How to make current data loader more efficient

Hi, I have a working data loader, however while I am training it on the gpu cluster the average time taken to run per epoch is too much (approx. 25 mins). I am using PNG images for training with an average dimension of the image around (2500x5000). I want to make the training more faster. Please suggest what more can i do in order to do the training fast. Attached is the code of the data loader.

class MammographyCBISDDSM(Dataset):

def __init__(self, excel_file, category, transform=None):
        excel_file (string): Path to the excel file with annotations.
        category (string) : 'Classification' for Benign and Malignant. 'Subtypes' for Subtype Classification
        transform (callable, optional): Optional transform to be applied
    self.mammography = pd.read_csv(excel_file,sep = ';')
    self.category = category
    self.transform = transform

def __len__(self):
    return len(self.mammography)

def class_name_to_labels(self, idx):
    if self.category == 'Classification':
        class_name = self.mammography.iloc[idx, 9]
        if class_name == 'MALIGNANT':
            labels = 0.0
        elif class_name == 'BENIGN':
            labels = 1.0
        elif class_name == 'BENIGN_WITHOUT_CALLBACK':
            labels = 2.0

    return labels

def __getitem__(self, idx):
    if torch.is_tensor(idx):
        idx = idx.tolist()

    img_folder = self.mammography.iloc[idx, 11]
    image_array = image_array.pixel_array
    image_array = image_array * 1.0 / image_array.max()
    image_array = torch.from_numpy(image_array.astype(np.float32))
    image_array = image_array.repeat(3, 1, 1)

    if self.transform:
        image_array = self.transform(image_array)

    labels = self.class_name_to_labels(idx)
    labels = torch.from_numpy(np.array(labels))
    return image_array, labels

This great post explains common issues causing a data loading bottleneck and has some suggestions to speed it up.

Thanks for the link, I went through the suggestions and it made sense. As I am new to python and pytorch environment, it would be great if you could tell me more precise reason in this case and what I can do to overcome the issue! Thanks once again!

Probably biggest bottleneck here is reading image from disk (you can use timeit to track if it is true).
First, I would check if I have enough RAM to pre-load all the images into dataset. If this is not an option, I’d move all data to NVMe/SSD drive. Also I would convert and save PNGs to numpy arrays anyway, even on HDD, since you don’t have to invoke cv2 file parser every time image is loaded.
Last thing is, 2500x5000 is really big size. Is there any particular reason to use images as is without scaling down?