_layers.py 文件源码

def __init__(
        self,
        inputs = None,
        train_labels = None,
        vocabulary_size = 80000,
        embedding_size = 200,
        num_sampled = 64,
        nce_loss_args = {},
        E_init = tf.random_uniform_initializer(minval=-1.0, maxval=1.0),
        E_init_args = {},
        nce_W_init = tf.truncated_normal_initializer(stddev=0.03),
        nce_W_init_args = {},
        nce_b_init = tf.constant_initializer(value=0.0),
        nce_b_init_args = {},
        name ='word2vec_layer',
    ):
        Layer.__init__(self, name=name)
        self.inputs = inputs
        self.n_units = embedding_size
        print("  tensorlayer:Instantiate Word2vecEmbeddingInputlayer %s: (%d, %d)" % (self.name, vocabulary_size, embedding_size))
        # Look up embeddings for inputs.
        # Note: a row of 'embeddings' is the vector representation of a word.
        # for the sake of speed, it is better to slice the embedding matrix
        # instead of transfering a word id to one-hot-format vector and then
        # multiply by the embedding matrix.
        # embed is the outputs of the hidden layer (embedding layer), it is a
        # row vector with 'embedding_size' values.
        with tf.variable_scope(name) as vs:
            embeddings = tf.get_variable(name='embeddings',
                                    shape=(vocabulary_size, embedding_size),
                                    initializer=E_init,
                                    **E_init_args)
            embed = tf.nn.embedding_lookup(embeddings, self.inputs)
            # Construct the variables for the NCE loss (i.e. negative sampling)
            nce_weights = tf.get_variable(name='nce_weights',
                                    shape=(vocabulary_size, embedding_size),
                                    initializer=nce_W_init,
                                    **nce_W_init_args)
            nce_biases = tf.get_variable(name='nce_biases',
                                    shape=(vocabulary_size),
                                    initializer=nce_b_init,
                                    **nce_b_init_args)

        # Compute the average NCE loss for the batch.
        # tf.nce_loss automatically draws a new sample of the negative labels
        # each time we evaluate the loss.
        self.nce_cost = tf.reduce_mean(
            tf.nn.nce_loss(weights=nce_weights, biases=nce_biases,
                           inputs=embed, labels=train_labels,
                           num_sampled=num_sampled, num_classes=vocabulary_size,
                           **nce_loss_args))

        self.outputs = embed
        self.normalized_embeddings = tf.nn.l2_normalize(embeddings, 1)

        self.all_layers = [self.outputs]
        self.all_params = [embeddings, nce_weights, nce_biases]
        self.all_drop = {}