PAIR/PAIR-Diffusion

1import torch2from torch import nn3 4 5class LitEma(nn.Module):6    def __init__(self, model, decay=0.9999, use_num_upates=True):7        super().__init__()8        if decay < 0.0 or decay > 1.0:9            raise ValueError('Decay must be between 0 and 1')10 11        self.m_name2s_name = {}12        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))13        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates14        else torch.tensor(-1, dtype=torch.int))15 16        for name, p in model.named_parameters():17            if p.requires_grad:18                # remove as '.'-character is not allowed in buffers19                s_name = name.replace('.', '')20                self.m_name2s_name.update({name: s_name})21                self.register_buffer(s_name, p.clone().detach().data)22 23        self.collected_params = []24 25    def reset_num_updates(self):26        del self.num_updates27        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int))28 29    def forward(self, model):30        decay = self.decay31 32        if self.num_updates >= 0:33            self.num_updates += 134            decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates))35 36        one_minus_decay = 1.0 - decay37 38        with torch.no_grad():39            m_param = dict(model.named_parameters())40            shadow_params = dict(self.named_buffers())41 42            for key in m_param:43                if m_param[key].requires_grad:44                    sname = self.m_name2s_name[key]45                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])46                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))47                else:48                    assert not key in self.m_name2s_name49 50    def copy_to(self, model):51        m_param = dict(model.named_parameters())52        shadow_params = dict(self.named_buffers())53        for key in m_param:54            if m_param[key].requires_grad:55                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)56            else:57                assert not key in self.m_name2s_name58 59    def store(self, parameters):60        """61        Save the current parameters for restoring later.62        Args:63          parameters: Iterable of `torch.nn.Parameter`; the parameters to be64            temporarily stored.65        """66        self.collected_params = [param.clone() for param in parameters]67 68    def restore(self, parameters):69        """70        Restore the parameters stored with the `store` method.71        Useful to validate the model with EMA parameters without affecting the72        original optimization process. Store the parameters before the73        `copy_to` method. After validation (or model saving), use this to74        restore the former parameters.75        Args:76          parameters: Iterable of `torch.nn.Parameter`; the parameters to be77            updated with the stored parameters.78        """79        for c_param, param in zip(self.collected_params, parameters):80            param.data.copy_(c_param.data)81