| 305 | |
| 306 | # generate dataset pipeline |
| 307 | def build_model_input(filename, batch_size, num_epochs): |
| 308 | def parse_csv(value): |
| 309 | tf.logging.info('Parsing {}'.format(filename)) |
| 310 | HASH_defaults = [[" "] for i in range(0, len(HASH_INPUTS))] |
| 311 | label_defaults = [[0] for i in range (0, len(LABEL_COLUMNS))] |
| 312 | column_headers = LABEL_COLUMNS + HASH_INPUTS |
| 313 | record_defaults = label_defaults + HASH_defaults |
| 314 | columns = tf.io.decode_csv(value, record_defaults=record_defaults) |
| 315 | all_columns = collections.OrderedDict(zip(column_headers, columns)) |
| 316 | labels = [] |
| 317 | for i in range(0, len(LABEL_COLUMNS)): |
| 318 | labels.append(all_columns.pop(LABEL_COLUMNS[i])) |
| 319 | label = tf.stack(labels, axis=1) |
| 320 | features = all_columns |
| 321 | return features, label |
| 322 | |
| 323 | def parse_parquet(value): |
| 324 | tf.logging.info('Parsing {}'.format(filename)) |
| 325 | labels = [] |
| 326 | for i in range(0, len(LABEL_COLUMNS)): |
| 327 | labels.append(value.pop(LABEL_COLUMNS[i])) |
| 328 | label = tf.stack(labels, axis=1) |
| 329 | features = value |
| 330 | return features, label |
| 331 | |
| 332 | '''Work Queue Feature''' |
| 333 | if args.workqueue and not args.tf: |
| 334 | from tensorflow.python.ops.work_queue import WorkQueue |
| 335 | work_queue = WorkQueue([filename], num_epochs=num_epochs) |
| 336 | # For multiple files： |
| 337 | # work_queue = WorkQueue([filename, filename1,filename2,filename3]) |
| 338 | files = work_queue.input_dataset() |
| 339 | else: |
| 340 | files = filename |
| 341 | # Extract lines from input files using the Dataset API. |
| 342 | if args.parquet_dataset and not args.tf: |
| 343 | from tensorflow.python.data.experimental.ops import parquet_dataset_ops |
| 344 | dataset = parquet_dataset_ops.ParquetDataset(files, batch_size=batch_size) |
| 345 | if args.parquet_dataset_shuffle: |
| 346 | dataset = dataset.shuffle(buffer_size=20000, |
| 347 | seed=args.seed) # fix seed for reproducing |
| 348 | if not args.workqueue: |
| 349 | dataset = dataset.repeat(num_epochs) |
| 350 | dataset = dataset.map(parse_parquet, num_parallel_calls=28) |
| 351 | else: |
| 352 | dataset = tf.data.TextLineDataset(files) |
| 353 | dataset = dataset.shuffle(buffer_size=20000, |
| 354 | seed=args.seed) # set seed for reproducing |
| 355 | if not args.workqueue: |
| 356 | dataset = dataset.repeat(num_epochs) |
| 357 | dataset = dataset.batch(batch_size) |
| 358 | dataset = dataset.map(parse_csv, num_parallel_calls=28) |
| 359 | |
| 360 | dataset = dataset.prefetch(2) |
| 361 | return dataset |
| 362 | |
| 363 | def build_feature_cols(): |
| 364 | feature_cols = [] |