How to get the inference time (speed) ? about up-detr

[dddzg]

Hi @ilem777 , it depends on what you are applying UP-DETR for.
If you are applying UP-DETR for object detection, it has the exact same inference time as DETR (28fps on V100 with standard COCO val2017 images). Checking more details about the inference time of DETR in facebookresearch/detr#110.

[zxsitu]

@dddzg Thank you for your prompt response. Yes, I am using custom object detection dataset in COCO format, I have finished training the model and now I want to train the model, I will try your suggestion.

[dddzg]

Good luck!

[zxsitu]

@dddzg Hello, how can I run the "computer_flops.py" file?
I changed PATH_TO_COCO in line 52 to my custom dataset path.
Then I run it, why does it automatically download a model to test it?
I think I should change lines 59~63
https://gist.github.com/fmassa/c0fbb9fe7bf53b533b5cc241f5c8234c#file-compute_flops-py-L59-L63

device = torch.device('cuda')
results = {}
for model_name in ['detr_resnet50']:
    model = torch.hub.load('facebookresearch/detr', model_name, pretrained=True)
    model.to(device)

How can I modify these lines to read my trained model? The path of my trained model is in "outputs/checkpint.pth". Thank you, this is really bothering me.