model.py

import time

import numpy as np
import tensorflow as tf

import config

class ChatBotModel:
    def __init__(self, forward_only, batch_size):
        """forward_only: if set, we do not construct the backward pass in the model.
        """
        print('Initialize new model')
        self.fw_only = forward_only
        self.batch_size = batch_size

    def _create_placeholders(self):
        # Feeds for inputs. It's a list of placeholders
        print('Create placeholders')
        self.encoder_inputs = [tf.placeholder(tf.int32, shape=[None], name='encoder{}'.format(i))
                               for i in range(config.BUCKETS[-1][0])]
        self.decoder_inputs = [tf.placeholder(tf.int32, shape=[None], name='decoder{}'.format(i))
                               for i in range(config.BUCKETS[-1][1] + 1)]
        self.decoder_masks = [tf.placeholder(tf.float32, shape=[None], name='mask{}'.format(i))
                              for i in range(config.BUCKETS[-1][1] + 1)]

        # Our targets are decoder inputs shifted by one (to ignore <GO> symbol)
        self.targets = self.decoder_inputs[1:]

    def _inference(self):
        print('Create inference')
        # If we use sampled softmax, we need an output projection.
        # Sampled softmax only makes sense if we sample less than vocabulary size.
        if config.NUM_SAMPLES > 0 and config.NUM_SAMPLES < config.DEC_VOCAB:
            w = tf.get_variable('proj_w', [config.HIDDEN_SIZE, config.DEC_VOCAB])
            b = tf.get_variable('proj_b', [config.DEC_VOCAB])
            self.output_projection = (w, b)

        def sampled_loss(logits, labels):
            labels = tf.reshape(labels, [-1, 1])
            return tf.nn.sampled_softmax_loss(weights=tf.transpose(w), 
                                              biases=b, 
                                              inputs=logits, 
                                              labels=labels, 
                                              num_sampled=config.NUM_SAMPLES, 
                                              num_classes=config.DEC_VOCAB)
        self.softmax_loss_function = sampled_loss

        single_cell = tf.contrib.rnn.GRUCell(config.HIDDEN_SIZE)
        self.cell = tf.contrib.rnn.MultiRNNCell([single_cell for _ in range(config.NUM_LAYERS)])

    def _create_loss(self):
        print('Creating loss... \nIt might take a couple of minutes depending on how many buckets you have.')
        start = time.time()
        def _seq2seq_f(encoder_inputs, decoder_inputs, do_decode):
            setattr(tf.contrib.rnn.GRUCell, '__deepcopy__', lambda self, _: self)
            setattr(tf.contrib.rnn.MultiRNNCell, '__deepcopy__', lambda self, _: self)
            return tf.contrib.legacy_seq2seq.embedding_attention_seq2seq(
                    encoder_inputs, decoder_inputs, self.cell,
                    num_encoder_symbols=config.ENC_VOCAB,
                    num_decoder_symbols=config.DEC_VOCAB,
                    embedding_size=config.HIDDEN_SIZE,
                    output_projection=self.output_projection,
                    feed_previous=do_decode)

        if self.fw_only:
            self.outputs, self.losses = tf.contrib.legacy_seq2seq.model_with_buckets(
                                        self.encoder_inputs, 
                                        self.decoder_inputs, 
                                        self.targets,
                                        self.decoder_masks, 
                                        config.BUCKETS, 
                                        lambda x, y: _seq2seq_f(x, y, True),
                                        softmax_loss_function=self.softmax_loss_function)
            # If we use output projection, we need to project outputs for decoding.
            if self.output_projection:
                for bucket in range(len(config.BUCKETS)):
                    self.outputs[bucket] = [tf.matmul(output, 
                                            self.output_projection[0]) + self.output_projection[1]
                                            for output in self.outputs[bucket]]
        else:
            self.outputs, self.losses = tf.contrib.legacy_seq2seq.model_with_buckets(
                                        self.encoder_inputs, 
                                        self.decoder_inputs, 
                                        self.targets,
                                        self.decoder_masks,
                                        config.BUCKETS,
                                        lambda x, y: _seq2seq_f(x, y, False),
                                        softmax_loss_function=self.softmax_loss_function)
        print('Time:', time.time() - start)

    def _creat_optimizer(self):
        print('Create optimizer... \nIt might take a couple of minutes depending on how many buckets you have.')
        with tf.variable_scope('training') as scope:
            self.global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')

            if not self.fw_only:
                self.optimizer = tf.train.GradientDescentOptimizer(config.LR)
                trainables = tf.trainable_variables()
                self.gradient_norms = []
                self.train_ops = []
                start = time.time()
                for bucket in range(len(config.BUCKETS)):
                    
                    clipped_grads, norm = tf.clip_by_global_norm(tf.gradients(self.losses[bucket], 
                                                                 trainables),
                                                                 config.MAX_GRAD_NORM)
                    self.gradient_norms.append(norm)
                    self.train_ops.append(self.optimizer.apply_gradients(zip(clipped_grads, trainables), 
                                                            global_step=self.global_step))
                    print('Creating opt for bucket {} took {} seconds'.format(bucket, time.time() - start))
                    start = time.time()


    def _create_summary(self):
        pass

    def build_graph(self):
        self._create_placeholders()
        self._inference()
        self._create_loss()
        self._creat_optimizer()
        self._create_summary()