Convert images to torch tensors if the model is running on a GPU?

[yinleon]

Apologies if this is obvious.

I am using the ocr_predictor model on a GPU

import torch
from doctr.io import DocumentFile
from doctr.models import ocr_predictor

model = ocr_predictor(
    reco_arch="crnn_mobilenet_v3_small",
    det_arch="linknet_resnet34",
    pretrained=True,
)

if torch.cuda.is_available():
    device = torch.device("cuda:0")
    model = model.to(device).half()

When I read the image(s), they appear to be numpy arrays. Do I need to convert the image into a torch tensor on the same machine?

Separately the images are 720, 1280 which I fear may be too large to hold in the GPU's memory.

Thank you!

[felixdittrich92]

Hi @yinleon 👋,

What's your GPU ?
Code is correct 👍

It shouldn't raise any memory issues - I was running it with 4k images without issues on a 16GB GPU because under the hood every image will be resized to 1024x1024 same shape the detection models was trained on.

Best regards,
Felix