Underdog

rewrite_training.py

@@ -1,237 +1,759 @@
 import os
-import polars as pl
+import json
+import math
 import numpy as np
+import polars as pl
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from torch.utils.data import Dataset, DataLoader, random_split
+import torch.optim as optim
-from concurrent.futures import ProcessPoolExecutor
+
+from torch.utils.data import Dataset, DataLoader
+from sklearn.model_selection import train_test_split
+from sklearn.preprocessing import LabelEncoder
+from multiprocessing import Pool, cpu_count
+from functools import partial
 from tqdm import tqdm
+from collections import Counter
 
-# --- CONFIGURATION ---
+# ===============================
-BASE_PATH = "asl_kaggle"
+# GPU CONFIGURATION
-TARGET_FRAMES = 22
+# ===============================
-LIPS = [61, 146, 91, 181, 84, 17, 314, 405, 321, 375, 291, 308, 324, 318, 402, 317, 14, 87, 178, 88, 95]
+print("=" * 60)
-HANDS = list(range(468, 543))
+print("GPU CONFIGURATION")
-SELECTED_INDICES = LIPS + HANDS
+print("=" * 60)
-NUM_FEATS = len(SELECTED_INDICES) * 3
 
-# Training hyperparameters
+if torch.cuda.is_available():
-BATCH_SIZE = 32
+    print(f"✓ CUDA available!")
-EPOCHS = 50
+    print(f"✓ GPU: {torch.cuda.get_device_name(0)}")
-LEARNING_RATE = 0.001
+    device = torch.device('cuda:0')
-TRAIN_SPLIT = 0.8
+    torch.backends.cudnn.benchmark = True
-CHECKPOINT_DIR = "checkpoints"
+    torch.backends.cudnn.enabled = True
+else:
+    print("✗ CUDA not available, using CPU")
+    device = torch.device('cpu')
 
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+print("=" * 60)
-print(f"Using device: {device}")
+
+# ===============================
+# SELECTED LANDMARK INDICES
+# ===============================
+IMPORTANT_FACE_INDICES = sorted(list(set([
+    0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
+    55, 65, 66, 105, 107, 336, 296, 334,
+    33, 133, 160, 159, 158, 144, 145, 153,
+    362, 263, 387, 386, 385, 373, 374, 380,
+    1, 2, 98, 327,
+    61, 185, 40, 39, 37, 0, 267, 269, 270, 409,
+    291, 146, 91, 181, 84, 17, 314, 405, 321, 375,
+    78, 191, 80, 81, 82, 13, 312, 311, 310, 415,
+    308, 324, 318, 402, 317, 14, 87, 178, 88, 95
+])))
+
+NUM_FACE_POINTS = len(IMPORTANT_FACE_INDICES)
+NUM_HAND_POINTS = 21 * 2
+TOTAL_POINTS_PER_FRAME = NUM_HAND_POINTS + NUM_FACE_POINTS
 
 
-# --- DATA PROCESSING ---
+# ===============================
-def load_kaggle_metadata(base_path):
+# DATA AUGMENTATION
-    return pl.read_csv(os.path.join(base_path, "train.csv"))
+# ===============================
+def augment_sequence(x, modality_mask):
+    """Apply random augmentations to training data"""
+    x = x.copy()
+
+    # Random temporal cropping (simulate different signing speeds)
+    if np.random.rand() < 0.3 and len(x) > 20:
+        start = np.random.randint(0, max(1, len(x) // 4))
+        x = x[start:]
+        modality_mask = modality_mask[start:]
+
+    # Random spatial scaling
+    if np.random.rand() < 0.5:
+        scale = np.random.uniform(0.85, 1.15)
+        x = x * scale
+
+    # Random rotation (around z-axis for x,y coordinates)
+    if np.random.rand() < 0.5:
+        angle = np.random.uniform(-0.3, 0.3)
+        cos_a, sin_a = np.cos(angle), np.sin(angle)
+
+        # Reshape to get xyz coordinates
+        x_reshaped = x.reshape(len(x), -1, 3)
+        x_rot = x_reshaped.copy()
+        x_rot[..., 0] = x_reshaped[..., 0] * cos_a - x_reshaped[..., 1] * sin_a
+        x_rot[..., 1] = x_reshaped[..., 0] * sin_a + x_reshaped[..., 1] * cos_a
+        x = x_rot.reshape(x.shape)
+
+    # Random masking (simulate occlusion) - only for some frames
+    if np.random.rand() < 0.3:
+        n_mask = int(len(x) * 0.15)  # mask 15% of frames
+        mask_indices = np.random.choice(len(x), n_mask, replace=False)
+        x[mask_indices] *= 0.1  # dim but don't completely zero
+
+    # Random noise
+    if np.random.rand() < 0.4:
+        noise = np.random.normal(0, 0.02, x.shape)
+        x = x + noise
+
+    # Random time warping (speed up or slow down)
+    if np.random.rand() < 0.3 and len(x) > 20:
+        speed = np.random.uniform(0.8, 1.2)
+        new_len = int(len(x) * speed)
+        new_len = min(new_len, len(x))
+        indices = np.linspace(0, len(x) - 1, new_len).astype(int)
+        x = x[indices]
+        modality_mask = modality_mask[indices]
+
+    return x, modality_mask
 
