Skip to content

Assignment inference improvment #3

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
tihsu99 wants to merge 11 commits into
base: main
Choose a base branch
from
Open

Assignment inference improvment #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
8ffbfaf
add project configuration files and update invisible padding logic
tihsu99 Jun 12, 2026
51500a4
refactor mask_jet function to handle valid indices and improve data a…
tihsu99 Jun 12, 2026
a2c8265
refactor assignment logic to incorporate scaled assignments and impro…
tihsu99 Jun 12, 2026
d7d24dd
fix bug
Jun 15, 2026
af1a2a7
add 2D score metrics and confusion detection plots for improved analy…
tihsu99 Jun 23, 2026
a679673
Merge remote-tracking branch 'origin/tihsu_assignment' into tihsu_ass…
tihsu99 Jun 23, 2026
b221e1a
refactor score metrics and confusion detection to use int64 for impro…
tihsu99 Jun 23, 2026
9965db7
refactor score metrics and confusion detection to use int64 for impro…
tihsu99 Jun 23, 2026
8e068b4
refactor detection metrics and visualization to enhance summary and p…
tihsu99 Jun 23, 2026
01221e9
refactor detection metrics and visualization to enhance summary and p…
tihsu99 Jun 23, 2026
927ed5a
refactor detection metrics and visualization to enhance summary and p…
tihsu99 Jun 23, 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
2 changes: 1 addition & 1 deletion network/evenet_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -386,7 +386,7 @@ def forward(
if 'x_invisible' in x and self.include_neutrino_generation:
invisible_point_cloud = x['x_invisible']
else:
invisible_point_cloud = torch.zeros(B, 1, self.invisible_input_dim, device=input_point_cloud.device)
invisible_point_cloud = torch.zeros(B, 1, self.sequential_input_dim, device=input_point_cloud.device)

invisible_point_cloud_mask = x['x_invisible_mask'].unsqueeze(
-1) if 'x_invisible_mask' in x else torch.zeros_like(input_point_cloud_mask[:, [0], :]).bool()
Expand Down
31 changes: 19 additions & 12 deletions network/loss/assignment.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -293,6 +293,16 @@ def loss_single_process(
## Detection Loss ##
####################

base_particle_balance_weight = torch.ones_like(targets_mask[0], dtype=torch.float32)
masks_for_balance = torch.stack(targets_mask).int()

if particle_index_tensor is not None:
class_indices = (masks_for_balance * particle_index_tensor.to(masks_for_balance.device).unsqueeze(1)).sum(0).int()
base_particle_balance_weight = (
base_particle_balance_weight *
particle_weights_tensor.to(masks_for_balance.device)[class_indices]
)

detections = detections
detections_target = targets_mask
detection_losses = []
Expand All @@ -301,7 +311,7 @@ def loss_single_process(
for symmetry_group in event_permutations:
for symmetry_element in symmetry_group:
symmetry_element = np.array(symmetry_element)
detection = detections[symmetry_element[0]]
detection = detections[symmetry_element[0]] # All the symmetry group detection/assignment outputs are just duplicated
detection_target = torch.stack([detections_target[symmetry_index] for symmetry_index in symmetry_element])
detection_target = detection_target.sum(0).long()

Expand All @@ -322,13 +332,15 @@ def loss_single_process(

process_masking = torch.stack(process_masking).float()
process_weighting = torch.stack(process_weighting).float()
detection_particle_balance_weight = base_particle_balance_weight.unsqueeze(0)

if event_weight is not None:
detection_losses = torch.stack(detection_losses) * process_masking * process_weighting * event_weight.view(-1, *([1] * (process_masking.dim() - 1)))
valid_process = torch.sum(process_masking * process_weighting * event_weight.view(-1, *([1] * (process_masking.dim() - 1))))
event_balance_weight = event_weight.view(-1, *([1] * (process_masking.dim() - 1)))
detection_losses = torch.stack(detection_losses) * process_masking * process_weighting * event_balance_weight * detection_particle_balance_weight
valid_process = torch.sum(process_masking * process_weighting * event_balance_weight * detection_particle_balance_weight)
else:
detection_losses = torch.stack(detection_losses) * process_masking * process_weighting
valid_process = torch.sum(process_masking * process_weighting)
detection_losses = torch.stack(detection_losses) * process_masking * process_weighting * detection_particle_balance_weight
valid_process = torch.sum(process_masking * process_weighting * detection_particle_balance_weight)

if valid_process > 0:
detection_loss = torch.sum(detection_losses) / valid_process
Expand All @@ -351,15 +363,10 @@ def loss_single_process(
focal_gamma
)

particle_balance_weight = torch.ones_like(symmetric_losses)
masks_for_balance = torch.stack(targets_mask).int()

if particle_balance_weight is not None and particle_index_tensor is not None:
class_indices = (masks_for_balance * particle_index_tensor.to(masks_for_balance.device).unsqueeze(1)).sum(0).int()
particle_balance_weight *= particle_weights_tensor.to(masks_for_balance.device)[class_indices]
particle_balance_weight = base_particle_balance_weight.to(symmetric_losses.device)

if process_weight[0] is not None:
particle_balance_weight *= process_weight[0].unsqueeze(0)
particle_balance_weight = particle_balance_weight * process_weight[0]

targets_mask_finite = torch.stack(targets_mask).float()
if not torch.isfinite(symmetric_losses).all():
Expand Down
Loading