C-H-Wong

Post-processing is essential for current detection scheme, however traditional **nms** is not the only choice, especially when dealing with enormous detection bboxes which is quite resource-consuming. We propose **large batch...

It looks like a normal fp to me, with score of 0.02, it's not a serious problem. It's just test process, we don't need to further filter the results. By...

Right, 1080ti

> Since am having the same error, I downgraded PyTorch from 1.9 to 1.8. Now it's throwing a CUDA error. Could you please specify which CUDA version to use? Requirements...

Weights and other parameters are automatically saved through training process, by changing the saving interval [here](https://github.com/Media-Smart/vedatad/blob/e4423b371dd11c0a6c31fc8b4aa49b7ce26827be/configs/trainval/daotad/daotad_i3d_r50_e700_thumos14_rgb.py#L199) you can save the weights of any epoch.

When evaluating different mAP, you only need to pass different **iou_thr** value to **evaluate** function [here](https://github.com/Media-Smart/vedatad/blob/e4423b371dd11c0a6c31fc8b4aa49b7ce26827be/vedatad/datasets/custom.py#L324) in tools/test.py [here](https://github.com/Media-Smart/vedatad/blob/e4423b371dd11c0a6c31fc8b4aa49b7ce26827be/tools/test.py#L77). We suggest to save the inference results first (set --out when...

You could check the parameters setting [here](https://github.com/Media-Smart/vedatad/blob/main/setup.py#L151) in setup.py