 
-def load_and_preprocess(path, base_path=BASE_PATH, target_frames=TARGET_FRAMES):
+# ===============================
-    parquet_path = os.path.join(base_path, path)
+# ENHANCED DATA EXTRACTION (POLARS)
-    df = pl.read_parquet(parquet_path)
+# ===============================
+def extract_multi_landmarks(path, min_valid_frames=3):
+    """
+    Extract both hands + selected face landmarks with modality flags
+    Returns: dict with 'landmarks', 'left_hand_valid', 'right_hand_valid', 'face_valid'
+    """
+    try:
+        df = pl.read_parquet(path)
+        seq = []
+        left_valid_frames = []
+        right_valid_frames = []
+        face_valid_frames = []
 
-    anchors = (
+        all_types = df.select("type").unique().to_series().to_list()
-        df.filter((pl.col("type") == "face") & (pl.col("landmark_index") == 0))
+        has_data = any(t in all_types for t in ["left_hand", "right_hand", "face"])
-        .select([pl.col("frame"), pl.col("x").alias("nx"), pl.col("y").alias("ny"), pl.col("z").alias("nz")])
-    )
 
-    processed = (
+        if not has_data:
-        df.join(anchors, on="frame", how="left")
+            return None
-        .with_columns([
-            (pl.col("x") - pl.col("nx")).fill_null(0.0),
-            (pl.col("y") - pl.col("ny")).fill_null(0.0),
-            (pl.col("z") - pl.col("nz")).fill_null(0.0),
-        ])
-        .sort(["frame", "type", "landmark_index"])
-    )
 
-    raw_tensor = processed.select(["x", "y", "z"]).to_numpy().reshape(-1, 543, 3)
+        frames = sorted(df.select("frame").unique().to_series().to_list())
-    reduced_tensor = raw_tensor[:, SELECTED_INDICES, :]
 
-    curr_len = reduced_tensor.shape[0]
+        if len(frames) < min_valid_frames:
-    indices = np.linspace(0, curr_len - 1, num=target_frames).round().astype(int)
+            return None
-    return reduced_tensor[indices]
+
+        for frame in frames:
+            frame_df = df.filter(pl.col("frame") == frame)
+            frame_points = np.full((TOTAL_POINTS_PER_FRAME, 3), np.nan, dtype=np.float32)
+
+            pos = 0
+            left_valid = False
+            right_valid = False
+            face_valid = False
+
+            # Left hand
+            left = frame_df.filter(pl.col("type") == "left_hand")
+            if left.height > 0:
+                valid_count = 0
+                for i in range(21):
+                    row = left.filter(pl.col("landmark_index") == i)
+                    if row.height > 0:
+                        coords = row.select(["x", "y", "z"]).row(0)
+                        if all(c is not None for c in coords):
+                            frame_points[pos] = coords
+                            valid_count += 1
+                    pos += 1
+                left_valid = (valid_count >= 10)
+            else:
+                pos += 21
+
+            # Right hand
+            right = frame_df.filter(pl.col("type") == "right_hand")
+            if right.height > 0:
+                valid_count = 0
+                for i in range(21):
+                    row = right.filter(pl.col("landmark_index") == i)
+                    if row.height > 0:
+                        coords = row.select(["x", "y", "z"]).row(0)
+                        if all(c is not None for c in coords):
+                            frame_points[pos] = coords
+                            valid_count += 1
+                    pos += 1
+                right_valid = (valid_count >= 10)
+            else:
+                pos += 21
+
+            # Face
+            face = frame_df.filter(pl.col("type") == "face")
+            if face.height > 0:
+                valid_count = 0
+                for idx in IMPORTANT_FACE_INDICES:
+                    row = face.filter(pl.col("landmark_index") == idx)
+                    if row.height > 0:
+                        coords = row.select(["x", "y", "z"]).row(0)
+                        if all(c is not None for c in coords):
+                            frame_points[pos] = coords
+                            valid_count += 1
+                    pos += 1
+                face_valid = (valid_count >= len(IMPORTANT_FACE_INDICES) * 0.3)
+
+            valid_ratio = 1 - np.isnan(frame_points).mean()
+            if valid_ratio >= 0.20:
+                frame_points = np.nan_to_num(frame_points, nan=0.0)
+                seq.append(frame_points)
+                left_valid_frames.append(left_valid)
+                right_valid_frames.append(right_valid)
+                face_valid_frames.append(face_valid)
+
+        if len(seq) < min_valid_frames:
+            return None
+
+        return {
+            'landmarks': np.stack(seq),
+            'left_hand_valid': np.array(left_valid_frames),
+            'right_hand_valid': np.array(right_valid_frames),
+            'face_valid': np.array(face_valid_frames)
+        }
+
+    except Exception as e:
+        return None
 
