How to use
1) Fine-tune on your dataset
python finetune_wms.py \
  --data_root /path/to/WMS_Cleaned \
  --resume train_log \
  --save_dir train_log_wms \
  --epochs 10 \
  --batch_size 8 \
  --crop 256 \
  --stride 1


Optional (adds motion-aware weighted L1 on top of RIFE’s loss if the optimizer is exposed by your Model wrapper):

  --use_motion_weight --motion_lambda 0.5

2) Inference on a pair (get the mid frame)
python infer_wms_pair.py \
  --img0 /path/to/folder/greyscale_rgb_cyl_20250825_1340.png \
  --img1 /path/to/folder/greyscale_rgb_cyl_20250825_1350.png \
  --model_dir train_log_wms \
  --out out_mid.png

3) Inference on a whole folder (generate frames or an MP4)
python infer_wms_sequence.py \
  --folder /path/to/folder \
  --model_dir train_log_wms \
  --exp 2 \
  --out_dir wms_interp_frames \
  --out_video wms_interp.mp4 \
  --fps 30


--exp 1 inserts 1 midframe between each pair (2× fps).

--exp 2 inserts 3 midframes (4× fps), etc.

Notes that match your setup

No renaming needed. The dataset loader recursively scans everything under your root (e.g., CRT_*, ETC_*, VIIRS, HS_*, GFS_*) and builds (I0, It, I1) triplets just by sorted filenames.

The val split is automatic and deterministic (hash of the middle file).

Works with RGB/greyscale; greyscale is converted to RGB for the model.

Models are saved to train_log_wms (same RIFE loader pattern as your inference_* scripts).



What it does

Recursively scans your dataset folders (no renaming needed), builds (I0, It, I1) triplets by natural filename order.

Loads two models: --model_a (baseline, e.g. train_log) and --model_b (fine-tuned, e.g. train_log_wms).

Computes PSNR, SSIM, and optional LPIPS (if installed, use --lpips).

Writes per-triplet metrics to a CSV, plus:

Paired t-test and Cohen’s d for each metric (B vs A).

Boxplots (matplotlib only, no seaborn) for PSNR/SSIM (and LPIPS if enabled).

Example usage
python eval_wms.py \
  --data_root /path/to/WMS_Cleaned \
  --stride 1 \
  --model_a train_log \
  --model_b train_log_wms \
  --out_dir wms_eval_outputs


With LPIPS (if you have lpips installed):

python eval_wms.py \
  --data_root /path/to/WMS_Cleaned \
  --model_a train_log \
  --model_b train_log_wms \
  --lpips \
  --out_dir wms_eval_outputs

Outputs

Inside wms_eval_outputs/:

wms_metrics.csv — per-triplet PSNR/SSIM (/LPIPS) for both models.

stats.txt — mean±std, Δ, paired t-test p-values, Cohen’s d for each metric.

box_psnr.png, box_ssim.png (and box_lpips.png if enabled).

If you want, I can also add a cloud-region evaluation (compute SSIM only inside cloud masks) or a quick qualitative triplet dump (baseline vs fine-tuned vs GT image grids) for the paper figures.