_cell, helper, self.encoder_state, output_layer=self.projection_layer) outputs, final_context_state, _ = tf.contrib.seq2seq.dynamic_decode(decoder, output_time_major=True, swap_memory=True) sample_id = outputs.sample_id logits = outputs.rnn_output #logging.debug("decoder_inputs_onehot_timemajor's shape: {}".format(self.decoder_inputs_onehot_timemajor.get_shape().as_list())) decoder_max_steps, decoder_batch_size, decoder_dim = tf.unstack(tf.shape(outputs.rnn_output)) self.decoder_targets_true_length = tf.transpose(self.decoder_outputs, perm=[1, 0])[:decoder_max_steps] logging.debug("rnn_outpus's shape: {}".format(logits.get_shape().as_list())) #target_weights = tf.sequence_mask( # self.decoder_inputs_actual_lengths, max_sequence_len, dtype=logits.dtype) self.mask = tf.to_float(tf.not_equal(self.decoder_targets_true_length, 0)) self.loss = tf.contrib.seq2seq.sequence_loss( outputs.rnn_output, self.decoder_targets_true_length, weights=self.mask) optimizer = tf.train.AdamOptimizer(name="AdamOptimizer") self.grads, self.vars = zip(*optimizer.compute_gradients(self.loss)) self.clip_grads, _ = tf.clip_by_global_norm(self.grads, 5) self.train_op = optimizer.apply_gradients(zip(self.clip_grads, self.vars), global_step=self.global_step) return self.train_op, self.loss def decoding_layer_inference(self): helper = tf.contrib.seq2seq.GreedyEmbeddingHelper( self.embedding_mat, tf.fill([batch_size], self.sent_sos_id), self.sent_eos_id) decoder = tf.contrib.seq2seq.BasicDecoder( self.decoder_cell, helper, self.encoder_state, output_layer = self.projection_layer) outputs, _ = tf.contrib.seq2seq.dynamic_decode( decoder, maximum_iterations=self.max_sequence_len) inference_id = outputs.sample_id