 
-# --- DATASET CLASS ---
+def get_features_sequence(landmarks_data, max_frames=100):
-class ASLDataset(Dataset):
+    """Enhanced feature extraction with separate modality processing"""
-    def __init__(self, tensors, labels):
+    if landmarks_data is None:
-        self.tensors = tensors
+        return None, None, None
-        self.labels = labels
 
-    def __len__(self):
+    landmarks_3d = landmarks_data['landmarks']
-        return len(self.tensors)
+    if len(landmarks_3d) == 0:
+        return None, None, None
 
-    def __getitem__(self, idx):
+    T, N, _ = landmarks_3d.shape
-        return self.tensors[idx], self.labels[idx]
+
+    # Separate modalities for independent normalization
+    left_hand = landmarks_3d[:, :21, :]
+    right_hand = landmarks_3d[:, 21:42, :]
+    face = landmarks_3d[:, 42:, :]
+
+    features_list = []
+
+    for modality, valid_mask in [
+        (left_hand, landmarks_data['left_hand_valid']),
+        (right_hand, landmarks_data['right_hand_valid']),
+        (face, landmarks_data['face_valid'])
+    ]:
+        valid_frames = modality[valid_mask] if valid_mask.any() else modality
+        if len(valid_frames) > 0:
+            center = np.mean(valid_frames, axis=(0, 1), keepdims=True)
+            spread = np.std(valid_frames, axis=(0, 1), keepdims=True).max()
+        else:
+            center = 0
+            spread = 1
+
+        modality_norm = (modality - center) / max(spread, 1e-6)
+        flat = modality_norm.reshape(T, -1)
+
+        # Deltas
+        deltas = np.zeros_like(flat)
+        if T > 1:
+            deltas[1:] = flat[1:] - flat[:-1]
+
+        features_list.append(flat)
+        features_list.append(deltas)
+
+    features = np.concatenate(features_list, axis=1)
+
+    modality_mask = np.stack([
+        landmarks_data['left_hand_valid'],
+        landmarks_data['right_hand_valid'],
+        landmarks_data['face_valid']
+    ], axis=1).astype(np.float32)
+
+    # Pad/truncate
+    if T < max_frames:
+        pad = np.zeros((max_frames - T, features.shape[1]), dtype=np.float32)
+        features = np.concatenate([features, pad], axis=0)
+
+        mask_pad = np.zeros((max_frames - T, 3), dtype=np.float32)
+        modality_mask = np.concatenate([modality_mask, mask_pad], axis=0)
+
+        frame_mask = np.zeros(max_frames, dtype=bool)
+        frame_mask[:T] = True
+    else:
+        features = features[:max_frames]
+        modality_mask = modality_mask[:max_frames]
+        frame_mask = np.ones(max_frames, dtype=bool)
+
+    features = np.nan_to_num(features, nan=0.0, posinf=0.0, neginf=0.0)
+    features = np.clip(features, -30, 30)
+
+    return features.astype(np.float32), frame_mask, modality_mask
 
