search

Nioolek / PPYOLOE_pytorch

An unofficial implementation of Pytorch version PP-YOLOE,based on Megvii YOLOX training code.
Apache License 2.0
181 stars 20 forks source link

cocoeval花的时间变长了 #13

Open vvhj opened 2 years ago

vvhj commented 2 years ago

这个是我在yolox上测试的时间 2022-01-22 08:28:29.327 | INFO | yolox.evaluators.coco_evaluator:evaluate_prediction:208 - Evaluate in main process... 2022-01-22 08:28:33.417 | INFO | yolox.evaluators.coco_evaluator:evaluate_prediction:241 - Loading and preparing results... 2022-01-22 08:28:39.083 | INFO | yolox.evaluators.coco_evaluator:evaluate_prediction:241 - DONE (t=5.67s)

这个是我用您这测试的时间 2022-06-13 05:28:58.913 | INFO | yolox.evaluators.coco_evaluator:evaluate_prediction:244 - Evaluate in main process... 2022-06-13 05:31:54.250 | INFO | yolox.evaluators.coco_evaluator:evaluate_prediction:277 - Loading and preparing results... 2022-06-13 05:33:08.532 | INFO | yolox.evaluators.coco_evaluator:evaluate_prediction:277 - DONE (t=74.28s)

用时差10倍

sdreamforchen commented 2 years ago

请问eval的时候是重参后的架构吗

vvhj commented 2 years ago

分析了一下，推理速度是ok的，感觉问题出在forward_eval 里 return torch.cat( # 目标是(1,8400,85) [ pred_bboxes, # torch.ones((b, pred_bboxes.shape[1], 1), device=pred_bboxes.device, dtype=pred_bboxes.dtype), cls_score_list.permute(0, 2, 1) ], axis=-1)

torch.ones((b, pred_bboxes.shape[1], 1), device=pred_bboxes.device, dtype=pred_bboxes.dtype),我怀疑是这行导致每个结果都要计算所以在走cocoeval的时候要花较多的时间