AI, we don't believe that you can fly bro pt 2

training.py

@@ -28,6 +28,7 @@ if torch.cuda.is_available():
     print(f"✓ GPU: {torch.cuda.get_device_name(0)}")
     device = torch.device('cuda:0')
     torch.backends.cudnn.benchmark = True
+    torch.backends.cudnn.enabled = True
 else:
     print("✗ CUDA not available, using CPU")
     device = torch.device('cpu')
@@ -36,7 +37,7 @@ print("=" * 60)
 
 
 # ===============================
-# DATA LOADING WITH NaN HANDLING
+# DATA LOADING - HANDLES PARTIAL NaN
 # ===============================
 def load_kaggle_asl_data(base_path):
     train_df = pd.read_csv(os.path.join(base_path, "train.csv"))
@@ -46,7 +47,7 @@ def load_kaggle_asl_data(base_path):
 
 
 def extract_hand_landmarks_from_parquet(path):
-    """Extract hand landmarks, handling NaN values properly"""
+    """Extract hand landmarks - ONLY uses frames with valid (non-NaN) data"""
     try:
         df = pd.read_parquet(path)
 
@@ -54,56 +55,70 @@ def extract_hand_landmarks_from_parquet(path):
         left = df[df["type"] == "left_hand"]
         right = df[df["type"] == "right_hand"]
 
-        # Choose hand with more non-NaN data
+        if len(left) == 0 and len(right) == 0:
-        left_valid = left[['x', 'y', 'z']].notna().all(axis=1).sum()
-        right_valid = right[['x', 'y', 'z']].notna().all(axis=1).sum()
-
-        if left_valid == 0 and right_valid == 0:
-            return None  # No valid hand data
-
-        hand = left if left_valid >= right_valid else right
-
-        if len(hand) == 0:
             return None
 
-        # Get frames with valid data
+        # Count valid (non-NaN) rows for each hand
+        left_valid = 0
+        right_valid = 0
+
+        if len(left) > 0:
+            left_valid = left[['x', 'y', 'z']].notna().all(axis=1).sum()
+        if len(right) > 0:
+            right_valid = right[['x', 'y', 'z']].notna().all(axis=1).sum()
+
+        # No valid data at all
+        if left_valid == 0 and right_valid == 0:
+            return None
+
+        # Choose hand with more valid data
+        hand = left if left_valid >= right_valid else right
+
+        # Get unique frames
         frames = sorted(hand['frame'].unique())
         landmarks_seq = []
 
         for frame in frames:
             lm_frame = hand[hand['frame'] == frame]
 
-            # Check if this frame has valid data
+            # Count how many valid landmarks this frame has
-            valid_rows = lm_frame[['x', 'y', 'z']].notna().all(axis=1)
+            valid_count = lm_frame[['x', 'y', 'z']].notna().all(axis=1).sum()
-            if valid_rows.sum() < 10:  # Need at least 10 valid landmarks
+
+            # Skip frames with too few valid landmarks
+            if valid_count < 10:
                 continue
 
-            lm_list = []
+            # Extract landmarks for this frame
-            frame_has_data = False
+            frame_landmarks = []
+            valid_landmarks_in_frame = 0
 
             for i in range(21):
                 lm = lm_frame[lm_frame['landmark_index'] == i]
+
                 if len(lm) == 0:
-                    lm_list.append([0.0, 0.0, 0.0])
+                    frame_landmarks.append([0.0, 0.0, 0.0])
                 else:
-                    x = lm['x'].iloc[0]
+                    x = float(lm['x'].iloc[0])
-                    y = lm['y'].iloc[0]
+                    y = float(lm['y'].iloc[0])
-                    z = lm['z'].iloc[0]
+                    z = float(lm['z'].iloc[0])
 
-                    # Check for NaN
+                    # Check if valid
-                    if pd.isna(x) or pd.isna(y) or pd.isna(z):
+                    if pd.notna(x) and pd.notna(y) and pd.notna(z):
-                        lm_list.append([0.0, 0.0, 0.0])
+                        frame_landmarks.append([x, y, z])
+                        valid_landmarks_in_frame += 1
                     else:
-                        lm_list.append([float(x), float(y), float(z)])
+                        frame_landmarks.append([0.0, 0.0, 0.0])
-                        frame_has_data = True
 
-            if frame_has_data:
+            # Only add frame if it has enough valid landmarks
-                landmarks_seq.append(lm_list)
+            if valid_landmarks_in_frame >= 10:
+                landmarks_seq.append(frame_landmarks)
 
-        if len(landmarks_seq) == 0:
+        # Need at least 3 valid frames
+        if len(landmarks_seq) < 3:
             return None
 
         return np.array(landmarks_seq, dtype=np.float32)
+
     except Exception as e:
         return None
 
@@ -113,43 +128,57 @@ def get_features_sequence(landmarks_seq, max_frames=100):
     if landmarks_seq is None or len(landmarks_seq) == 0:
         return None, None
 
