Title: Query on EfficientSAM's Adaptability for Low-Resource Environments about efficientsam HOT 7 OPEN

@yihong1120, thanks for your interest in EfficientSAM!

Can you give some detailed information on the hardware platform for your use case, e.g, memory constraints, latency constraints, and CPU type?

from efficientsam.

When I download to model and export to A40 GPU, seems to use 35Gb GPU memory and sam without prompt, use over than 20s.
Maybe it's too slow to use.

from efficientsam.

@WanBenLe, can you kill all the jobs you are running to check?

from efficientsam.

I used it:
efficient_sam_vits_model = build_efficient_sam_vits()
efficient_sam_vits_model.to('cuda:3')

In def predict_masks, take_along_dim will lead to out of GPU memory.
iou_predictions = torch.take_along_dim(iou_predictions, sorted_ids, dim=2)

So I have to
output_masks=output_masks.to('cpu')
iou_predictions=iou_predictions.to('cpu')

I also use cuda to run in batched_nms process_small_region.
masks = torch.cat(new_masks, dim=0).to('cuda:3')
boxes = batched_mask_to_box(masks)
keep_by_nms = batched_nms(
boxes.float(),
torch.as_tensor(scores).to('cuda:3'),
torch.zeros_like(boxes[:, 0]).to('cuda:3'), # categories
iou_threshold=nms_thresh,
)

33gb GPU Mem of model.
41gb GPU Mem of model + batched_nms.

Maybe I had something wrong?

from efficientsam.

@WanBenLe What is the input for your case?

from efficientsam.

@yformer
Image, which resize 512*512

with zipfile.ZipFile("weights/efficient_sam_vits.pt.zip", 'r') as zip_ref:
zip_ref.extractall("weights")
efficient_sam_vits_model = build_efficient_sam_vits()
efficient_sam_vits_model.to('cuda:2')
efficient_sam_vits_model.device
rle = run_everything_ours(cv2.resize(image,(512,512)), efficient_sam_vits_model)

from efficientsam.

Me too, OOM occured using v100-32G.

from efficientsam.