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

Jump to bottom

Implement break-down mode for cross entropy calculation #1257

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Implement break-down mode for cross entropy calculation #1257

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e1fc5d3
change operation order in cross-entropy preprocess
penkoba Apr 18, 2022
fba0cd7
implement break-down mode for cross-entropy calculation
penkoba Apr 18, 2022
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
30 changes: 24 additions & 6 deletions yolox/models/yolo_head.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -487,12 +487,30 @@ def get_assignments(

with torch.cuda.amp.autocast(enabled=False):
cls_preds_ = (
cls_preds_.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
* obj_preds_.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
)
pair_wise_cls_loss = F.binary_cross_entropy(
cls_preds_.sqrt_(), gt_cls_per_image, reduction="none"
).sum(-1)
cls_preds_.float().unsqueeze(0).sigmoid_()
* obj_preds_.float().unsqueeze(0).sigmoid_()
).sqrt_()

try:
# may fail to allocate GPU memory when num_gt is large
pair_wise_cls_loss = F.binary_cross_entropy(
cls_preds_.repeat(num_gt, 1, 1), gt_cls_per_image,
reduction="none"
).sum(-1)
except RuntimeError as e:
# TODO: the string might change, consider a better way
if mode == "cpu" or "CUDA out of memory. " not in str(e):
raise # RuntimeError might not caused by CUDA OOM
# to work with less GPU memory
print('--- break-down mode for cross-entropy calculations ---')
pair_wise_cls_loss = torch.empty(
(num_gt, torch.sum(fg_mask)),
dtype=cls_preds_.dtype, device=cls_preds_.device)
for i in range(num_gt):
pair_wise_cls_loss[i] = F.binary_cross_entropy(
cls_preds_, gt_cls_per_image[i].unsqueeze(0),
reduction="none"
).sum(-1)
del cls_preds_

cost = (
Expand Down