Closed kfxw closed 4 years ago
Fixed most of problems mentioned in the last review, except the 'throwout_param', which is used as input for custom op. Custom op cannot accept class type as input. Also the training is further speedup, with quite compatible running speed with the pytorch counterpart. Memory usage is ~300MB larger.
I've reimplemented the FCOS model. The ResNet-50 based one has the same performance on coco minival2014 set with mAP of 36.6, which behaves similar with the one trained with official pytorch code. Besides, the training speed is still kind of slower than official code. More refinement may be needed in the future.