def __init__(self, dataset='mnist.dataset', batch_size=16, fold='train', shuffle=True, last_batch=False, example_count=None, **kwargs): self.dsf = DatasetFile(dataset, example_count=example_count) self.fold = fold if self.fold == "train": self.transform = transforms.Compose( [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) # self.transform = transforms.Compose( # [transforms.RandomCrop(64, padding=4), # transforms.RandomHorizontalFlip(), # transforms.ToTensor(), # transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) else: self.transform = transforms.Compose( [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) self.img_conv = ImageConverter(dataset=self.dsf, transform=self.transform, **kwargs) self.lab_conv = LabelConverter(self.dsf, **kwargs) self.batch_size = batch_size self.last_batch = last_batch self.shuffle = shuffle self.num_classes = self.lab_conv.num_classes self.image_tensor = None self.label_tensor = None
class CustomDataloader(object): def __init__(self, dataset='mnist.dataset', batch_size=16, fold='train', shuffle=True, last_batch=False, example_count=None, img_format=None, **kwargs): self.dsf = DatasetFile(dataset, example_count=example_count) if img_format is None: img_format = ImageConverter self.img_conv = img_format(self.dsf, **kwargs) self.label_conv = QValueConverter(self.dsf, **kwargs) self.batch_size = batch_size self.fold = fold self.last_batch = last_batch self.shuffle = shuffle self.num_classes = self.label_conv.num_classes def get_batch(self, **kwargs): batch = self.dsf.get_batch(fold=self.fold, batch_size=self.batch_size, **kwargs) images, labels = self.convert(batch) return images, labels def __iter__(self): print('dataloader {} getting batches'.format(self)) self.batcher = self.dsf.get_all_batches(fold=self.fold, batch_size=self.batch_size, shuffle=self.shuffle, last_batch=self.last_batch) return self def __next__(self): batch = next(self.batcher) images, labels = self.convert(batch) return images, labels def convert(self, batch): images = self.img_conv(batch) labels = self.label_conv(batch) qvals, masks = labels[:, 0], labels[:, 1] images = torch.FloatTensor(images).cuda() masks = torch.FloatTensor(masks).cuda() qvals = torch.FloatTensor(qvals).cuda() return images, (qvals, masks) def __len__(self): return math.floor(self.dsf.count(self.fold) / self.batch_size) def count(self): return self.dsf.count(self.fold) def class_name(self, idx): return self.label_conv.labels[idx]
class CustomDataloader(object): def __init__(self, dataset='mnist.dataset', batch_size=16, fold='train', shuffle=True, last_batch=False, example_count=None, **kwargs): self.dsf = DatasetFile(dataset, example_count=example_count) self.img_conv = ImageConverter(self.dsf, **kwargs) self.lab_conv = LabelConverter(self.dsf, **kwargs) self.batch_size = batch_size self.fold = fold self.last_batch = last_batch self.shuffle = shuffle self.num_classes = self.lab_conv.num_classes self.image_tensor = None self.label_tensor = None self.batcher = None def get_batch(self, **kwargs): batch = self.dsf.get_batch(fold=self.fold, batch_size=self.batch_size, **kwargs) images, labels = self.convert(batch) return images, labels def __iter__(self): return self def __next__(self): if self.batcher is None: self.batcher = self.dsf.get_all_batches(fold=self.fold, batch_size=self.batch_size, shuffle=self.shuffle, last_batch=self.last_batch) batch = next(self.batcher) images, labels = self.convert(batch) return images, labels def convert(self, batch): images = self.img_conv(batch) labels = self.lab_conv(batch) images = torch.FloatTensor(images).cuda() labels = torch.LongTensor(labels).cuda() return images, labels def __len__(self): return math.floor(self.dsf.count(self.fold) / self.batch_size) def count(self): return self.dsf.count(self.fold) def class_name(self, idx): return self.lab_conv.labels[idx]
