training_1.py

import torch
from torchvision import transforms

# define some re-usable stuff
IMAGE_SIZE = 224
NUM_CLASSES = 1103
BATCH_SIZE = 32
device = torch.device("cuda:0")
IMG_MEAN = model_ft.mean
IMG_STD = model_ft.std

# make some augmentations on training data
train_transform = transforms.Compose([
    transforms.RandomResizedCrop((IMAGE_SIZE, IMAGE_SIZE)),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize(IMG_MEAN, IMG_STD)
])


# use the collections dataset class we created earlier
train_dataset = CollectionsDataset(csv_file='../input/folds.csv',
                                   root_dir='../input/train/',
                                   num_classes=NUM_CLASSES,
                                   transform=train_transform)

# create the pytorch data loader
train_dataset_loader = torch.utils.data.DataLoader(train_dataset,
                                                   batch_size=BATCH_SIZE,
                                                   shuffle=True,
                                                   num_workers=4)
# push model to device
model_ft = model_ft.to(device)