Closed johnnylu305 closed 4 years ago
Yeah. When you refer to the caffe version, the only provided init.model is pretrained on COCO, if you can find the init model without COCO, please tell me. I really think use this model is not a good choice.
Yeah. We are working on this problem. We will inform you if we find any solution.
Since you said that your setting of DeepLabv2-resnet101 was the same as the paper, I wonder whether you used COCO pretrain weight or not?