Coverage for transformer_lens/components/unembed.py: 100%

1"""Hooked Transformer Unembed Component. 

2 

3This module contains all the component :class:`Unembed`. 

4""" 

5 

6from typing import Dict, Union 

7 

8import torch 

9import torch.nn as nn 

10from jaxtyping import Float 

11 

12from transformer_lens.HookedTransformerConfig import HookedTransformerConfig 

13from transformer_lens.utilities.addmm import batch_addmm 

14 

15 

16class Unembed(nn.Module): 

17 def __init__(self, cfg: Union[Dict, HookedTransformerConfig]): 

18 super().__init__() 

19 self.cfg = HookedTransformerConfig.unwrap(cfg) 

20 # Note that there's a separate variable for d_vocab_out and d_vocab (the input vocab size). For language tasks these are always the same, but for algorithmic tasks we may want them to be different. 

21 self.W_U: Float[torch.Tensor, "d_model d_vocab_out"] = nn.Parameter( 

22 torch.empty(self.cfg.d_model, self.cfg.d_vocab_out, dtype=self.cfg.dtype) 

23 ) 

24 self.b_U: Float[torch.Tensor, "d_vocab_out"] = nn.Parameter( 

25 torch.zeros(self.cfg.d_vocab_out, dtype=self.cfg.dtype) 

26 ) 

27 

28 def forward( 

29 self, residual: Float[torch.Tensor, "batch pos d_model"] 

30 ) -> Float[torch.Tensor, "batch pos d_vocab_out"]: 

31 return batch_addmm(self.b_U, self.W_U, residual)