Corrected HRNet 300W code, v2

Main architectural changes

1. config.py is now the single source of truth for paths and experiment settings.

   • Dataset path
   • Checkpoint paths
   • Output directories
   • Training hyperparameters
   • Evaluation thresholds
   • NME normalization mode
   • Alignment template and output size

2. DeepLake300W.__getitem__() now returns three values:

img, heatmaps, pts

pts is the true landmark coordinate after crop and resize. Evaluation should use pts directly as GT.

3. evaluate.py no longer decodes GT points from GT heatmaps.

Old behavior:

gt_pts = heatmaps_to_pts(gt_hm, IMG_SIZE)

Corrected behavior:

for imgs, _gt_hm, gt_pts in loader:
    ...

4. evaluate.py now reports:

   • Mean NME
   • Median NME
   • Failure rate
   • AUC
   • Per-region pixel error
   • Per-region normalized error

5. predict_align.py now reads real GT landmarks directly and adds quantitative alignment metrics.

Important config entries

Change these in config.py instead of editing scripts:

EVAL_CHECKPOINT_PATH = CHECKPOINT_DIR / "hrnet_epoch_200.pth"
EVAL_OUT_DIR = TEST_DIR / "full_eval_epoch_200_fixed"
NME_NORM_TYPE = "inter_ocular"
FAILURE_THRESHOLD = 0.08
AUC_THRESHOLD = 0.08

ALIGN_CHECKPOINT_PATH = EVAL_CHECKPOINT_PATH
ALIGN_OUT_DIR = TEST_DIR / "single_predict_fixed"
ALIGN_OUTPUT_SIZE = 112

How to run

python train.py
python evaluate.py
python predict_align.py

If another script still has:

img, heatmaps = dataset[index]

change it to:

img, heatmaps, pts = dataset[index]

or:

img, heatmaps, _ = dataset[index]