| import torch |
| from torch import nn |
|
|
|
|
| class LitEma(nn.Module): |
| def __init__(self, model, decay=0.9999, use_num_upates=True): |
| super().__init__() |
| if decay < 0.0 or decay > 1.0: |
| raise ValueError('Decay must be between 0 and 1') |
|
|
| self.m_name2s_name = {} |
| self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) |
| self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates |
| else torch.tensor(-1, dtype=torch.int)) |
|
|
| for name, p in model.named_parameters(): |
| if p.requires_grad: |
| |
| s_name = name.replace('.', '') |
| self.m_name2s_name.update({name: s_name}) |
| self.register_buffer(s_name, p.clone().detach().data) |
|
|
| self.collected_params = [] |
|
|
| def reset_num_updates(self): |
| del self.num_updates |
| self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int)) |
|
|
| def forward(self, model): |
| decay = self.decay |
|
|
| if self.num_updates >= 0: |
| self.num_updates += 1 |
| decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates)) |
|
|
| one_minus_decay = 1.0 - decay |
|
|
| with torch.no_grad(): |
| m_param = dict(model.named_parameters()) |
| shadow_params = dict(self.named_buffers()) |
|
|
| for key in m_param: |
| if m_param[key].requires_grad: |
| sname = self.m_name2s_name[key] |
| shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) |
| shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) |
| else: |
| assert not key in self.m_name2s_name |
|
|
| def copy_to(self, model): |
| """ |
| Copying the ema state (i.e., buffers) to the targeted model |
| Input: |
| model: targeted model |
| """ |
| m_param = dict(model.named_parameters()) |
| shadow_params = dict(self.named_buffers()) |
| for key in m_param: |
| if m_param[key].requires_grad: |
| m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) |
| else: |
| assert not key in self.m_name2s_name |
|
|
| def store(self, parameters): |
| """ |
| Save the parameters of the targeted model into the temporary pool for restoring later. |
| Args: |
| parameters: parameters of the targeted model. |
| Iterable of `torch.nn.Parameter`; the parameters to be temporarily stored. |
| """ |
| self.collected_params = [param.clone() for param in parameters] |
|
|
| def restore(self, parameters): |
| """ |
| Restore the parameters from the temporaty pool (stored with the `store` method). |
| Useful to validate the model with EMA parameters without affecting the |
| original optimization process. Store the parameters before the |
| `copy_to` method. After validation (or model saving), use this to |
| restore the former parameters. |
| Args: |
| parameters: Iterable of `torch.nn.Parameter`; the parameters to be |
| updated with the stored parameters. |
| """ |
| for c_param, param in zip(self.collected_params, parameters): |
| param.data.copy_(c_param.data) |
|
|
| def resume(self, ckpt, num_updates): |
| """ |
| Resume from the targeted checkpoint, i.e., copying the checkpoints to ema buffers |
| Input: |
| model: targerted model |
| """ |
| self.register_buffer('num_updates', torch.tensor(num_updates, dtype=torch.int)) |
|
|
| shadow_params = dict(self.named_buffers()) |
| for key, value in ckpt.items(): |
| try: |
| shadow_params[self.m_name2s_name[key]].data.copy_(value.data) |
| except: |
| if key.startswith('module') and key not in shadow_params: |
| key = key[7:] |
| shadow_params[self.m_name2s_name[key]].data.copy_(value.data) |
|
|
