Revert .pth model to FP32

quantization
1

The .pth was train with mixed precision. How to I make the model into FP32 again?

2

generally my understanding is that training with mixed precision occurs as follows (pseudocode):

class Model:
  <model code>

def train(net):
  with torch.cuda.autocast()
    <train loop>

main:
  net = Model()
  train(net)
  save(net)

in which case, your model should be saved in fp32, since the autocast part just alters the calculations in there, not the model itself. If you did something different, i’m not sure if its possible to revert. I would double check the documentation thought to make sure you are doing things as expected:

https://pytorch.org/docs/stable/amp.html
https://pytorch.org/docs/stable/notes/amp_examples.html