pikaliov 9200772bea Fix NaN loss: revert MONA to fp32, fix loss logging ...

- MONA fp16 causes NaN (gamma=1e-6 underflows in fp16 min subnormal ~6e-8)
- Revert MONA forward to fp32 with autocast(enabled=False), cast output back
- Fix loss CSV: save raw_loss before backward() (tensor consumed after backward)
- Verified: loss=3.78, no NaN, bs=48 peak=21.4 GB

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-21 22:09:49 +03:00

..

datasets

Remove projections (1024 native), add satellite text, dual GatedFusion

2026-04-21 19:01:30 +03:00

eval

Rewrite: GatedFusion architecture + UAV-GeoLoc dataset

2026-04-17 17:13:00 +03:00

losses

Remove projections (1024 native), add satellite text, dual GatedFusion

2026-04-21 19:01:30 +03:00

models

Fix NaN loss: revert MONA to fp32, fix loss logging

2026-04-21 22:09:49 +03:00

training

Fix NaN loss: revert MONA to fp32, fix loss logging

2026-04-21 22:09:49 +03:00

__init__.py

Initial commit: caption quality test on UAV-VisLoc

2026-04-17 00:04:46 +03:00