 
-# --- MODEL ARCHITECTURE ---
+def process_row(row_data, base_path, max_frames=100):
-class ASLClassifier(nn.Module):
+    """Process a single row"""
-    def __init__(self, num_classes, target_frames=TARGET_FRAMES, num_feats=NUM_FEATS):
+    path_rel, sign = row_data
+    path = os.path.join(base_path, path_rel)
+    if not os.path.exists(path):
+        return None, None, None, None
+
+    try:
+        lm_data = extract_multi_landmarks(path)
+        if lm_data is None:
+            return None, None, None, None
+
+        feat, frame_mask, modality_mask = get_features_sequence(lm_data, max_frames)
+        if feat is None:
+            return None, None, None, None
+
+        return feat, frame_mask, modality_mask, sign
+
+    except Exception:
+        return None, None, None, None
+
+
+# ===============================
+# MIXUP AUGMENTATION
+# ===============================
+def mixup_data(x, frame_mask, modality_mask, y, alpha=0.2):
+    """Mixup augmentation"""
+    if alpha > 0:
+        lam = np.random.beta(alpha, alpha)
+    else:
+        lam = 1
+
+    batch_size = x.size(0)
+    index = torch.randperm(batch_size).to(x.device)
+
+    mixed_x = lam * x + (1 - lam) * x[index]
+    mixed_frame_mask = frame_mask | frame_mask[index]  # Union of valid frames
+    mixed_modality_mask = torch.max(modality_mask, modality_mask[index])
+
+    y_a, y_b = y, y[index]
+    return mixed_x, mixed_frame_mask, mixed_modality_mask, y_a, y_b, lam
+
+
+# ===============================
+# ENHANCED MODEL WITH ATTENTION POOLING
+# ===============================
+class PositionalEncoding(nn.Module):
+    def __init__(self, d_model, max_len=128):
         super().__init__()
-        self.conv1 = nn.Conv1d(num_feats, 256, kernel_size=3, padding=1)
+        pe = torch.zeros(max_len, d_model)
-        self.bn1 = nn.BatchNorm1d(256)
+        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
-        self.conv2 = nn.Conv1d(256, 512, kernel_size=3, padding=1)
+        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
-        self.bn2 = nn.BatchNorm1d(512)
+        pe[:, 0::2] = torch.sin(position * div_term)
-        self.pool = nn.MaxPool1d(2)
+        pe[:, 1::2] = torch.cos(position * div_term)
-        self.dropout = nn.Dropout(0.5)
+        self.register_buffer('pe', pe.unsqueeze(0))
-        self.fc = nn.Linear(512, num_classes)
 
     def forward(self, x):
-        x = x.view(x.shape[0], x.shape[1], -1)
+        return x + self.pe[:, :x.size(1)]
-        x = x.transpose(1, 2)
-        x = F.relu(self.bn1(self.conv1(x)))
-        x = self.pool(x)
-        x = F.relu(self.bn2(self.conv2(x)))
-        x = self.pool(x)
-        x = F.adaptive_avg_pool1d(x, 1).squeeze(-1)
-        x = self.dropout(x)
-        return self.fc(x)
 
