search

ucla-mobility / V2V4Real

[CVPR2023 Highlight] The official codebase for paper "V2V4Real: A large-scale real-world dataset for Vehicle-to-Vehicle Cooperative Perception"
Other
192 stars 12 forks source link

tensor data type mismatch #30

Closed drowning-in-codes closed 2 months ago

drowning-in-codes commented 2 months ago

Thanks for your great codebase. I'm a little confused about some of the code in train_da.py:

with torch.cuda.amp.autocast():  
          ouput_dict = model(total_batch_data) # data type float16 
da_loss = DA_module(ouput_dict)

When using autocast in domain adaptation, the output type will be float16, and then the output will be passed through DA_module, which is float32, which will result in a data type mismatch, and my solution is to convert it to float32 after autocasting , and it is also possible to put DA_module() directly in the region of autocast.