Performance discrepancy when training from scratch w/ PyTorch 1.4

mileyan / simple_shot

MIT License

129 stars 27 forks source link

Performance discrepancy when training from scratch w/ PyTorch 1.4 #5

Open bertinetto opened 4 years ago

bertinetto commented 4 years ago

Hi, Thanks for the nice work and for sharing the code.

I have tried replicating the results of the paper (training from scratch) but with no luck. I have followed the instructions of the readme and tried both pytorch/cuda-toolkit 1.4/10.0. For ResNet-10 and ResNet-18, on miniImageNet I am getting a discrepancy between 2% and 3% (absolute).

Thanks in advance for the help!

bertinetto commented 4 years ago

Hi, For some reason, the problem only occurred when using PyTorch 1.4. Using PyTorch 1.0 results are reproduced. Will post here when/if I can find a fix for PyTorch 1.4

mileyan commented 4 years ago

Thanks for pointing it out. I will look at the difference between 1.4 and 1.0.