Shape2Force / experiments /README.md
kaveh's picture
S2F v1.2.0
d2a9154
|
Raw History Blame Contribute Delete
3.16 kB
# Baselines and ablations
Code for the evaluation of the released checkpoints and for the baseline and ablation models.
Run every command from the repository root; a GPU is used when available.
```
experiments/
layout.py paths and split-manifest helpers
manifest_data.py data loader driven by a split manifest
ablation_models.py U-Net and S2F generator variants
losses.py reconstruction losses
configs.py training protocol and variants
train.py train one variant
evaluate.py evaluate checkpoints and trained runs on all manifests
run_all.py run the steps in order
splits/ split manifests
results/eval/ evaluation tables
```
## Data
The image pairs are available on Zenodo (doi: 10.5281/zenodo.18457879). Place them in `dataset/` or set
`S2F_DATASET_ROOT`. Each manifest in `splits/` lists one sample per row (folder relative to the dataset root,
condition, Young's modulus, pixel-size setting, colour-scale limit and substrate setting):
- `alltune_train` (280) and `alltune_val` (81): single-cell training and validation sets;
- `test_fibro_*`: independent single-cell test sets;
- `spheroid_train` (129) and `spheroid_val` (28).
## Usage
```bash
python -m experiments.run_all --steps eval_released
python -m experiments.run_all --steps train,eval_runs --variants s2f_no_substrate,unet_no_substrate --seeds 0,1
python -m experiments.train --variant unet --seed 0
```
Runs are written to `experiments/runs/<variant>_seed<seed>/` and resume when restarted. Add
`--epochs 1 --max-batches 2` for a quick test.
## Training protocol
All variants use the training and validation sets above, batch size 2, 1024 × 1024 px, the augmentations of
`data/augmentations.py` and Adam (β1 = 0.5, β2 = 0.999) with a cosine schedule and 10 % warm-up per phase.
| phase | epochs | reconstruction loss | other terms | lr |
|---|---|---|---|---|
| 1 | 100 | 100 × weighted MSE + soft Dice | 1 × adversarial | 1e-4 |
| 2 | 40 | 10 × RME | 1 × adversarial, 100 × traction-stress consistency | 1e-5 |
Phase 2 starts from the best phase-1 generator. In each phase the checkpoint with the lowest validation loss is
kept; the final model is the best phase-2 checkpoint.
## Variants reported in the paper
| model | variant |
|---|---|
| U-Net | `unet_no_substrate` |
| pix2pix | `unet_gan_no_substrate` |
| S2F | `s2f_no_substrate` |
| S2F (+ substrate) | `s2f_staged` |
| without multi-scale input module | `s2f_no_substrate_no_multiscale` |
| without bottleneck context and attention modules | `s2f_no_substrate_bridge_plain` |
| without attention gates | `s2f_no_substrate_no_attention_gates` |
| without discriminator | `s2f_no_substrate_no_gan` |
| without traction-stress consistency loss | `s2f_staged_no_force` |
The input-free baseline (average training map) is evaluated with `python -m experiments.evaluate --meanmap`.
Other variants are listed in `configs.py`.
## Outputs
`evaluate.py` computes per-image MSE, MS-SSIM, pixel-wise Pearson r, RME and mean traction stress (Pa), and writes
`results/eval/per_image.csv`, `summary.csv` and `model_params.csv`.