add vlm_models

Author: aparnabg

src/sailsprep/vlm_models_constraint_activity/Constraint/README.md

@@ -0,0 +1,90 @@
+# V-JEPA2 Constraint Prediction with Temporal Windowing
+
+Complete pipeline for training V-JEPA2 to predict child constraint status from videos using temporal windowing and majority voting.
+
+## Overview
+
+This package predicts:
+1. **Constraint Status**: y (constrained) / n (not constrained) / partial
+2. **Constraint Type**: highchair, carseat, stroller, etc. (only if constrained)
+
+**Method**: Temporal windowing with majority voting
+- Divides video into N windows
+- Samples frames from each window
+- Gets prediction for each window  
+- Final prediction = majority vote across windows
+
+### Prerequisites
+- Python 3.10+
+- Your CSV file with columns: `BidsProcessed`, `Child_constrained`, `Constraint_type`
+
+
+
+### **Step 1: Prepare Data**
+
+```bash
+python prepare_data.py \
+    --csv your_data.csv \
+    --video_col BidsProcessed \
+    --constrained_col Child_constrained \
+    --type_col Constraint_type \
+    --val_split 0.3 \
+    --test_split 0.3 \
+    --output prepared_data.pkl
+```
+
+**Output:**
+- `prepared_data.pkl` with train/val/test splits 
+
+---
+
+### **Step 2: Train Model**
+
+#### **Option A: Without Windowing (Faster)**
+```bash
+python finetune_vjepa_windowing.py \
+    --data prepared_data.pkl \
+    --model_cache ./models \
+    --output ./finetuned_vjepa \
+    --epochs 5 \
+    --batch_size 32 \
+    --lr 1e-4 \
+    --num_workers 8
+```
+
+#### **Option B: With Windowing (More Accurate)**
+```bash
+python finetune_vjepa_windowing.py \
+    --data prepared_data.pkl \
+    --model_cache ./models \
+    --output ./finetuned_vjepa \
+    --epochs 5 \
+    --batch_size 16 \
+    --lr 1e-4 \
+    --num_workers 8 \
+    --use_windowing \
+    --num_windows 3
+```
+
+**Output:**
+- `./finetuned_vjepa/best_model.pt` - Trained model
+- `./finetuned_vjepa/training_history.json` - Training metrics
+
+---
+
+### **Step 3: Evaluate on Test Set**
+
+```bash
+python evaluate_vjepa_windowing.py \
+    --data prepared_data.pkl \
+    --model ./finetuned_vjepa \
+    --split test \
+    --output ./test_results
+```
+
+**Output:**
+- `./test_results/test_predictions.csv` - All predictions
+- `./test_results/test_metrics.json` - Accuracy, F1, confusion matrix
+
+---
+