selsa vid fp16 training error

[fcakyon]

I only add fp16 settings into original selsa fasterrcnnr 50 training config:

# fp16 settings
fp16 = dict(loss_scale=512.)

When I try to train selsa vid model on fp16 mode, I get this error:

Traceback (most recent call last):
  File "tools/train.py", line 168, in <module>
    main()
  File "tools/train.py", line 157, in main
    train_model(
  File "mmtracking\mmtrack\apis\train.py", line 135, in train_model
    runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\mmcv\runner\epoch_based_runner.py", line 125, in run
    epoch_runner(data_loaders[i], **kwargs)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\mmcv\runner\epoch_based_runner.py", line 50, in train
    self.run_iter(data_batch, train_mode=True)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\mmcv\runner\epoch_based_runner.py", line 29, in run_iter
    outputs = self.model.train_step(data_batch, self.optimizer,
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\mmcv\parallel\data_parallel.py", line 67, in train_step
    return self.module.train_step(*inputs[0], **kwargs[0])
  File "mmtracking\mmtrack\models\vid\base.py", line 215, in train_step
    losses = self(**data)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\torch\nn\modules\module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\mmcv\runner\fp16_utils.py", line 84, in new_func
    return old_func(*args, **kwargs)
  File "mmtracking\mmtrack\models\vid\base.py", line 149, in forward
    return self.forward_train(img, img_metas, **kwargs)
  File "mmtracking\mmtrack\models\vid\selsa.py", line 137, in forward_train
    all_x = self.detector.extract_feat(all_imgs)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\mmdet\models\detectors\two_stage.py", line 82, in extract_feat       
    x = self.backbone(img)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\torch\nn\modules\module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\mmdet\models\backbones\resnet.py", line 627, in forward
    x = self.conv1(x)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\torch\nn\modules\module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\torch\nn\modules\conv.py", line 423, in forward
    return self._conv_forward(input, self.weight)
  File "C:\Users\FCA\Miniconda3\envs\mmtracking\lib\site-packages\torch\nn\modules\conv.py", line 419, in _conv_forward
    return F.conv2d(input, weight, self.bias, self.stride,
RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same

[OceanPang]

MMTracking does not support FP16 training yet. It will be supported in the future.

[GT9505]

Hi, @fcakyon , sorry for the late. We have already supported fp16 training in #230 . You can give it a try if you still need it.