Skip to content

fix: seperate loggits by task #420

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

Merged
chungongyu merged 2 commits into from
Jul 2, 2026
Merged

fix: seperate loggits by task #420

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a969bde
fix: seperate loggits by task
chungongyu Jul 1, 2026
f45705e
refactor: works when data has no batch dim
chungongyu Jul 1, 2026
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
89 changes: 51 additions & 38 deletions profold2/model/head.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
import sys
import logging
import math

import torch
from torch import nn
Expand Down Expand Up @@ -1352,7 +1353,6 @@ def __init__(
self.return_motifs = env('profold2_fitness_return_motifs', defval=True, dtype=bool)

def predict(self, variant_logit, variant_mask, gating=None):
variant_logit = variant_logit[..., None]
if exists(gating):
variant_logit = variant_logit * gating
if self.pooling == 'sum':
Expand Down Expand Up @@ -1380,31 +1380,33 @@ def forward(self, headers, representations, batch):
else:
wab = None

def _hamiton_run(variant):
variant = F.one_hot(variant.long(), num_class)
def _hamiton_run(variant, variant_mask):
variant = (
F.one_hot(variant.long(), num_class) * variant_mask[..., None]
).float()
if exists(wab):
hi = torch.einsum(
'... m j d,... i j q,q c d,d -> ... m i c', variant.float(), eij, wab, self.mask
'... m j t d,... i j q,q c d,d -> ... m i t c', variant, eij, wab, self.mask
)
else:
hi = torch.einsum(
'... m j d,... i j c d,d -> ... m i c', variant.float(), eij, self.mask
'... m j t d,... i j c d,d -> ... m i t c', variant, eij, self.mask
)
motifs = rearrange(ei, '... i c -> ... () i c') + hi
motifs = rearrange(ei, '... i c -> ... () i () c') + hi
if self.log_softmax:
logits = torch.einsum(
'... m i c,... m i c -> ... m i', F.log_softmax(motifs, dim=-1), variant.float()
'... i t c,... i t c -> ... i t', F.log_softmax(motifs, dim=-1), variant
)
else:
logits = torch.einsum('... m i c,... m i c -> ... m i', motifs, variant.float())
logits = torch.einsum('... i t c,... i t c -> ... i t', motifs, variant)
if self.return_motifs:
return motifs, logits
return None, logits

def _hamiton_cat(logits):
motifs, logits = zip(*logits)
motifs = torch.cat(motifs, dim=1) if self.return_motifs else None
logits = torch.cat(logits, dim=1)
motifs = torch.cat(motifs, dim=-4) if self.return_motifs else None
logits = torch.cat(logits, dim=-3)
return motifs, logits

if 'variant' in batch:
Expand All @@ -1415,8 +1417,12 @@ def _hamiton_cat(logits):
variant = batch['seq'][..., None, :]
variant_mask = batch['mask'][..., None, :]

variant, variant_mask = variant[..., None], variant_mask[..., None]
if 'variant_task_mask' in batch:
variant_mask = batch['variant_task_mask']

motifs, logits = functional.sharded_apply(
_hamiton_run, [variant],
_hamiton_run, [variant, variant_mask],
shard_size=None if self.training else self.shard_size,
shard_dim=1,
cat_dim=_hamiton_cat
Expand All @@ -1429,12 +1435,8 @@ def _hamiton_cat(logits):
gating = None
r.update(gating=gating)
if not self.training:
if 'variant_task_mask' in batch:
variant_mask = batch['variant_task_mask']
else:
variant_mask = variant_mask[..., None]
variant_mask = variant_mask * variant_mask[:, :1, ...]
variant_logit = logits - logits[:, :1, ...]
variant_mask = variant_mask * variant_mask[..., :1, :, :]
variant_logit = logits - logits[..., :1, :, :]
variant_logit = self.predict(
variant_logit, variant_mask, gating=gating[..., None, :, :]
)
Expand Down Expand Up @@ -1462,14 +1464,20 @@ def loss(self, value, batch):

