|
--- |
|
license: mit |
|
--- |
|
|
|
ghostmax/softmax1 trained on tinystories for 500M tokens. |
|
|
|
ckpt will be out_* files trained 2 models 1 with regular softmax |
|
|
|
REPO: https://github.com/joey00072/llama2.c/tree/ghostmax |
|
|
|
softmax: https://wandb.ai/shubhamchoudhari00072/ghostmax/reports/loss-val-23-08-25-21-48-58---Vmlldzo1MjM0MjUw <br> |
|
ghostmax: https://wandb.ai/shubhamchoudhari00072/ghostmax/reports/loss-val-23-08-25-21-50-15---Vmlldzo1MjM0MjYw <br> |
|
<br> |
|
|
|
```python |
|
def softmax(x, dim=None): |
|
e_x = torch.exp(x - torch.max(x, dim=dim, keepdim=True)[0]) |
|
return e_x / e_x.sum(dim=dim, keepdim=True) |
|
|
|
def ghostmax(x, dim=None): |
|
e_x = torch.exp(x - torch.max(x, dim=dim, keepdim=True)[0]) |
|
return e_x / (1+e_x.sum(dim=dim, keepdim=True) ) |
|
|
|
``` |
|
<br> |
|
<iframe src="https://wandb.ai/shubhamchoudhari00072/ghostmax/runs/cbeei9uh?workspace=user-shubhamchoudhari00072" style="border:none;height:1024px;width:100%"> |
|
<br> |
|
|
|
|