Update modeling_super_linear.py
Browse files- modeling_super_linear.py +24 -2
modeling_super_linear.py
CHANGED
|
@@ -208,9 +208,31 @@ class RLinear(nn.Module):
|
|
| 208 |
final_scaling = original_norm / new_norm if new_norm.item() != 0 else 1.0
|
| 209 |
#final_scaling = 1
|
| 210 |
new_W = new_W * final_scaling
|
|
|
|
| 211 |
|
| 212 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 213 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 214 |
|
| 215 |
def forward(self, x):
|
| 216 |
# x: [Batch, Input length,Channel]
|
|
@@ -218,7 +240,7 @@ class RLinear(nn.Module):
|
|
| 218 |
if x.shape[1] < self.seq_len:
|
| 219 |
if self.zero_shot_Linear is None:
|
| 220 |
#print(F"new Lookkback : {x.shape[1]}")
|
| 221 |
-
self.transform_model(x.shape[1],
|
| 222 |
|
| 223 |
x = x.clone()
|
| 224 |
#x = x * (x.shape[1]/512)
|
|
|
|
| 208 |
final_scaling = original_norm / new_norm if new_norm.item() != 0 else 1.0
|
| 209 |
#final_scaling = 1
|
| 210 |
new_W = new_W * final_scaling
|
| 211 |
+
self.zero_shot_Linear = new_W
|
| 212 |
|
| 213 |
+
else:
|
| 214 |
+
W = self.Linear.weight.detach()
|
| 215 |
+
target_indices = torch.linspace(0, self.seq_len - 1, steps=new_lookback, device=W.device)
|
| 216 |
+
source_indices = torch.arange(0, self.seq_len, device=W.device).float()
|
| 217 |
+
|
| 218 |
+
# Initialize the new weight matrix
|
| 219 |
+
new_W = torch.zeros((W.size(0), new_lookback), device=W.device)
|
| 220 |
+
|
| 221 |
+
# Linear interpolation for each row
|
| 222 |
+
for i in range(W.size(0)):
|
| 223 |
+
new_W[i] = torch.tensor([torch.sum(W[i] * (1 - torch.abs(idx - source_indices) / self.seq_len).clamp(min=0))
|
| 224 |
+
for idx in target_indices], device=W.device)
|
| 225 |
+
|
| 226 |
+
# Maintain the same norm as the original weights
|
| 227 |
+
original_norm = torch.norm(W, p=2)
|
| 228 |
+
new_norm = torch.norm(new_W, p=2)
|
| 229 |
+
final_scaling = original_norm / new_norm if new_norm.item() != 0 else 1.0
|
| 230 |
+
new_W = new_W * final_scaling
|
| 231 |
|
| 232 |
+
self.zero_shot_Linear = new_W
|
| 233 |
+
|
| 234 |
+
|
| 235 |
+
|
| 236 |
|
| 237 |
def forward(self, x):
|
| 238 |
# x: [Batch, Input length,Channel]
|
|
|
|
| 240 |
if x.shape[1] < self.seq_len:
|
| 241 |
if self.zero_shot_Linear is None:
|
| 242 |
#print(F"new Lookkback : {x.shape[1]}")
|
| 243 |
+
self.transform_model(x.shape[1],3)
|
| 244 |
|
| 245 |
x = x.clone()
|
| 246 |
#x = x * (x.shape[1]/512)
|