def __init__(self, dataset='mnist.dataset', batch_size=16, fold='train', shuffle=True, last_batch=False, example_count=None, **kwargs): self.dsf = DatasetFile(dataset, example_count=example_count) self.img_conv = ImageConverter(self.dsf, **kwargs) self.lab_conv = LabelConverter(self.dsf, **kwargs) self.batch_size = batch_size self.fold = fold self.last_batch = last_batch self.shuffle = shuffle self.num_classes = self.lab_conv.num_classes self.image_tensor = None self.label_tensor = None
class CustomDataloader(object): def __init__(self, dataset='mnist.dataset', batch_size=16, fold='train', shuffle=True, last_batch=False, example_count=None, img_format=None, **kwargs): self.dsf = DatasetFile(dataset, example_count=example_count) if img_format is None: img_format = ImageConverter self.img_conv = img_format(self.dsf, **kwargs) self.batch_size = batch_size self.fold = fold self.last_batch = last_batch self.shuffle = shuffle def get_batch(self, **kwargs): batch = self.dsf.get_batch(fold=self.fold, batch_size=self.batch_size, **kwargs) images, labels = self.convert(batch) return images, labels def __iter__(self): self.batcher = self.dsf.get_all_batches(fold=self.fold, batch_size=self.batch_size, shuffle=self.shuffle, last_batch=self.last_batch) return self def __next__(self): batch = next(self.batcher) images, labels = self.convert(batch) return images, labels def convert(self, batch): images = self.img_conv(batch) images = torch.FloatTensor(images).cuda() labels = None return images, labels def __len__(self): return math.floor(self.dsf.count(self.fold) / self.batch_size) def count(self): return self.dsf.count(self.fold)
def __init__(self, dataset='mnist.dataset', batch_size=16, fold='train', shuffle=True, last_batch=False, example_count=None, img_format=None, **kwargs): self.dsf = DatasetFile(dataset, example_count=example_count) if img_format is None: img_format = ImageConverter self.img_conv = img_format(self.dsf, **kwargs) self.batch_size = batch_size self.fold = fold self.last_batch = last_batch self.shuffle = shuffle
class CustomDataloader(object): def __init__(self, dataset='mnist.dataset', batch_size=16, fold='train', shuffle=True, last_batch=False, example_count=None, **kwargs): self.dsf = DatasetFile(dataset, example_count=example_count) self.img_conv = ImageConverter(self.dsf, **kwargs) self.lab_conv = LabelConverter(self.dsf, **kwargs) self.batch_size = batch_size self.fold = fold self.last_batch = last_batch self.shuffle = shuffle self.num_classes = self.lab_conv.num_classes self.image_tensor = None self.label_tensor = None def get_batch(self, **kwargs): batch = self.dsf.get_batch(fold=self.fold, batch_size=self.batch_size, **kwargs) images, labels = self.convert(batch) return images, labels def __iter__(self): batcher = self.dsf.get_all_batches( fold=self.fold, batch_size=self.batch_size, shuffle=self.shuffle, last_batch=self.last_batch) """ # TODO: Multithreading improves throughput by 10-20% # It must be implemented safely, however- not like this # In particular, ensure no deadlocks, interactivity and logging should still work import queue import threading q = queue.Queue(maxsize=1) def yield_batch_worker(): for batch in batcher: images, labels = self.convert(batch) q.put((images, labels)) q.put('finished') t = threading.Thread(target=yield_batch_worker) t.start() while True: result = q.get() if result == 'finished': break yield result t.join() """ for batch in batcher: images, labels = self.convert(batch) yield images, labels def convert(self, batch): images = self.img_conv(batch) labels = self.lab_conv(batch) # images = torch.FloatTensor(images).cuda() images = torch.stack(images, dim=0).cuda() labels = torch.LongTensor(labels).cuda() return images, labels def __len__(self): return math.floor(self.dsf.count(self.fold) / self.batch_size) def count(self): return self.dsf.count(self.fold) def class_name(self, idx): return self.lab_conv.labels[idx]