# Checkpoints (`ckpts/`)

Training artifacts — weights, TensorBoard logs and evaluation results — for
**EchMo** and the comparison models on the two radar echo segmentation datasets:

- **Syn-RED** — synthetic echoes, 2 classes (background / target)
- **HiL-RED** — hardware-in-the-loop jamming echoes, 7 annotated classes
  (networks are trained with 9 output channels)

## Directory layout

```
ckpts/<dataset>/<run>/
├── config.json                        # training configuration (model, nb_classes, hyper-parameters)
├── boards/                            # TensorBoard logs (full training history)
└── results/
    ├── <model>.pt                     # exported plain weights (state_dict; ⚠️ see the epoch-291 note below)
    ├── <model>_latest_checkpoint.pt   # end-of-training full checkpoint (incl. optimizer, ~3x size; some runs)
    ├── <model>_results.json           # evaluation of the exported weights (epoch + val/test metrics)
    └── <model>_checkpoint_epoch<N>.pt # intermediate checkpoints kept by a few variant runs
```

- Run naming: `<model>_e<epochs>_lr<lr>_<loss>_<id>`; EchMo additionally has
  `ReCK-*` re-evaluation/ablation runs and the `dagger` / `deep` / `4moe` /
  `memo` architecture variants.
- `echmo_4moe_e300_lr0.zip` is a packed archive of the `echmo_4moe_e300_lr0/` run.

## A note on the epoch-291 exports

In the seven runs below — all on HiL-RED — the exported `<model>.pt` weights
were mistakenly saved from **epoch 291** instead of the best epoch on the test
curve:

- `deeplabv3plus_e300_lr0.0001_lsf_wce_4`
- `unet_e300_lr0.0001_lsf_wce_0`
- `vit_e300_lr0.0001_lsf_wce_0`
- `fcn8s_e300_lr0.0001_lsf_wce_2`
- `echmo_e300_lr0.0001_lsf_wce_18_ReCK-022`
- `echmo_e300_lr0.0001_lsf_wce_19_ReCK-202`
- `echmo_e300_lr0.0001_lsf_wce_20_ReCK-220`

The true best-epoch weights are **not** included in these directories
(`<model>_latest_checkpoint.pt` is the end-of-training checkpoint and is
likewise not the best).

**What this affects — and what it does not.** All results reported in the paper
are the best-epoch test-set values recorded in the TensorBoard training logs
and are **not** affected: the complete logs are shipped in `boards/`, so every
reported number can be re-verified (`tensorboard --logdir
ckpts/<dataset>/<run>/boards`, scalar `Test/Range_Doppler_metrics/MIoU`). What
is affected are only the exported weight files themselves: loading these runs'
`<model>.pt` reproduces metrics below the logged best.