-    # Center on wrist
+    # Center on wrist (landmark 0)
-    landmarks_seq = landmarks_seq - landmarks_seq[:, 0:1, :]
+    wrist = landmarks_seq[:, 0:1, :].copy()
+    landmarks_seq = landmarks_seq - wrist
 
-    # Scale using wrist → middle finger MCP distance
+    # Scale normalization using wrist to middle finger MCP (landmark 9)
-    scale = np.linalg.norm(landmarks_seq[:, 0] - landmarks_seq[:, 9], axis=1, keepdims=True)
+    scale = np.linalg.norm(landmarks_seq[:, 9, :] - np.zeros(3), axis=1, keepdims=True)
-    scale = np.maximum(scale, 1e-6)
+    scale = np.maximum(scale, 1e-6)  # Avoid division by zero
     landmarks_seq = landmarks_seq / scale[:, np.newaxis, :]
 
-    # Replace any remaining NaN/Inf with 0
+    # Clean up any remaining NaN/Inf
     landmarks_seq = np.nan_to_num(landmarks_seq, nan=0.0, posinf=0.0, neginf=0.0)
 
-    # Finger curl distances
+    # Clip extreme values
-    tips = [4, 8, 12, 16, 20]
+    landmarks_seq = np.clip(landmarks_seq, -10, 10)
-    bases = [1, 5, 9, 13, 17]
+
+    # Calculate finger curl features
+    tips = [4, 8, 12, 16, 20]  # Thumb, index, middle, ring, pinky tips
+    bases = [1, 5, 9, 13, 17]  # Corresponding base joints
+
     curl_features = []
     for b, t in zip(bases, tips):
         curl = np.linalg.norm(landmarks_seq[:, t] - landmarks_seq[:, b], axis=1)
         curl_features.append(curl)
-    curl_features = np.stack(curl_features, axis=1)
+    curl_features = np.stack(curl_features, axis=1)  # (T, 5)
 
-    # Temporal deltas
+    # Temporal deltas (motion)
     deltas = np.zeros_like(landmarks_seq)
+    if len(landmarks_seq) > 1:
         deltas[1:] = landmarks_seq[1:] - landmarks_seq[:-1]
 
-    # Flatten features
+    # Combine all features
-    seq = np.concatenate([landmarks_seq, deltas, curl_features[:, :, np.newaxis]], axis=2)
+    seq = np.concatenate([
+        landmarks_seq,  # (T, 21, 3)
+        deltas,  # (T, 21, 3)
+        curl_features[:, :, np.newaxis]  # (T, 5, 1)
+    ], axis=2)
+
+    # Flatten spatial dimensions: (T, 21*3 + 21*3 + 5) = (T, 131)
     seq = seq.reshape(seq.shape[0], -1)
 
-    # Pad or truncate
+    # Pad or truncate to max_frames
     T, F = seq.shape
     if T < max_frames:
+        # Pad with zeros
         pad = np.zeros((max_frames - T, F), dtype=np.float32)
         seq = np.concatenate([seq, pad], axis=0)
     elif T > max_frames:
+        # Truncate
         seq = seq[:max_frames, :]
 
-    # Create mask
+    # Create attention mask (True for valid positions)
     valid_mask = np.zeros(max_frames, dtype=bool)
     valid_mask[:min(T, max_frames)] = True
 
@@ -157,26 +186,30 @@ def get_features_sequence(landmarks_seq, max_frames=100):
 
 
 def process_row(row, base_path, max_frames=100):
-    """Process a single row"""
+    """Process a single row - worker function for multiprocessing"""
     path = os.path.join(base_path, row["path"])
+
     if not os.path.exists(path):
         return None, None, None
 
     try:
+        # Extract landmarks
         lm = extract_hand_landmarks_from_parquet(path)
         if lm is None:
             return None, None, None
 
+        # Get features
         feat, mask = get_features_sequence(lm, max_frames)
         if feat is None:
             return None, None, None
 
-        # Final NaN check
+        # Final safety check
         if np.isnan(feat).any() or np.isinf(feat).any():
             return None, None, None
 
         return feat, mask, row["sign"]
-    except:
+
+    except Exception as e:
         return None, None, None
 
 
@@ -198,12 +231,15 @@ class PositionalEncoding(nn.Module):
 
 
 class TransformerASL(nn.Module):
-    def __init__(self, input_dim=68, num_classes=250, d_model=256, nhead=8, num_layers=4):
+    def __init__(self, input_dim, num_classes, d_model=256, nhead=8, num_layers=4):
         super().__init__()
+
+        # Input projection
         self.proj = nn.Linear(input_dim, d_model)
         self.norm_in = nn.LayerNorm(d_model)
-        self.pos = PositionalEncoding(d_model)
+        self.pos = PositionalEncoding(d_model, max_len=128)
 
