Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix gated forward functions #295

Merged
merged 5 commits into from
Sep 20, 2024
Merged

Fix gated forward functions #295

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
0a57e47
support seqpos slicing
callummcdougall Sep 18, 2024
89714c8
fix forward functions for gated
callummcdougall Sep 18, 2024
d38622e
remove seqpos changes
callummcdougall Sep 18, 2024
d9ea96a
fix formatting (remove my changes)
callummcdougall Sep 18, 2024
c16366c
format
jbloomAus Sep 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 13 additions & 6 deletions sae_lens/sae.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -427,6 +427,13 @@ def forward(
with torch.no_grad():
x = x.to(self.dtype)
sae_in = self.reshape_fn_in(x) # type: ignore

# handle run time activation normalization if needed
sae_in = self.run_time_activation_norm_fn_in(sae_in)

# apply b_dec_to_input if using that method.
sae_in = sae_in - (self.b_dec * self.cfg.apply_b_dec_to_input)

gating_pre_activation = sae_in @ self.W_enc + self.b_gate
active_features = (gating_pre_activation > 0).float()

Expand Down Expand Up @@ -455,10 +462,10 @@ def forward(
sae_in = self.reshape_fn_in(x) # type: ignore

# handle run time activation normalization if needed
x = self.run_time_activation_norm_fn_in(x)
sae_in = self.run_time_activation_norm_fn_in(sae_in)

# apply b_dec_to_input if using that method.
sae_in = x - (self.b_dec * self.cfg.apply_b_dec_to_input)
sae_in = sae_in - (self.b_dec * self.cfg.apply_b_dec_to_input)

# "... d_in, d_in d_sae -> ... d_sae",
hidden_pre = sae_in @ self.W_enc + self.b_enc
Expand Down Expand Up @@ -495,11 +502,11 @@ def encode_gated(
magnitude_pre_activation = self.hook_sae_acts_pre(
sae_in @ (self.W_enc * self.r_mag.exp()) + self.b_mag
)
feature_magnitudes = self.hook_sae_acts_post(
self.activation_fn(magnitude_pre_activation)
)
feature_magnitudes = self.activation_fn(magnitude_pre_activation)

return active_features * feature_magnitudes
feature_acts = self.hook_sae_acts_post(active_features * feature_magnitudes)

return feature_acts

def encode_jumprelu(
self, x: Float[torch.Tensor, "... d_in"]
Expand Down
Loading