tbhugging/singleview_cnn_fourier

What this model does

Given a single 128×128 projection image from the M8 GummyBear Tomography dataset, the model predicts

• particle_x
• particle_y
• particle_z

coordinates of the embedded particle.

Architecture

The model uses Fourier pooling in place of conventional max-pooling layers as described in the M8 study (see https://github.com/tbgitoo/gummybear-tomography).

Single-view CNN + Fourier pooling particle localiser from the gummybear-tomography Final Report M8 Step 3 (train → validation/test on (particle_x, particle_y, particle_z)).

• Hub: tbhugging/singleview_cnn_fourier
• Companion dataset: tbhugging/gummybear-tomography
• Runnable report: GummyBearTomography_Final_Report.ipynb (M8 Step 3)
• Source study checkpoint: checkpoints/m8/m08_train_val_test_xyz.pt (arch fourier only)

Training configuration

• Input field: anomaly_ref
• Normalisation: per_image_zscore
• Camera: 180.0° (single view)
• Targets: particle_x, particle_y, particle_z
• Library class: LocalizerSingleViewFourier
• Trainable parameters: 31811
• Stage learning rate: 0.03

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 M8 Step 3 (single-view 180°, anomaly_ref, per_image_zscore, targets particle_x, particle_y, particle_z)
• Source checkpoint: checkpoints/m8/m08_train_val_test_xyz.pt (arch fourier 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). This matches the Final Report M8 Step 3 bars (not element-wise MSE). Hub metric id: rmse (display name RMSE_total (Euclidean xyz)).

Results

Source: Final Report M8 Step 3.

Load

import torch
# libraries from https://github.com/tbgitoo/gummybear-tomography
from tomography_ml.studies.single_view_m8 import make_m8_single_view_model
from tomography_ml.localization.builders import materialize_lazy_modules

model = make_m8_single_view_model('fourier', n_outputs=3, device='cpu')
materialize_lazy_modules(model, torch.zeros(1, 1, 128, 128))
state = torch.load('pytorch_model.bin', map_location='cpu', weights_only=True)
model.load_state_dict(state)
model.eval()

Do not treat this checkpoint as multi-view (M9) or multi-illumination (M10).

Inference

For an example with worked download, model instanciation and inference, see: 11_1_test_singleview_cnn_fourier.ipynb in the gummybear-tomography repository.