(examples)
| 407 | label_norm = EncodeBitMap(n=math.ceil(math.log(dataset_cls.num_classes, 2)), ignore_label=dataset_cls.ignore_label) |
| 408 | |
| 409 | def preprocess_train(examples): |
| 410 | images = [image.convert("RGB") for image in examples[image_column]] |
| 411 | labels = [label for label in examples[label_column]] |
| 412 | pixel_values = [] |
| 413 | label_pixel_values = [] |
| 414 | for image, label in zip(images, labels): |
| 415 | # resize |
| 416 | image = img_resize(image) |
| 417 | label = lbl_resize(label) |
| 418 | # crop |
| 419 | image = crop(image) |
| 420 | label = crop(label) |
| 421 | # flip |
| 422 | if args.random_flip and random.random() < 0.5: |
| 423 | image = flip(image) |
| 424 | label = flip(label) |
| 425 | # to tensor |
| 426 | image = image2tensor(image) |
| 427 | label = label2tensor(label) |
| 428 | # norm |
| 429 | image = image_norm(image) |
| 430 | label = label_norm(label)[0] |
| 431 | label = 2. * label - 1. |
| 432 | |
| 433 | pixel_values.append(image) |
| 434 | label_pixel_values.append(label) |
| 435 | |
| 436 | examples["pixel_values"] = pixel_values |
| 437 | examples["label_pixel_values"] = label_pixel_values |
| 438 | return examples |
| 439 | |
| 440 | with accelerator.main_process_first(): |
| 441 | train_dataset = dataset["train"].with_transform(preprocess_train) |
nothing calls this directly
no outgoing calls
no test coverage detected