CRAFT-pytorch
CRAFT-pytorch copied to clipboard
clovaai
pytorch vs tensorrt
hi, i tried to get max performance and fps but at the end i found that pytorch on small fixed size is better than tensorrt time for example i had images in sizes about: 150 * 40 to 400 * 200 if i use tensorrt i get 75 * 20 images as result that text boxes are shade and near together but when i used pytorch in fixed size 150*40 i have minimum decrease accuracy with maximum speed, hope this help you , if i am wrong and there is a better way please feel free for any comment or message. thank you.
Hello， I have the same problem! I use Tesla V100(32G) GPU to inference, and fix input size 312801280. There is no speedup for tensorrt FP32 and FP16. When I use tensorrt int8, the precision、recall is 0. I am so confused.
Besides, for torch model, I found that batch inference does not speedup too. Please help!
I have another result more likely to yours(perhaps) with Tensorrt, when I gave big images (2000 * 1000) to craft, i could get best performance ( 1/4 exe time) but when my image was small as 200* 400 , then I didn't observe any speedup
but you said you in torch << batch inference does not speedup too>> , what do you mean?
Hi @sabrabano0 , for the original torch model, the fps of batch1 and batch16 is same
@sabrabano0 Could you tell me how to convert onnx model file to tensorRT model file ? I have yield a onnx model file but it occurs a mistake when yield a tensorRT model file . Looking forward to hearing from you. Thank you very much.
How do we convert this craft torch model (craft_mlt_25k.pth) to TensorRT format or ONNX format?
Hello， I have the same problem! I use Tesla V100(32G) GPU to inference, and fix input size 3_1280_1280. There is no speedup for tensorrt FP32 and FP16. When I use tensorrt int8, the precision、recall is 0. I am so confused.
Besides, for torch model, I found that batch inference does not speedup too. Please help!
@SabraHashemi Please can you refer any repository/blog to convert CROFT into tensorrt from pytorch ?
How do we convert this craft torch model (craft_mlt_25k.pth) to TensorRT format or ONNX format?
HI @kishcs , did you found any solution for it ? if yes then please guide me.