Bottleneck问题

[Albertchamberlain]

由Basicblock切换成Bottleneck运行模型报错

[Albertchamberlain]

直接切换运行即可复现

[TosaSato2]

I changed the block to "BottleNeck" to change the backbone to a ResNet-50, and it worked, however I had to make the following modifications to some custom data LSTR scripts:

1. In the config/LSTR.json script I changed:

   • res_layers by those of ResNet-50 to [3, 4, 6, 3]
   • res_dims by those of ResNet-50 to [64, 128, 256, 512]

2. In the script models/py_utils/kp.py on line 151:

self.input_proj = nn.Conv2d(res_dims[-1], hidden_dim, kernel_size=1) # the same as channel of self.layer4

I changed the res_dims[-1] to 2048 which is the same as channel of self.layer4 for ResNet-50.