Compute the class logit predictions. Args: is_train: bool, are we in train mode embeddings: `embedding_utils.PretrainedWordEmbeddings` to embed with premise_tensors: batched premise tensors from `build_tensorize_text_fn` hypothesis_tensors: batched hypothesis tensors from `b
(is_train, embeddings, premise_tensors, hypothesis_tensors)
| 317 | |
| 318 | |
| 319 | def predict(is_train, embeddings, premise_tensors, hypothesis_tensors): |
| 320 | """Compute the class logit predictions. |
| 321 | |
| 322 | Args: |
| 323 | is_train: bool, are we in train mode |
| 324 | embeddings: `embedding_utils.PretrainedWordEmbeddings` to embed with |
| 325 | premise_tensors: batched premise tensors from `build_tensorize_text_fn` |
| 326 | hypothesis_tensors: batched hypothesis tensors from |
| 327 | `build_tensorize_text_fn` |
| 328 | |
| 329 | Returns: |
| 330 | logits: <tf.float32>[batch, 3] Scores for each class (for checkpoint |
| 331 | comparability, the BoolQ models also produce a neutral logit) |
| 332 | """ |
| 333 | with tf.variable_scope("embed"): |
| 334 | premise = embed_text(premise_tensors, embeddings) |
| 335 | premise_lens = premise_tensors["len"] |
| 336 | with tf.variable_scope("embed", reuse=True): |
| 337 | hypothesis = embed_text(hypothesis_tensors, embeddings) |
| 338 | hypothesis_len = hypothesis_tensors["len"] |
| 339 | |
| 340 | e_drop = FLAGS.embed_dropout_rate |
| 341 | premise = variational_dropout(premise, e_drop, is_train) |
| 342 | hypothesis = variational_dropout(hypothesis, e_drop, is_train) |
| 343 | |
| 344 | with tf.variable_scope("embed/encode-text/layer-1"): |
| 345 | premise = apply_highway_lstm(premise, premise_lens) |
| 346 | with tf.variable_scope("embed/encode-text/layer-1", reuse=True): |
| 347 | hypothesis = apply_highway_lstm(hypothesis, hypothesis_len) |
| 348 | |
| 349 | with tf.variable_scope("fuse"): |
| 350 | with tf.variable_scope("attention"): |
| 351 | atten = compute_attention(premise, hypothesis) |
| 352 | atten = mask_attention(atten, premise_lens, hypothesis_len) |
| 353 | |
| 354 | attended_h = tf.matmul(tf.nn.softmax(atten), hypothesis) |
| 355 | attended_p = tf.matmul( |
| 356 | tf.nn.softmax(tf.transpose(atten, [0, 2, 1])), premise) |
| 357 | premise = tf.concat([premise, attended_h, attended_h * premise], 2) |
| 358 | hypothesis = tf.concat([hypothesis, attended_p, attended_p * hypothesis], 2) |
| 359 | |
| 360 | with tf.variable_scope("post-process/layer-0"): |
| 361 | premise = apply_highway_lstm(premise, premise_lens) |
| 362 | with tf.variable_scope("post-process/layer-0", reuse=True): |
| 363 | hypothesis = apply_highway_lstm(hypothesis, hypothesis_len) |
| 364 | |
| 365 | drop = FLAGS.dropout_rate |
| 366 | premise = variational_dropout(premise, drop, is_train) |
| 367 | hypothesis = variational_dropout(hypothesis, drop, is_train) |
| 368 | |
| 369 | with tf.variable_scope("pool/atten-pool"): |
| 370 | premise = pool(premise, premise_lens) |
| 371 | with tf.variable_scope("pool/atten-pool", reuse=True): |
| 372 | hypothesis = pool(hypothesis, hypothesis_len) |
| 373 | |
| 374 | joint_embed = tf.concat([premise, hypothesis], 1) |
| 375 | with tf.variable_scope("post-processs-pooled/layer-1"): |
| 376 | joint_embed = tf.layers.dense( |
no test coverage detected
searching dependent graphs…