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

Remove zero init and constant biasing for to_gate #304

Merged
merged 2 commits into from
Oct 3, 2024
Merged

Remove zero init and constant biasing for to_gate #304

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
64066cb
Update attention.py
amorehead Oct 3, 2024
9b96216
Update attention.py
amorehead Oct 3, 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
18 changes: 9 additions & 9 deletions alphafold3_pytorch/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
from __future__ import annotations
from beartype.typing import NamedTuple, Tuple
from functools import partial

import torch
from torch import nn, Tensor
Expand All @@ -18,6 +19,10 @@
typecheck
)

# alias

LinearNoBias = partial(nn.Linear, bias = False)

# helpers

def exists(val):
Expand Down Expand Up @@ -178,7 +183,6 @@ def __init__(
num_memory_kv: int = 0,
enable_attn_softclamp = False,
attn_softclamp_value = 50.,
init_gate_bias = -2.,
softmax_full_precision = False
):
super().__init__()
Expand Down Expand Up @@ -209,8 +213,8 @@ def __init__(
self.merge_heads = Rearrange('b h n d -> b n (h d)')

self.to_q = nn.Linear(dim, dim_inner, bias = query_bias)
self.to_kv = nn.Linear(dim, dim_inner * 2, bias = False)
self.to_out = nn.Linear(dim_inner, dim, bias = False)
self.to_kv = LinearNoBias(dim, dim_inner * 2)
self.to_out = LinearNoBias(dim_inner, dim)

self.memory_kv = None

Expand All @@ -224,11 +228,7 @@ def __init__(
self.to_gates = None

if gate_output:
gate_linear = nn.Linear(dim, dim_inner)
nn.init.zeros_(gate_linear.weight)
nn.init.constant_(gate_linear.bias, init_gate_bias)

self.to_gates = gate_linear
self.to_gates = nn.Sequential(LinearNoBias(dim, dim_inner), nn.Sigmoid())

@typecheck
def forward(
Expand Down Expand Up @@ -266,7 +266,7 @@ def forward(

if exists(self.to_gates):
gates = self.to_gates(seq)
out = out * gates.sigmoid()
out = out * gates

# combine heads

Expand Down