Model function. Atttention seq2seq model, augmented with an encoder over the targets of the nearest neighbors. Args: features: Dictionary of input Tensors. mode: train or eval. Keys from tf.estimator.ModeKeys. vocab: A list of strings of words in the vocabulary. hps: Hyperpar
(features, mode, vocab, hps)
| 67 | |
| 68 | |
| 69 | def encoder(features, mode, vocab, hps): |
| 70 | """Model function. |
| 71 | |
| 72 | Atttention seq2seq model, augmented with an encoder |
| 73 | over the targets of the nearest neighbors. |
| 74 | |
| 75 | Args: |
| 76 | features: Dictionary of input Tensors. |
| 77 | mode: train or eval. Keys from tf.estimator.ModeKeys. |
| 78 | vocab: A list of strings of words in the vocabulary. |
| 79 | hps: Hyperparams. |
| 80 | |
| 81 | Returns: |
| 82 | Encoder outputs. |
| 83 | """ |
| 84 | |
| 85 | # [batch_size, src_len] |
| 86 | src_inputs = features["src_inputs"] |
| 87 | src_len = features["src_len"] |
| 88 | |
| 89 | with tf.variable_scope("embeddings"): |
| 90 | scale = (3.0 / hps.emb_dim) ** 0.5 |
| 91 | embeddings = tf.get_variable( |
| 92 | "embeddings", |
| 93 | [vocab.size(), hps.emb_dim], |
| 94 | dtype=tf.float32, |
| 95 | initializer=tf.random_uniform_initializer(minval=-scale, maxval=scale)) |
| 96 | |
| 97 | # [batch_size, src_len, emb_dim] |
| 98 | src_input_emb = tf.nn.embedding_lookup(embeddings, src_inputs) |
| 99 | |
| 100 | if mode == tf_estimator.ModeKeys.TRAIN and hps.emb_drop > 0.: |
| 101 | src_input_emb = tf.nn.dropout( |
| 102 | src_input_emb, keep_prob=1.0-hps.emb_drop) |
| 103 | src_att_context, neighbor_att_context = None, None |
| 104 | src_copy_context, neighbor_copy_context = None, None |
| 105 | with tf.variable_scope("src_encoder"): |
| 106 | |
| 107 | # 2 * [batch_size, src_len, encoder_dim] |
| 108 | src_encoder_outputs, src_encoder_states = tf.nn.bidirectional_dynamic_rnn( |
| 109 | cell_fw=get_rnn_cell( |
| 110 | mode=mode, hps=hps, |
| 111 | input_dim=hps.emb_dim, |
| 112 | num_units=hps.encoder_dim, |
| 113 | num_layers=hps.num_encoder_layers, |
| 114 | dropout=hps.encoder_drop, |
| 115 | cell_type="lstm"), |
| 116 | cell_bw=get_rnn_cell( |
| 117 | mode=mode, hps=hps, |
| 118 | input_dim=hps.emb_dim, |
| 119 | num_units=hps.encoder_dim, |
| 120 | num_layers=hps.num_encoder_layers, |
| 121 | dropout=hps.encoder_drop, |
| 122 | cell_type="lstm"), |
| 123 | inputs=src_input_emb, |
| 124 | dtype=tf.float32, |
| 125 | sequence_length=src_len) |
| 126 |
no test coverage detected
searching dependent graphs…