drop in replacement of nn.Linear
| 14 | |
| 15 | |
| 16 | class ActLinear(nn.Module): |
| 17 | """ |
| 18 | drop in replacement of nn.Linear |
| 19 | """ |
| 20 | |
| 21 | def __init__(self, base: nn.Linear): |
| 22 | super().__init__() |
| 23 | self.base = base |
| 24 | # self.register_buffer('activation_norms', torch.zeros([base.in_features], device=self.base.weight.device, requires_grad=False)) |
| 25 | self.activation_norms = torch.zeros( |
| 26 | [base.in_features], device=self.base.weight.device, requires_grad=False |
| 27 | ) |
| 28 | self.n_samples = 0 |
| 29 | self.record_activation = True |
| 30 | |
| 31 | def clear_act_buffer(self): |
| 32 | self.activation_norms.fill_(0.0) |
| 33 | self.n_samples = 0 |
| 34 | |
| 35 | def forward(self, x): |
| 36 | # TODO: normalize for numerical stability |
| 37 | # TODO: remove this after pruning |
| 38 | |
| 39 | # DEBUG: |
| 40 | # print("input zero percentage", (x==0).sum() / x.numel() ) |
| 41 | |
| 42 | if self.record_activation: |
| 43 | if hasattr(self, "mask") and self.mask is not None: |
| 44 | x_ = x[self.mask] |
| 45 | else: |
| 46 | x_ = x |
| 47 | |
| 48 | bs = x_.nelement() // x_.shape[-1] |
| 49 | self.activation_norms = self.activation_norms * ( |
| 50 | self.n_samples / (self.n_samples + bs) |
| 51 | ) + (x_ * x_).view(-1, x_.shape[-1]).sum(dim=0) * ( |
| 52 | 1.0 / (self.n_samples + bs) |
| 53 | ) |
| 54 | self.n_samples += bs |
| 55 | |
| 56 | out = self.base(x) |
| 57 | return out |
| 58 | |
| 59 | |
| 60 | class no_act_recording: |