Wang Xinjiang
Wang Xinjiang
Both the training time and memory only increase by around 5%-10% compared to DINO. I have also encountered a heavy slowdown with around 2x as much time per iteration as...
Sorry. It does not work for multiple ratios at present. But you may try to change the shape size in this line of code to support this feature.
Thanks for your interest. The initial code only set `num_gt = int(scores.sum() + 0.5)` as default. After submitting the paper, we also tried to add another `dynamic_ratio` to control the...
Sorry, this config is supposed to be run on 8 GPUs with `2x8` meaning samples_per_gpu=2 and the total number of GPUs = 8.
Hi, Thanks for your interest. Only the mean gmm_thr value for all categories is logged to be printed out. Bounding boxes are filtered with a per-class threshold during training. Please...
Hi thanks for your contribution. This work is impressive. There are a few comments after an overview of this PR. - I note that the improvement of 1x training +...
The [ema hook](https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/ema.py) may help you to restructure your code. We may target on wrapping all pipelines in a single hook class.