 
-# --- TRAINING FUNCTIONS ---
+class ModalityAwareTransformer(nn.Module):
-def train_epoch(model, dataloader, criterion, optimizer, device):
+    def __init__(self, input_dim, num_classes, d_model=512, nhead=8, num_layers=6, dropout=0.15):
+        super().__init__()
+
+        # Main projection
+        self.proj = nn.Linear(input_dim, d_model)
+
+        # Modality embedding (3 modalities: left_hand, right_hand, face)
+        self.modality_embed = nn.Linear(3, d_model)
+
+        self.norm_in = nn.LayerNorm(d_model)
+        self.pos = PositionalEncoding(d_model)
+
+        enc_layer = nn.TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=d_model * 4,
+            dropout=dropout,
+            activation='gelu',
+            batch_first=True,
+            norm_first=True
+        )
+        self.encoder = nn.TransformerEncoder(enc_layer, num_layers=num_layers)
+
+        # Attention pooling
+        self.attention_pool = nn.Linear(d_model, 1)
+
+        self.head = nn.Sequential(
+            nn.LayerNorm(d_model),
+            nn.Dropout(0.3),
+            nn.Linear(d_model, d_model // 2),
+            nn.GELU(),
+            nn.Dropout(0.2),
+            nn.Linear(d_model // 2, num_classes)
+        )
+
+    def forward(self, x, modality_mask=None, key_padding_mask=None):
+        # Project features
+        x = self.proj(x)
+
+        # Add modality information
+        if modality_mask is not None:
+            mod_embed = self.modality_embed(modality_mask)
+            x = x + mod_embed
+
+        x = self.norm_in(x)
+        x = self.pos(x)
+        x = self.encoder(x, src_key_padding_mask=key_padding_mask)
+
+        # Attention-based pooling
+        attn_weights = self.attention_pool(x)  # (B, T, 1)
+        if key_padding_mask is not None:
+            attn_weights = attn_weights.masked_fill(key_padding_mask.unsqueeze(-1), -1e9)
+        attn_weights = F.softmax(attn_weights, dim=1)
+        x = (x * attn_weights).sum(dim=1)
+
+        return self.head(x)
+
+
+def load_kaggle_asl_data(base_path):
+    """Load training metadata using Polars"""
+    train_path = os.path.join(base_path, "train.csv")
+    train_df = pl.read_csv(train_path)
+    return train_df, None
+
+
+# ===============================
+# DATASET WITH AUGMENTATION
+# ===============================
+class ASLMultiDataset(Dataset):
+    def __init__(self, X, frame_masks, modality_masks, y, training=False, max_frames=100):
+        self.X = X
+        self.frame_masks = frame_masks
+        self.modality_masks = modality_masks
+        self.y = y
+        self.training = training
+        self.max_frames = max_frames
+
+    def __len__(self):
+        return len(self.X)
+
+    def __getitem__(self, idx):
+        x = self.X[idx].copy()
+        frame_mask = self.frame_masks[idx].copy()
+        modality_mask = self.modality_masks[idx].copy()
+        y = self.y[idx]
+
+        if self.training:
+            # Apply augmentation
+            x, modality_mask = augment_sequence(x, modality_mask)
+
+            # Re-pad if needed after augmentation
+            if len(x) < self.max_frames:
+                pad = np.zeros((self.max_frames - len(x), x.shape[1]), dtype=np.float32)
+                x = np.concatenate([x, pad], axis=0)
+
+                mask_pad = np.zeros((self.max_frames - len(x), 3), dtype=np.float32)
+                modality_mask = np.concatenate([modality_mask, mask_pad], axis=0)
+
+                frame_mask = np.zeros(self.max_frames, dtype=bool)
+                frame_mask[:len(x)] = True
+            else:
+                x = x[:self.max_frames]
+                modality_mask = modality_mask[:self.max_frames]
+                frame_mask = np.ones(self.max_frames, dtype=bool)
+
+        return (
+            torch.from_numpy(x).float(),
+            torch.from_numpy(frame_mask).bool(),
+            torch.from_numpy(modality_mask).float(),
+            torch.tensor(y, dtype=torch.long)
+        )
+
+
+# ===============================
+# TRAINING SINGLE MODEL
+# ===============================
+def train_model(X_tr, fm_tr, mm_tr, y_tr, X_te, fm_te, mm_te, y_te,
+                num_classes, input_dim, model_idx=0, epochs=80):
+    """Train a single model"""
+
+    # Set different seed for each model
+    torch.manual_seed(42 + model_idx)
+    np.random.seed(42 + model_idx)
+
+    batch_size = 64 if device.type == 'cuda' else 32
+
+    train_dataset = ASLMultiDataset(X_tr, fm_tr, mm_tr, y_tr, training=True, max_frames=100)
+    test_dataset = ASLMultiDataset(X_te, fm_te, mm_te, y_te, training=False, max_frames=100)
+
+    train_loader = DataLoader(
+        train_dataset,
+        batch_size=batch_size, shuffle=True,
+        num_workers=4, pin_memory=device.type == 'cuda'
+    )
+
+    test_loader = DataLoader(
+        test_dataset,
+        batch_size=batch_size * 2, shuffle=False,
+        num_workers=4, pin_memory=device.type == 'cuda'
+    )
+
+    # Enhanced model
+    model = ModalityAwareTransformer(
+        input_dim=input_dim,
+        num_classes=num_classes,
+        d_model=512,
+        nhead=8,
+        num_layers=6,
+        dropout=0.15
+    ).to(device)
+
+    print(f"\n[Model {model_idx + 1}] Parameters: {sum(p.numel() for p in model.parameters()):,}")
+
+    criterion = nn.CrossEntropyLoss(label_smoothing=0.1)
+    optimizer = optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-4)
+
+    # OneCycleLR scheduler
+    scheduler = optim.lr_scheduler.OneCycleLR(
+        optimizer,
+        max_lr=1e-3,
+        steps_per_epoch=len(train_loader),
+        epochs=epochs,
+        pct_start=0.1,
+        anneal_strategy='cos'
+    )
+
+    best_acc = 0.0
+    save_path = f"best_asl_model_{model_idx}.pth"
+
+    print(f"\n{'=' * 60}")
+    print(f"TRAINING MODEL {model_idx + 1}")
+    print(f"{'=' * 60}")
+
+    for epoch in range(epochs):
         model.train()
-    running_loss = 0.0
+        total_loss = correct = total = 0
-    correct = 0
-    total = 0
 
