加--fp16训练老师模型后，再训练学生模型报错

[LiangYongAI]

@xiao-hua-sheng 在训练老师模型时加--fp16，然后再用老师模型指导学生模型训练,也加--fp16，在训练开始的时候报错，Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same，完整错误如下所示：

File "/opt/YOLOX-Distill/yolox/core/trainer.py", line 105, in train_one_iter
self.T_pre = self.T_model(inps)
│ │ │ │ └ tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ │ │ │ [113., 111., 111., ..., 73., 75., 75.],
│ │ │ │ [114., ...
│ │ │ └ YOLOX(
│ │ │ (backbone): YOLOPAFPN(
│ │ │ (backbone): CSPDarknet(
│ │ │ (stem): Focus(
│ │ │ (conv): BaseConv(
│ │ │ (conv): ...
│ │ └ <yolox.core.trainer.Trainer object at 0x7f4baf145ed0>
│ └ None
└ <yolox.core.trainer.Trainer object at 0x7f4baf145ed0>

File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
│ │ └ {}
│ └ (tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ [113., 111., 111., ..., 73., 75., 75.],
│ [114.,...
└ <bound method YOLOX.forward of YOLOX(
(backbone): YOLOPAFPN(
(backbone): CSPDarknet(
(stem): Focus(
(conv...

File "/opt/YOLOX-Distill/yolox/models/yolox.py", line 31, in forward
fpn_outs = self.backbone(x)
│ └ tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ [113., 111., 111., ..., 73., 75., 75.],
│ [114., ...
└ YOLOX(
(backbone): YOLOPAFPN(
(backbone): CSPDarknet(
(stem): Focus(
(conv): BaseConv(
(conv): ...

File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
│ │ └ {}
│ └ (tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ [113., 111., 111., ..., 73., 75., 75.],
│ [114.,...
└ <bound method YOLOPAFPN.forward of YOLOPAFPN(
(backbone): CSPDarknet(
(stem): Focus(
(conv): BaseConv(
(c...

File "/opt/YOLOX-Distill/yolox/models/yolo_pafpn.py", line 93, in forward
out_features = self.backbone(input)
│ └ tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ [113., 111., 111., ..., 73., 75., 75.],
│ [114., ...
└ YOLOPAFPN(
(backbone): CSPDarknet(
(stem): Focus(
(conv): BaseConv(
(conv): Conv2d(12, 48, kernel_size=(3...

File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
│ │ └ {}
│ └ (tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ [113., 111., 111., ..., 73., 75., 75.],
│ [114.,...
└ <bound method CSPDarknet.forward of CSPDarknet(
(stem): Focus(
(conv): BaseConv(
(conv): Conv2d(12, 48, kernel_si...

File "/opt/YOLOX-Distill/yolox/models/darknet.py", line 170, in forward
x = self.stem(x)
│ └ tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ [113., 111., 111., ..., 73., 75., 75.],
│ [114., ...
└ CSPDarknet(
(stem): Focus(
(conv): BaseConv(
(conv): Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1...

File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
│ │ └ {}
│ └ (tensor([[[[112., 112., 112., ..., 77., 77., 77.],
│ [113., 111., 111., ..., 73., 75., 75.],
│ [114.,...
└ <bound method Focus.forward of Focus(
(conv): BaseConv(
(conv): Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), paddi...

File "/opt/YOLOX-Distill/yolox/models/network_blocks.py", line 210, in forward
return self.conv(x)
│ └ tensor([[[[112., 112., 113., ..., 77., 77., 77.],
│ [114., 112., 114., ..., 76., 76., 75.],
│ [115., ...
└ Focus(
(conv): BaseConv(
(conv): Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn)...

File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
│ │ └ {}
│ └ (tensor([[[[112., 112., 113., ..., 77., 77., 77.],
│ [114., 112., 114., ..., 76., 76., 75.],
│ [115.,...
└ <bound method BaseConv.forward of BaseConv(
(conv): Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=...

File "/opt/YOLOX-Distill/yolox/models/network_blocks.py", line 51, in forward
return self.act(self.bn(self.conv(x)))
│ │ │ └ tensor([[[[112., 112., 113., ..., 77., 77., 77.],
│ │ │ [114., 112., 114., ..., 76., 76., 75.],
│ │ │ [115., ...
│ │ └ BaseConv(
│ │ (conv): Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
│ │ (bn): BatchNorm2d(48, eps...
│ └ BaseConv(
│ (conv): Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
│ (bn): BatchNorm2d(48, eps...
└ BaseConv(
(conv): Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(48, eps...

File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
│ │ └ {}
│ └ (tensor([[[[112., 112., 113., ..., 77., 77., 77.],
│ [114., 112., 114., ..., 76., 76., 75.],
│ [115.,...
└ <bound method Conv2d.forward of Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)>
File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/conv.py", line 447, in forward
return self._conv_forward(input, self.weight, self.bias)
│ │ │ │ └ Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
│ │ │ └ Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
│ │ └ tensor([[[[112., 112., 113., ..., 77., 77., 77.],
│ │ [114., 112., 114., ..., 76., 76., 75.],
│ │ [115., ...
│ └ <function Conv2d._conv_forward at 0x7f4bc0e67d40>
└ Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
File "/home/ubuntu/anaconda3/envs/yolox/lib/python3.7/site-packages/torch/nn/modules/conv.py", line 444, in _conv_forward
self.padding, self.dilation, self.groups)
│ │ │ │ │ └ 1
│ │ │ │ └ Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
│ │ │ └ (1, 1)
│ │ └ Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
│ └ (1, 1)
└ Conv2d(12, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)

RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same

[xiao-hua-sheng]

--fp16暂时只能为False，工作毕竟忙，后面有时间在修改