Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,26 @@
|
|
1 |
---
|
2 |
license: mit
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: mit
|
3 |
---
|
4 |
+
|
5 |
+
ghostmax/softmax1 trained on tinystories for 500M tokens.
|
6 |
+
|
7 |
+
ckpt will be out_* files trained 2 models 1 with regular softmax
|
8 |
+
|
9 |
+
REPO: https://github.com/joey00072/llama2.c/tree/ghostmax
|
10 |
+
|
11 |
+
softmax: https://wandb.ai/shubhamchoudhari00072/ghostmax/reports/loss-val-23-08-25-21-48-58---Vmlldzo1MjM0MjUw
|
12 |
+
ghostmax: https://wandb.ai/shubhamchoudhari00072/ghostmax/reports/loss-val-23-08-25-21-50-15---Vmlldzo1MjM0MjYw
|
13 |
+
<iframe src="https://wandb.ai/shubhamchoudhari00072/ghostmax/runs/cbeei9uh?workspace=user-shubhamchoudhari00072" style="border:none;height:1024px;width:100%">
|
14 |
+
|
15 |
+
```python
|
16 |
+
def softmax(x, dim=None):
|
17 |
+
e_x = torch.exp(x - torch.max(x, dim=dim, keepdim=True)[0])
|
18 |
+
return e_x / e_x.sum(dim=dim, keepdim=True)
|
19 |
+
|
20 |
+
def ghostmax(x, dim=None):
|
21 |
+
e_x = torch.exp(x - torch.max(x, dim=dim, keepdim=True)[0])
|
22 |
+
return e_x / (1+e_x.sum(dim=dim, keepdim=True) )
|
23 |
+
|
24 |
+
```
|
25 |
+
|
26 |
+
|