Yutao Xu

Results 14 comments of Yutao Xu

The results for ResNet-50 is 33.7 without gn head, currently. Still in debugging

I haven't tried the python version yet. Maybe you can refer to the official version.

I use pytorch 1.1.0 in my code. If you use the latest version, you can either omit the warning or replace torch.uint8 with torch.bool.

iou_loss 是通过偏移计算出来的,这部分我参照了官方代码 https://github.com/tianzhi0549/FCOS

"load": false, "save": true, "pretrain": true, "freeze_bn": true, "freeze_stages": 1, "epoches": 30, "nbatch_train": 28, "nbatch_eval": 28, "device": [0,1,2,3], "num_workers": 14, "lr_base": 0.01, "lr_gamma": 0.1, "lr_schedule": [16000, 22000], "momentum": 0.9, "weight_decay":...

This configuration can achieve 77(map50) in voc2007 test

我的代码没有将正例周围的点忽略,所以需要用梯度裁剪,你可以把clip_grad的值调低一点。

评估指标可以用MSCOCO数据集提供的api,在我的run_cocoeval.py里有,需要按照它的数据格式组织。如果你需要一个粗略的测评方法,可以参照fcos/utils_box/eval_csv.py,使用VOC的测评方法。

update google drive link, see readme.

这个代码目前没有在1x达到37的精度(包括之前按照论文配置的复现),可能是resize问题或者忽略了一些trick。