Skip to content

Reduce augmentation probabilities to fix underfitting #19

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

Open
sridhs21 wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Reduce augmentation probabilities to fix underfitting #19

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
665fb2b
Reduce augmentation probabilities to fix underfitting
sridhs21 Nov 19, 2025
9fbdff2
Fix augmentation bug: apply brightness/contrast globally to preserve …
sridhs21 Dec 2, 2025
7ce3e14
reverting percentages back to original for testing purposes
sridhs21 Dec 25, 2025
56c7448
Fixed issue where on-the-fly augmentation was mutating caches base fr…
sridhs21 Jan 27, 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
21 changes: 12 additions & 9 deletions XPointMLTest.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -404,18 +404,19 @@ def _apply_augmentation(self, all_data, mask):
noise = torch.randn_like(all_data) * noise_std
all_data = all_data + noise

# 5. Random brightness/contrast adjustment per channel (30% chance)
# Helps model become invariant to intensity variations
# 5. Random brightness/contrast adjustment (30% chance)
# CHANGED: Applied globally across channels to preserve physical relationships
# (e.g., keeping the derivative relationship between psi and B fields)
if self.rng.random() < 0.3:
for c in range(all_data.shape[0]):
brightness = self.rng.uniform(-0.1, 0.1)
contrast = self.rng.uniform(0.9, 1.1)
mean = all_data[c].mean()
all_data[c] = contrast * (all_data[c] - mean) + mean + brightness
brightness = self.rng.uniform(-0.1, 0.1)
contrast = self.rng.uniform(0.9, 1.1)
# Apply same transformation to all channels
mean = all_data.mean(dim=(-2, -1), keepdim=True)
all_data = contrast * (all_data - mean) + mean + brightness

# 6. Cutout/Random erasing (20% chance)
# Prevents model from relying too heavily on specific spatial features
if self.rng.random() < 0.2:
if self.rng.random() < 0.20:
h, w = all_data.shape[-2:]
cutout_size = int(min(h, w) * self.rng.uniform(0.1, 0.25))
if cutout_size > 0:
Expand Down Expand Up @@ -451,7 +452,9 @@ def __getitem__(self, _):
want_pos = (attempt / self.retries) < self.pos_ratio

if has_pos == want_pos or attempt == self.retries - 1:
all_crop = self._crop(frame["all"], y0, x0)
# Clone to avoid in-place augmentation modifying cached base frames
all_crop = self._crop(frame["all"], y0, x0).clone()
crop_mask = crop_mask.clone()

# Apply augmentation if enabled
all_crop, crop_mask = self._apply_augmentation(all_crop, crop_mask)
Expand Down