Update modeling_super_linear.py
Browse files- modeling_super_linear.py +18 -2
modeling_super_linear.py
CHANGED
|
@@ -214,6 +214,22 @@ class RLinear(nn.Module):
|
|
| 214 |
def forward(self, x):
|
| 215 |
# x: [Batch, Input length,Channel]
|
| 216 |
x_shape = x.shape
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 217 |
if len(x_shape) == 2:
|
| 218 |
x = x.unsqueeze(-1)
|
| 219 |
|
|
@@ -575,8 +591,8 @@ class SuperLinearForCausalLM(PreTrainedModel, GenerationMixin):
|
|
| 575 |
# backbone expects (B, C, L)
|
| 576 |
x_enc = inputs_embeds
|
| 577 |
|
| 578 |
-
if x_enc.shape[1] < 512:
|
| 579 |
-
x_enc = self.fourier_interp_dim1(x_enc)
|
| 580 |
|
| 581 |
|
| 582 |
# backbone returns (B, pred_len, C)
|
|
|
|
| 214 |
def forward(self, x):
|
| 215 |
# x: [Batch, Input length,Channel]
|
| 216 |
x_shape = x.shape
|
| 217 |
+
if x.shape[1] < self.seq_len:
|
| 218 |
+
if self.zero_shot_Linear is None:
|
| 219 |
+
print(F"new Lookkback : {x.shape[1]}")
|
| 220 |
+
self.transform_model(x.shape[1])
|
| 221 |
+
|
| 222 |
+
if len(x_shape) == 2:
|
| 223 |
+
x = x.unsqueeze(-1)
|
| 224 |
+
x = x.clone()
|
| 225 |
+
x = self.revin_layer(x, 'norm')
|
| 226 |
+
x = F.linear(x.permute(0,2,1), self.zero_shot_Linear).permute(0,2,1).clone()
|
| 227 |
+
x = self.revin_layer(x, 'denorm')
|
| 228 |
+
if len(x_shape) == 2:
|
| 229 |
+
x = x.squeeze(-1)
|
| 230 |
+
return x
|
| 231 |
+
|
| 232 |
+
|
| 233 |
if len(x_shape) == 2:
|
| 234 |
x = x.unsqueeze(-1)
|
| 235 |
|
|
|
|
| 591 |
# backbone expects (B, C, L)
|
| 592 |
x_enc = inputs_embeds
|
| 593 |
|
| 594 |
+
'''if x_enc.shape[1] < 512:
|
| 595 |
+
x_enc = self.fourier_interp_dim1(x_enc)'''
|
| 596 |
|
| 597 |
|
| 598 |
# backbone returns (B, pred_len, C)
|