-    progress_bar = tqdm(dataloader, desc="Training")
+        for x, frame_mask, modality_mask, yb in tqdm(train_loader, desc=f"Epoch {epoch + 1}", leave=False):
-    for inputs, labels in progress_bar:
+            x = x.to(device)
-        inputs, labels = inputs.to(device), labels.to(device)
+            frame_mask = frame_mask.to(device)
+            modality_mask = modality_mask.to(device)
+            yb = yb.to(device)
+
+            # Apply mixup
+            if np.random.rand() < 0.5:
+                x, frame_mask, modality_mask, y_a, y_b, lam = mixup_data(
+                    x, frame_mask, modality_mask, yb, alpha=0.2
+                )
+
+                key_padding_mask = ~frame_mask
+                optimizer.zero_grad(set_to_none=True)
+                logits = model(x, modality_mask=modality_mask, key_padding_mask=key_padding_mask)
+                loss = lam * criterion(logits, y_a) + (1 - lam) * criterion(logits, y_b)
+
+                # Use original labels for accuracy
+                correct += (logits.argmax(-1) == yb).sum().item()
+            else:
+                key_padding_mask = ~frame_mask
+                optimizer.zero_grad(set_to_none=True)
+                logits = model(x, modality_mask=modality_mask, key_padding_mask=key_padding_mask)
+                loss = criterion(logits, yb)
+                correct += (logits.argmax(-1) == yb).sum().item()
 
-        optimizer.zero_grad()
-        outputs = model(inputs)
-        loss = criterion(outputs, labels)
             loss.backward()
+            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
             optimizer.step()
+            scheduler.step()
 
-        running_loss += loss.item()
+            total_loss += loss.item()
-        _, predicted = torch.max(outputs.data, 1)
+            total += yb.size(0)
-        total += labels.size(0)
-        correct += (predicted == labels).sum().item()
 
-        progress_bar.set_postfix({
+        train_acc = correct / total * 100
-            'loss': running_loss / (progress_bar.n + 1),
-            'acc': 100 * correct / total
-        })
 
-    epoch_loss = running_loss / len(dataloader)
+        # Eval
-    epoch_acc = 100 * correct / total
-    return epoch_loss, epoch_acc
-
-
-def validate(model, dataloader, criterion, device):
         model.eval()
-    running_loss = 0.0
+        correct = total = 0
-    correct = 0
-    total = 0
-
         with torch.no_grad():
-        for inputs, labels in tqdm(dataloader, desc="Validation"):
+            for x, frame_mask, modality_mask, yb in test_loader:
-            inputs, labels = inputs.to(device), labels.to(device)
+                x = x.to(device)
-            outputs = model(inputs)
+                frame_mask = frame_mask.to(device)
-            loss = criterion(outputs, labels)
+                modality_mask = modality_mask.to(device)
+                yb = yb.to(device)
 
-            running_loss += loss.item()
+                key_padding_mask = ~frame_mask
-            _, predicted = torch.max(outputs.data, 1)
+                logits = model(x, modality_mask=modality_mask, key_padding_mask=key_padding_mask)
-            total += labels.size(0)
+                correct += (logits.argmax(-1) == yb).sum().item()
-            correct += (predicted == labels).sum().item()
+                total += yb.size(0)
 
-    val_loss = running_loss / len(dataloader)
+        test_acc = correct / total * 100
-    val_acc = 100 * correct / total
+
-    return val_loss, val_acc
+        print(f"[{epoch + 1:2d}/{epochs}] Loss: {total_loss / len(train_loader):.4f} | "
+              f"Train: {train_acc:.2f}% | Test: {test_acc:.2f}%", end="")
+
+        if test_acc > best_acc:
+            best_acc = test_acc
+            torch.save(model.state_dict(), save_path)
+            print("  → saved")
+        else:
+            print()
+
+    print(f"\nModel {model_idx + 1} - Best test accuracy: {best_acc:.2f}%")
+    return save_path, best_acc
 
