| 358 | |
| 359 | # generate dataset pipline |
| 360 | def build_model_input(filename, batch_size, num_epochs): |
| 361 | def parse_csv(value): |
| 362 | tf.logging.info('Parsing {}'.format(filename)) |
| 363 | cont_defaults = [[0.0] for i in range(1, 14)] |
| 364 | cate_defaults = [[' '] for i in range(1, 27)] |
| 365 | label_defaults = [[0]] |
| 366 | column_headers = TRAIN_DATA_COLUMNS |
| 367 | record_defaults = label_defaults + cont_defaults + cate_defaults |
| 368 | columns = tf.io.decode_csv(value, record_defaults=record_defaults) |
| 369 | all_columns = collections.OrderedDict(zip(column_headers, columns)) |
| 370 | labels = all_columns.pop(LABEL_COLUMN[0]) |
| 371 | features = all_columns |
| 372 | return features, labels |
| 373 | |
| 374 | def parse_parquet(value): |
| 375 | tf.logging.info('Parsing {}'.format(filename)) |
| 376 | labels = value.pop(LABEL_COLUMN[0]) |
| 377 | features = value |
| 378 | return features, labels |
| 379 | |
| 380 | '''Work Queue Feature''' |
| 381 | if args.workqueue and not args.tf: |
| 382 | from tensorflow.python.ops.work_queue import WorkQueue |
| 383 | work_queue = WorkQueue([filename], num_epochs=num_epochs) |
| 384 | # For multiple files： |
| 385 | # work_queue = WorkQueue([filename, filename1,filename2,filename3]) |
| 386 | files = work_queue.input_dataset() |
| 387 | else: |
| 388 | files = filename |
| 389 | # Extract lines from input files using the Dataset API. |
| 390 | if args.parquet_dataset and not args.tf: |
| 391 | from tensorflow.python.data.experimental.ops import parquet_dataset_ops |
| 392 | dataset = parquet_dataset_ops.ParquetDataset(files, batch_size=batch_size) |
| 393 | if args.parquet_dataset_shuffle: |
| 394 | dataset = dataset.shuffle(buffer_size=20000, |
| 395 | seed=args.seed) # fix seed for reproducing |
| 396 | if not args.workqueue: |
| 397 | dataset = dataset.repeat(num_epochs) |
| 398 | dataset = dataset.map(parse_parquet, num_parallel_calls=28) |
| 399 | else: |
| 400 | dataset = tf.data.TextLineDataset(files) |
| 401 | dataset = dataset.shuffle(buffer_size=20000, |
| 402 | seed=args.seed) # fix seed for reproducing |
| 403 | if not args.workqueue: |
| 404 | dataset = dataset.repeat(num_epochs) |
| 405 | dataset = dataset.batch(batch_size) |
| 406 | dataset = dataset.map(parse_csv, num_parallel_calls=28) |
| 407 | dataset = dataset.prefetch(2) |
| 408 | return dataset |
| 409 | |
| 410 | |
| 411 | # generate feature columns |