with accelerator.autocast(enabled=False):
errors = softmax_cross_entropy(
labels=labels,
labels=labels[..., None, :],
logits=motifs.float(),
mask=self.mask,
gammar=self.focal_loss
)
motif_mask = ((variant_label > self.label_threshold) &
variant_label_mask) | (~variant_label_mask)
motif_mask = torch.all(motif_mask, dim=-1, keepdim=True) * variant_mask
if 'variant_task_mask' in batch:
motif_mask = batch['variant_task_mask'] * rearrange(
(variant_label > self.label_threshold) & variant_label_mask, '... m t -> ... m () t'
)
else:
motif_mask = ((variant_label > self.label_threshold) &
variant_label_mask) | (~variant_label_mask)
motif_mask = torch.all(motif_mask, dim=-1, keepdim=True) * variant_mask
motif_mask = motif_mask[..., None]
avg_error_motif = functional.masked_mean(value=errors, mask=motif_mask)
logger.info('FitnessHead.motifs.loss: %s', avg_error_motif)
avg_error_motif = self.alpha * avg_error_motif
Expand All @@ -1494,22 +1502,27 @@ def loss(self, value, batch):

if self.num_var_as_ref > 0:
if num_var_as_ref > 0:
b, _, n = logits.shape
*b, m, n, t = logits.shape
# sample reference based on variant_label
ref_idx = torch.multinomial(
variant_weight + 1e-3, num_var_as_ref, replacement=False
ref_idx = torch.reshape(
torch.multinomial(
torch.reshape(variant_weight + 1e-3, (math.prod(b + [m]) // m, m)),
num_var_as_ref,
replacement=False
),
b + [num_var_as_ref]
)
ref_idx = torch.cat(
(torch.zeros((b, 1), dtype=ref_idx.dtype, device=logits.device), ref_idx),
(torch.zeros(*b, 1, dtype=ref_idx.dtype, device=logits.device), ref_idx),
dim=-1
)
logger.debug('FitnessHead.ref_idx: %s', ref_idx)

logits_ref = torch.gather(
logits, 1, repeat(ref_idx, '... m -> ... m i', i=logits.shape[-1])
logits, -3, repeat(ref_idx, '... m -> ... m i t', i=n, t=t)
)
mask_ref = torch.gather(
variant_mask, 1,
variant_mask, -3,
repeat(
ref_idx,
'... m -> ... m i t',
Expand All @@ -1518,25 +1531,25 @@ def loss(self, value, batch):
)
)
label_ref = torch.gather(
variant_label, 1, repeat(ref_idx, '... m -> ... m t', t=self.task_num)
variant_label, -3, repeat(ref_idx, '... m -> ... m t', t=self.task_num)
)
label_mask_ref = torch.gather(
variant_label_mask, 1, repeat(ref_idx, '... m -> ... m t', t=self.task_num)
variant_label_mask, -3, repeat(ref_idx, '... m -> ... m t', t=self.task_num)
)
else:
variant_mask = rearrange(torch.zeros_like(batch['mask']), 'b i -> b () i ()')
b = variant_mask.shape[0]
variant_label = torch.ones((b, 1, self.task_num), device=variant_mask.device)
variant_label_mask = torch.zeros(b, 1, self.task_num, device=variant_mask.device)
variant_mask = rearrange(torch.zeros_like(batch['mask']), '... i -> ... () i ()')
b = variant_mask.shape[:-2]
variant_label = torch.ones(*b, self.task_num, device=variant_mask.device)
variant_label_mask = torch.zeros(*b, self.task_num, device=variant_mask.device)

if not exists(logits_ref):
logits_ref = logits[:, :1, ...]
mask_ref = variant_mask[:, :1, ...]
label_ref = variant_label[:, :1, ...]
label_mask_ref = variant_label_mask[:, :1, ...]
logits_ref = logits[..., :1, :, :]
mask_ref = variant_mask[..., :1, :, :]
label_ref = variant_label[..., :1, :]
label_mask_ref = variant_label_mask[..., :1, :]

# pairwise logistic loss
variant_logit = logits[..., :, None, :] - logits_ref[..., None, :, :]
variant_logit = logits[..., :, None, :, :] - logits_ref[..., None, :, :, :]
variant_mask = variant_mask[..., :, None, :, :] * mask_ref[..., None, :, :, :]
variant_label = variant_label[..., :, None, :] - label_ref[..., None, :, :]
variant_label = torch.sign(variant_label
Expand Down
Loading