fix: resolve ruff lint warnings in vjepa scripts

Author: aparnabg

src/sailsprep/action_model_testing/vjepa/clips_fixed_length/vjepa_clip_level_ablation.py

@@ -1,26 +1,13 @@
 """
 VJEPA2 Clip-Level Classification — Head Ablation
 =================================================
-Same clip-level pipeline as vjepa_clip_level.py (SlowFast-style chunking,
-no NA class), but adds --head to ablate five classification heads:
 
   linear       : mean-pool tokens -> single Linear layer
   mlp_small    : mean-pool -> LayerNorm -> 512-dim MLP -> classifier
   mlp_large    : mean-pool -> LayerNorm -> 1024 -> 512-dim MLP -> classifier
   attentive    : cross-attention query token -> Linear  (original)
   transformer  : 2-layer Transformer encoder -> CLS token -> Linear
 
-Features are extracted ONCE per (label, run) and cached to disk.
-Each head is then trained & evaluated independently.
-
-Clipping rules (mirrors SlowFast exactly):
-  < 15 frames              -> skip
-  15-44 frames             -> 1 clip
-  45-59 frames             -> 2 clips
-  >= 60 frames             -> 30-frame chunks (last kept if >= 15 frames)
-
-NA frames break a run. NA is NOT a class.
-
 Usage:
     # Single head
     python vjepa_clip_level_ablation.py --label loco --head attentive
@@ -32,8 +19,6 @@
     # Skip encoder inference if features already cached
     python vjepa_clip_level_ablation.py --label loco --head all --skip_extraction
 
-CSV columns expected:
-    video_path, label_path, interpolated_anno_h5, split
 """
 
 import argparse
@@ -530,7 +515,7 @@ def run_inference(probe, te_feats, test_samples, label_map, device,
     rows   = []
     idx    = 0
 
-    for feats, labels in loader:
+    for feats, _labels in loader:
         probs = softmax(probe(feats.to(device)))
         for i in range(probs.shape[0]):
             top  = int(probs[i].argmax().item())
@@ -642,7 +627,8 @@ def main():
     labels = sorted({s["label_str"] for s in all_s})
     lmap   = {lab: i for i, lab in enumerate(labels)}
     print(f"\nLabel map: {lmap}")
-    json.dump(lmap, open(os.path.join(base_dir, "label_mapping.json"), "w"), indent=2)
+    with open(os.path.join(base_dir, "label_mapping.json"), "w") as f:
+        json.dump(lmap, f, indent=2)
     pd.DataFrame(test_s).to_csv(os.path.join(base_dir, "test_split.csv"), index=False)
 
     # Inverse-frequency class weights from train split

src/sailsprep/action_model_testing/vjepa/clips_without_coi_crop/extract_features.py

@@ -20,8 +20,8 @@
 
 try:
     from decord import VideoReader, cpu
-except ImportError:
-    raise ImportError("Please install decord: pip install eva-decord")
+except ImportError as e:
+    raise ImportError("Please install decord: pip install eva-decord") from e
 
 # ============================================================
 # CONFIG

src/sailsprep/action_model_testing/vjepa/clips_without_coi_crop/rmm/extract_features.py

@@ -23,8 +23,8 @@
 
 try:
     from decord import VideoReader, cpu
-except ImportError:
-    raise ImportError("Please install decord: pip install eva-decord")
+except ImportError as e:
+    raise ImportError("Please install decord: pip install eva-decord") from e
 
 # ============================================================
 # CONFIG

src/sailsprep/action_model_testing/vjepa/clips_without_coi_crop/rmm/train_probe_ablation.py

@@ -392,8 +392,8 @@ def main():
     print(f"  Clips    : {len(labels)}",    flush=True)
     print(f"  Labels   : {label_map}",      flush=True)
 
-    json.dump(label_map,
-              open(os.path.join(seed_dir, "label_mapping.json"), "w"), indent=2)
+    with open(os.path.join(seed_dir, "label_mapping.json"), "w") as f:
+        json.dump(label_map, f, indent=2)
 
     # --- Train / test split (identical to original train_probe.py) ---
     indices = np.arange(len(labels))

src/sailsprep/action_model_testing/vjepa/clips_without_coi_crop/train_probe_ablation.py

@@ -298,7 +298,7 @@ def run_inference(probe, test_features, test_labels_enc, original_labels,
     loader  = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=False)
 
     results, idx = [], 0
-    for feats, labels in loader:
+    for feats, _labels in loader:
         probs = softmax(probe(feats.to(device)))
         for i in range(probs.shape[0]):
             top  = int(probs[i].argmax().item())
@@ -381,8 +381,8 @@ def main():
     print(f"  Clips    : {len(labels)}")
     print(f"  Labels   : {label_map}")
 
-    json.dump(label_map,
-              open(os.path.join(seed_dir, "label_mapping.json"), "w"), indent=2)
+    with open(os.path.join(seed_dir, "label_mapping.json"), "w") as f:
+        json.dump(label_map, f, indent=2)
 
     # --- Train / test split (same logic as train_probe.py) ---
     indices = np.arange(len(labels))

src/sailsprep/action_model_testing/vjepa/coi_crop/finetune_vjepa2_h5bbox.py

@@ -1,12 +1,6 @@
 """End-to-end V-JEPA 2 fine-tuning on H5-bbox-cropped action segments.
 
-Mirrors the SlowFast pipeline:
-  - Same split CSV (video_path, label_path, h5_file_path)
-  - Same on-the-fly bbox crop from interpolated H5
-  - Same action segment extraction (label runs >= MIN_RUN frames)
-  - Same stratified split, class weights, metrics
-
-Default mode: frozen encoder + attentive probe (recommended for small data).
+Default mode: frozen encoder + attentive probe.
 Use --full_finetune to unfreeze the encoder.
 """
 

src/sailsprep/action_model_testing/vjepa/full_video/train_probe_framelevel.py

@@ -303,7 +303,7 @@ def run_inference_per_video(probe, splits_csv, task_column,
         )
 
         rows = []
-        for windows, labels, lbl_strs, frame_idxs in vid_loader:
+        for windows, _labels, lbl_strs, frame_idxs in vid_loader:
             windows = windows.to(device)
             logits  = probe(windows)
             probs   = softmax(logits)

src/sailsprep/action_model_testing/vjepa/full_video/train_probe_framelevel_hierarchical.py

@@ -344,7 +344,7 @@ def run_inference_per_video(probe, splits_csv, task_column,
         )
 
         rows = []
-        for windows, s1_lbls, s2_lbls, lbl_strs, frame_idxs in vid_loader:
+        for windows, _s1_lbls, _s2_lbls, lbl_strs, frame_idxs in vid_loader:
             windows = windows.to(device)
             l1, l2  = probe(windows)
             p1      = softmax(l1)

src/sailsprep/action_model_testing/vjepa/full_video/two_stage.py

@@ -384,7 +384,7 @@ def run_inference(probe, test_windows_raw, stage1_map, stage2_map,
     loader  = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=False)
 
     rows = []
-    for feats, s1_lbls, s2_lbls, lbl_strs in loader:
+    for feats, _s1_lbls, _s2_lbls, lbl_strs in loader:
         feats = feats.to(device)
 
         logits1, logits2 = probe(feats)