| from torch.optim import Optimizer | |
| class PassThroughOptimizer(Optimizer): | |
| """ | |
| A dummy PyTorch optimizer that does nothing during the step() function. | |
| This can be used for testing purposes or for training Lightning modules with manual | |
| parameter updates. In Lightning, we need to call `opt.step()` so that | |
| trainer.global_step is incremented, even if we are doing manual optimization. | |
| """ | |
| def __init__(self, params, lr=0.01): | |
| defaults = dict(lr=lr) | |
| super(PassThroughOptimizer, self).__init__(params, defaults) | |
| def step(self, closure=None): | |
| return None | |
| def zero_grad(self): | |
| for group in self.param_groups: | |
| for p in group["params"]: | |
| if p.grad is not None: | |
| p.grad.detach_() | |
| p.grad.zero_() | |
| def state_dict(self): | |
| # needed for lightning checkpointing | |
| return {} | |
| def load_state_dict(self, state_dict): | |
| # needed for lightning checkpointing | |
| pass | |