 
-def save_checkpoint(model, optimizer, epoch, train_loss, val_loss, val_acc, checkpoint_dir):
+# ===============================
-    os.makedirs(checkpoint_dir, exist_ok=True)
+# ENSEMBLE PREDICTION
-    checkpoint = {
+# ===============================
-        'epoch': epoch,
+def ensemble_predict(model_paths, test_loader, num_classes, input_dim):
-        'model_state_dict': model.state_dict(),
+    """Make predictions using ensemble of models"""
-        'optimizer_state_dict': optimizer.state_dict(),
+    all_preds = []
-        'train_loss': train_loss,
+
-        'val_loss': val_loss,
+    for model_path in model_paths:
-        'val_acc': val_acc,
+        model = ModalityAwareTransformer(
-    }
+            input_dim=input_dim,
-    path = os.path.join(checkpoint_dir, f'checkpoint_epoch_{epoch}.pt')
+            num_classes=num_classes,
-    torch.save(checkpoint, path)
+            d_model=512,
-    print(f"Checkpoint saved: {path}")
+            nhead=8,
+            num_layers=6
+        ).to(device)
+
+        model.load_state_dict(torch.load(model_path))
+        model.eval()
+
+        preds = []
+        with torch.no_grad():
+            for x, frame_mask, modality_mask, _ in test_loader:
+                x = x.to(device)
+                frame_mask = frame_mask.to(device)
+                modality_mask = modality_mask.to(device)
+
+                key_padding_mask = ~frame_mask
+                logits = model(x, modality_mask=modality_mask, key_padding_mask=key_padding_mask)
+                preds.append(F.softmax(logits, dim=-1))
+
+        all_preds.append(torch.cat(preds, dim=0))
+
+    # Average predictions
+    ensemble_pred = torch.stack(all_preds).mean(0)
+    return ensemble_pred.argmax(-1).cpu().numpy()
+
+
+# ===============================
+# MAIN
+# ===============================
+def main():
+    base_path = "asl_kaggle"
+    max_frames = 100
+    MIN_SAMPLES_PER_CLASS = 3  # Relaxed from 5
+    NUM_ENSEMBLE_MODELS = 3
+    EPOCHS = 80
+
+    print("\nLoading metadata...")
+    train_df, _ = load_kaggle_asl_data(base_path)
+    print(f"Total samples in train.csv: {train_df.height}")
+
+    rows = [(row[0], row[1]) for row in train_df.select(["path", "sign"]).iter_rows()]
+
+    print("\nProcessing sequences with BOTH hands + FACE (enhanced)...")
+    print("This may take a few minutes...")
+
+    with Pool(cpu_count()) as pool:
+        results = list(tqdm(
+            pool.imap(
+                partial(process_row, base_path=base_path, max_frames=max_frames),
+                rows,
+                chunksize=80
+            ),
+            total=len(rows),
+            desc="Landmarks extraction"
+        ))
+
+    X_list, frame_masks_list, modality_masks_list, y_list = [], [], [], []
+    failed_count = 0
+    for feat, frame_mask, modality_mask, sign in results:
+        if feat is not None and frame_mask is not None:
+            X_list.append(feat)
+            frame_masks_list.append(frame_mask)
+            modality_masks_list.append(modality_mask)
+            y_list.append(sign)
+        else:
+            failed_count += 1
+
+    if not X_list:
+        print(f"\n❌ No valid sequences extracted!")
+        print(f"Failed to process: {failed_count}/{len(results)} files")
+        return
+
+    print(f"\n✓ Successfully processed: {len(X_list)}/{len(results)} files")
+    print(f"✗ Failed: {failed_count}/{len(results)} files")
+
+    X = np.stack(X_list)
+    frame_masks = np.stack(frame_masks_list)
+    modality_masks = np.stack(modality_masks_list)
+
+    print(f"\nExtracted {len(X):,} sequences")
+    print(f"Feature shape: {X.shape[1:]}   (input_dim = {X.shape[2]})")
+
+    # Global normalization
+    X = np.clip(X, -30, 30)
+    mean = X.mean(axis=(0, 1), keepdims=True)
+    std = X.std(axis=(0, 1), keepdims=True) + 1e-8
+    X = (X - mean) / std
+
+    # Labels
+    le = LabelEncoder()
+    y = le.fit_transform(y_list)
+
+    # Filter rare classes
+    counts = Counter(y)
+    valid = [k for k, v in counts.items() if v >= MIN_SAMPLES_PER_CLASS]
+    mask = np.isin(y, valid)
+
+    X = X[mask]
+    frame_masks = frame_masks[mask]
+    modality_masks = modality_masks[mask]
+    y = y[mask]
+
+    le = LabelEncoder()
+    y = le.fit_transform(y)
+
+    print(f"After filtering: {len(X):,} samples | {len(le.classes_)} classes")
+
+    # Split
+    X_tr, X_te, fm_tr, fm_te, mm_tr, mm_te, y_tr, y_te = train_test_split(
+        X, frame_masks, modality_masks, y, test_size=0.15, stratify=y, random_state=42
+    )
+
+    # Train ensemble of models
+    model_paths = []
+    best_accs = []
+
+    for i in range(NUM_ENSEMBLE_MODELS):
+        model_path, best_acc = train_model(
+            X_tr, fm_tr, mm_tr, y_tr,
+            X_te, fm_te, mm_te, y_te,
+            num_classes=len(le.classes_),
+            input_dim=X.shape[2],
+            model_idx=i,
+            epochs=EPOCHS
+        )
+        model_paths.append(model_path)
+        best_accs.append(best_acc)
+
+    # Ensemble evaluation
+    print("\n" + "=" * 60)
+    print("ENSEMBLE EVALUATION")
+    print("=" * 60)
+
+    test_dataset = ASLMultiDataset(X_te, fm_te, mm_te, y_te, training=False)
+    test_loader = DataLoader(
+        test_dataset,
+        batch_size=128,
+        shuffle=False,
+        num_workers=4,
+        pin_memory=device.type == 'cuda'
+    )
+
+    ensemble_preds = ensemble_predict(model_paths, test_loader, len(le.classes_), X.shape[2])
+    ensemble_acc = (ensemble_preds == y_te).mean() * 100
+
+    print(f"\nIndividual model accuracies:")
+    for i, acc in enumerate(best_accs):
+        print(f"  Model {i + 1}: {acc:.2f}%")
+
+    print(f"\n🎯 Ensemble accuracy: {ensemble_acc:.2f}%")
+    print(f"   Improvement: +{ensemble_acc - max(best_accs):.2f}% over best single model")
+
+    print("\n" + "=" * 60)
+    print(f"TRAINING COMPLETE")
+    print("=" * 60)
 
