tbhugging/camera_orbit_compact_09_2b

What this model does

Given 6 camera-orbit views (0°, 60°, 120°, 180°, 240°, 300°) of 128×128 projections from the GummyBear Tomography dataset, the model predicts

• particle_x
• particle_y
• particle_z

coordinates of the embedded particle.

Architecture

Per-view GAP (pooled) CNN trunk → camera sin/cos geometry tokens → compact fusion MLP (e2e_pooled_geometry_fusion).

Final Report M9 Step 2 / 09_2B — compact head (f2_state).

• Hub: tbhugging/camera_orbit_compact_09_2b
• Companion dataset: tbhugging/gummybear-tomography
• Runnable report: GummyBearTomography_Final_Report.ipynb (M9 Step 2 — 09_2B pooled ladder)
• Source study checkpoint: checkpoints/m9/m09_e2e_pooled_geometry_fusion.pt (f2_state / m09_2_e2e_pooled_geometry_fusion)

Training configuration

• Protocol: 09_2B (pooled GAP; excludes 09_2A Fourier)
• Backbone: pooled_gap (pooled trunk)
• Input field: anomaly_ref
• Normalisation: per_image_zscore
• Camera orbit: 0°, 60°, 120°, 180°, 240°, 300° (6 views)
• Geometry: sin_theta, cos_theta (concat)
• Fusion: hidden 128, depth 1
• Targets: particle_x, particle_y, particle_z
• Builder: GeometryAwareFourierFusionLocalizer.for_09_2_pooled()
• Variant: m09_2_e2e_pooled_geometry_fusion
• Trainable parameters: 132358
• Stage-A learning rate: 0.001

Evaluation Results

Structured scores for the Hub widget are declared in the YAML model-index / metrics metadata (Model Cards — Evaluation Results).

Testing Data

• Dataset: tbhugging/gummybear-tomography (config m8_1)
• Splits: validation, test
• Protocol: Final Report M9 Step 2 / 09_2B (camera orbit 0°, 60°, 120°, 180°, 240°, 300°, anomaly_ref, per_image_zscore, GAP pooled trunk only, compact geometry fusion; excludes 09_2A Fourier)
• Source checkpoint: checkpoints/m9/m09_e2e_pooled_geometry_fusion.pt (f2_state only)

Metrics

Reported error is Euclidean RMSE over particle (x,y,z): d_i = ||pred_i - y_i||_2, then RMSE_total = sqrt(mean_i d_i^2). Hub metric id: rmse (display name RMSE_total (Euclidean xyz)).

Results

Scores match the Final Report M9 Step 2 09_2B pooled GAP bar (not 09_2A Fourier; not element-wise MSE).

Source: Final Report M9 Step 2 / 09_2B.

Load

import torch
# libraries from https://github.com/tbgitoo/gummybear-tomography
# Historical class name — use .for_09_2_pooled() for this GAP checkpoint only.
from tomography_ml.localization.localize_multiview import (
    GeometryAwareFourierFusionLocalizer,
)

n_views = 6
view_angles_deg = [0.0, 60.0, 120.0, 180.0, 240.0, 300.0]
model = GeometryAwareFourierFusionLocalizer.for_09_2_pooled(
    n_views=n_views,
    view_angles_deg=view_angles_deg,
)
views = torch.zeros(1, n_views, 1, 128, 128)
model(views)  # materialise lazy layers
state = torch.load('pytorch_model.bin', map_location='cpu', weights_only=True)
model.load_state_dict(state)
model.eval()
xyz = model(views)

Input tensor shape: [batch, n_views, channels, height, width]. Do not load with .for_09_2() (that builds the 09_2A Fourier trunk). Also excludes single-view M8, 09_3 large fusion, and M10 illumination stacks.

Inference

For an example with worked download, model instanciation and inference, see: 11_2_test_camera_orbit_compact_09_2b.ipynb in the github.com/tbgitoo/gummybear-tomography repository.