timm_grad_scaler.py

import torch

try:
    from apex import amp
    has_apex = True
except ImportError:
    amp = None
    has_apex = False

from timm.utils.clip_grad import dispatch_clip_grad

class NativeScalerWithGradAccum:
    state_dict_key = "amp_scaler"

    def __init__(self):
        self._scaler = torch.cuda.amp.GradScaler()

    def __call__(self, loss, optimizer, clip_grad=None, clip_mode='norm', parameters=None, create_graph=False, update_grad=True):
        self._scaler.scale(loss).backward(create_graph=create_graph)

        if update_grad:
            if clip_grad is not None:
                assert parameters is not None
                self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place
                dispatch_clip_grad(parameters, clip_grad, mode=clip_mode)
            else:
                self._scaler.unscale_(optimizer)
            self._scaler.step(optimizer)
            self._scaler.update()

    def state_dict(self):
        return self._scaler.state_dict()

    def load_state_dict(self, state_dict):
        self._scaler.load_state_dict(state_dict)