A complete example for imagenet data loading

I've been trying to use your FFCV data loader for imagenet training. I find the provided example hard to follow as you use progressive resizing. I wonder if you could provide a complete example with the most commonly used resolution 224.

I have also coded it up myself, but I found the validation accuracy is significantly lower than the training accuracy in my case (see attached code snippet below). For example, after 3 epochs, the training ACC is around 40%, but the validation is only 15%.

def get_ffcv_trainloader(train_dataset, device, batch_size, num_workers=12, in_memory=True):
    train_path = Path(train_dataset)
    assert train_path.is_file()

    decoder = RandomResizedCropRGBImageDecoder((224, 224))
    image_pipeline: List[Operation] = [
        decoder,
        RandomHorizontalFlip(),
        ToTensor(),
        ToDevice(device, non_blocking=True),
        ToTorchImage(),
        NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float32)
    ]

    label_pipeline: List[Operation] = [
        IntDecoder(),
        ToTensor(),
        Squeeze(),
        ToDevice(device, non_blocking=True)
    ]

    order = OrderOption.QUASI_RANDOM
    loader = Loader(train_dataset,
                    batch_size=batch_size,
                    num_workers=num_workers,
                    order=order,
                    os_cache=in_memory,
                    drop_last=True,
                    pipelines={
                        'image': image_pipeline,
                        'label': label_pipeline
                    })

    return loader

def get_ffcv_valloader(val_dataset, device, batch_size, num_workers=12):
    val_path = Path(val_dataset)
    assert val_path.is_file()
    cropper = CenterCropRGBImageDecoder((224, 224), ratio=224/256)
    image_pipeline = [
        cropper,
        ToTensor(),
        ToDevice(device, non_blocking=True),
        ToTorchImage(),
        NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float32)
    ]

    label_pipeline = [
        IntDecoder(),
        ToTensor(),
        Squeeze(),
        ToDevice(device, non_blocking=True)
    ]

    loader = Loader(val_dataset,
                    batch_size=batch_size,
                    num_workers=num_workers,
                    order=OrderOption.SEQUENTIAL,
                    drop_last=False,
                    pipelines={
                        'image': image_pipeline,
                        'label': label_pipeline
                    })
    return loader

libffcv / ffcv-imagenet

A complete example for imagenet data loading #2