| 426 | |
| 427 | # generate feature columns |
| 428 | def build_feature_columns(): |
| 429 | user_column = [] |
| 430 | item_column = [] |
| 431 | tag_column = [] |
| 432 | key_column = [] |
| 433 | if args.group_embedding and not args.tf: |
| 434 | with tf.feature_column.group_embedding_column_scope(name="categorical"): |
| 435 | for column_name in INPUT_FEATURES: |
| 436 | if column_name in TAG_COLUMN: |
| 437 | # parse_sequence_feature |
| 438 | categorical_column = tf.feature_column.sequence_categorical_column_with_hash_bucket( |
| 439 | column_name, |
| 440 | hash_bucket_size=HASH_BUCKET_SIZES[column_name], |
| 441 | dtype=tf.string) |
| 442 | else: |
| 443 | # parse_id_feature |
| 444 | categorical_column = tf.feature_column.categorical_column_with_hash_bucket( |
| 445 | column_name, |
| 446 | hash_bucket_size=HASH_BUCKET_SIZES[column_name], |
| 447 | dtype=tf.string) |
| 448 | |
| 449 | if not args.tf: |
| 450 | '''Feature Elimination of EmbeddingVariable Feature''' |
| 451 | if args.ev_elimination == 'gstep': |
| 452 | # Feature elimination based on global steps |
| 453 | evict_opt = tf.GlobalStepEvict(steps_to_live=4000) |
| 454 | elif args.ev_elimination == 'l2': |
| 455 | # Feature elimination based on l2 weight |
| 456 | evict_opt = tf.L2WeightEvict(l2_weight_threshold=1.0) |
| 457 | else: |
| 458 | evict_opt = None |
| 459 | '''Feature Filter of EmbeddingVariable Feature''' |
| 460 | if args.ev_filter == 'cbf': |
| 461 | # CBF-based feature filter |
| 462 | filter_option = tf.CBFFilter( |
| 463 | filter_freq=3, |
| 464 | max_element_size=2**30, |
| 465 | false_positive_probability=0.01, |
| 466 | counter_type=tf.int64) |
| 467 | elif args.ev_filter == 'counter': |
| 468 | # Counter-based feature filter |
| 469 | filter_option = tf.CounterFilter(filter_freq=3) |
| 470 | else: |
| 471 | filter_option = None |
| 472 | ev_opt = tf.EmbeddingVariableOption( |
| 473 | evict_option=evict_opt, filter_option=filter_option) |
| 474 | |
| 475 | if args.ev: |
| 476 | '''Embedding Variable Feature''' |
| 477 | categorical_column = tf.feature_column.categorical_column_with_embedding( |
| 478 | column_name, dtype=tf.string, ev_option=ev_opt) |
| 479 | elif args.adaptive_emb: |
| 480 | ''' Adaptive Embedding Feature Part 2 of 2 |
| 481 | Expcet the follow code, a dict, 'adaptive_mask_tensors', is need as the input of |
| 482 | 'tf.feature_column.input_layer(adaptive_mask_tensors=adaptive_mask_tensors)'. |
| 483 | For column 'COL_NAME',the value of adaptive_mask_tensors['$COL_NAME'] is a int32 |
| 484 | tensor with shape [batch_size]. |
| 485 | ''' |