 
-# --- EXECUTION ---
 if __name__ == "__main__":
-    # Load metadata
+    main()
-    asl_data = load_kaggle_metadata(BASE_PATH)
-
-    # Create label mapping
-    unique_signs = sorted(asl_data["sign"].unique().to_list())
-    label_to_idx = {sign: idx for idx, sign in enumerate(unique_signs)}
-    labels = torch.tensor([label_to_idx[sign] for sign in asl_data["sign"].to_list()])
-
-    print(f"Number of classes: {len(unique_signs)}")
-
-    # Process data in parallel
-    paths = asl_data["path"].to_list()
-    print(f"Processing {len(paths)} files in parallel...")
-
-    with ProcessPoolExecutor() as executor:
-        results = list(tqdm(executor.map(load_and_preprocess, paths), total=len(paths)))
-
-    dataset_tensor = torch.tensor(np.array(results), dtype=torch.float32)
-    print(f"Final Tensor Shape: {dataset_tensor.shape}")
-
-    # Create dataset and split
-    full_dataset = ASLDataset(dataset_tensor, labels)
-    train_size = int(TRAIN_SPLIT * len(full_dataset))
-    val_size = len(full_dataset) - train_size
-    train_dataset, val_dataset = random_split(full_dataset, [train_size, val_size])
-
-    train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=0)
-    val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)
-
-    print(f"Train samples: {train_size}, Validation samples: {val_size}")
-
-    # Initialize model, loss, optimizer
-    model = ASLClassifier(num_classes=len(unique_signs)).to(device)
-    criterion = nn.CrossEntropyLoss()
-    optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
-    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=5)
-
-    # Training loop
-    best_val_acc = 0.0
-
-    print("\n" + "=" * 50)
-    print("Starting Training")
-    print("=" * 50 + "\n")
-
-    for epoch in range(EPOCHS):
-        print(f"\nEpoch [{epoch + 1}/{EPOCHS}]")
-        print("-" * 50)
-
-        train_loss, train_acc = train_epoch(model, train_loader, criterion, optimizer, device)
-        val_loss, val_acc = validate(model, val_loader, criterion, device)
-
-        scheduler.step(val_loss)
-
-        print(f"\nEpoch {epoch + 1} Summary:")
-        print(f"  Train Loss: {train_loss:.4f} | Train Acc: {train_acc:.2f}%")
-        print(f"  Val Loss:   {val_loss:.4f} | Val Acc:   {val_acc:.2f}%")
-        print(f"  Learning Rate: {optimizer.param_groups[0]['lr']:.6f}")
-
-        # Save checkpoint if validation accuracy improved
-        if val_acc > best_val_acc:
-            best_val_acc = val_acc
-            save_checkpoint(model, optimizer, epoch + 1, train_loss, val_loss, val_acc, CHECKPOINT_DIR)
-            print(f"  ✓ New best validation accuracy: {best_val_acc:.2f}%")
-
-    print("\n" + "=" * 50)
-    print("Training Complete!")
-    print(f"Best Validation Accuracy: {best_val_acc:.2f}%")
-    print("=" * 50)