+        # Transformer encoder
         enc_layer = nn.TransformerEncoderLayer(
             d_model=d_model,
             nhead=nhead,
@@ -215,6 +251,7 @@ class TransformerASL(nn.Module):
         )
         self.encoder = nn.TransformerEncoder(enc_layer, num_layers=num_layers)
 
+        # Classification head
         self.head = nn.Sequential(
             nn.LayerNorm(d_model),
             nn.Dropout(0.25),
@@ -222,11 +259,17 @@ class TransformerASL(nn.Module):
         )
 
     def forward(self, x, key_padding_mask=None):
+        # x: (batch, seq_len, input_dim)
+        # key_padding_mask: (batch, seq_len) - True for padding positions
+
         x = self.proj(x)
         x = self.norm_in(x)
         x = self.pos(x)
         x = self.encoder(x, src_key_padding_mask=key_padding_mask)
+
+        # Global average pooling over valid positions
         x = x.mean(dim=1)
+
         return self.head(x)
 
 
@@ -236,7 +279,7 @@ class TransformerASL(nn.Module):
 def main():
     base_path = "asl_kaggle"
     max_frames = 100
-    MIN_SAMPLES_PER_CLASS = 6
+    MIN_SAMPLES_PER_CLASS = 5
 
     print("\nLoading metadata...")
     train_df, sign_to_idx = load_kaggle_asl_data(base_path)
@@ -244,7 +287,10 @@ def main():
 
     rows = [row for _, row in train_df.iterrows()]
 
-    print("\nProcessing sequences with NaN handling...")
+    print("\nProcessing sequences (this will take a few minutes)...")
+    print("Expected: ~36,000 valid sequences based on diagnostic")
+
+    # Process with multiprocessing
     with Pool(cpu_count()) as pool:
         results = list(tqdm(
             pool.imap(
@@ -259,27 +305,30 @@ def main():
     # Filter valid results
     X_list, masks_list, y_list = [], [], []
     for feat, mask, sign in results:
-        if feat is not None and feat.shape[0] == max_frames:
+        if feat is not None and mask is not None and sign is not None:
+            if feat.shape[0] == max_frames:
                 X_list.append(feat)
                 masks_list.append(mask)
                 y_list.append(sign)
 
-    print(f"\n✓ Valid sequences: {len(X_list)} out of {len(train_df)}")
+    print(f"\n✓ Successfully extracted: {len(X_list)} valid sequences")
+    print(f"  Success rate: {len(X_list) / len(train_df) * 100:.1f}%")
 
-    if not X_list:
+    if len(X_list) < 100:
-        print("❌ No valid sequences found!")
+        print("❌ Too few valid sequences found!")
-        print("\nPossible issues:")
+        print("   This shouldn't happen - please share this output for debugging")
-        print("  1. Most files contain only NaN hand landmarks")
-        print("  2. Hand detection failed in most videos")
-        print("  3. Dataset might be corrupted")
         return
 
+    # Stack into arrays
     X = np.stack(X_list)
     masks = np.stack(masks_list)
-    print(f"Data shape: {X.shape}")
+
+    print(f"\nData shape: {X.shape}")
+    print(f"Feature dimension: {X.shape[2]}")
 
     # Global normalization
-    X = np.clip(X, -5.0, 5.0)
+    print("Normalizing features...")
+    X = np.clip(X, -10.0, 10.0)
     mean = X.mean(axis=(0, 1), keepdims=True)
     std = X.std(axis=(0, 1), keepdims=True) + 1e-8
     X = (X - mean) / std
@@ -292,22 +341,29 @@ def main():
     counts = Counter(y)
     valid_classes = [cls for cls, cnt in counts.items() if cnt >= MIN_SAMPLES_PER_CLASS]
     mask_valid = np.isin(y, valid_classes)
+
     X = X[mask_valid]
     masks = masks[mask_valid]
     y = y[mask_valid]
 
-    # Re-encode
+    # Re-encode after filtering
     le = LabelEncoder()
     y = le.fit_transform(y)
 
-    print(f"Final dataset: {len(X)} samples | {len(le.classes_)} classes")
+    print(f"\nFinal dataset after filtering:")
+    print(f"  Samples: {len(X):,}")
+    print(f"  Classes: {len(le.classes_)}")
+    print(f"  Sign examples: {list(le.classes_[:10])}")
 
     # Train-test split
     X_train, X_test, masks_train, masks_test, y_train, y_test = train_test_split(
         X, masks, y, test_size=0.15, stratify=y, random_state=42
     )
 
-    # Dataset
+    print(f"\nTrain set: {len(X_train):,} samples")
+    print(f"Test set: {len(X_test):,} samples")
+
+    # Dataset wrapper
     class ASLSequenceDataset(Dataset):
         def __init__(self, X, masks, y):
             self.X = torch.from_numpy(X).float()
@@ -320,18 +376,27 @@ def main():
         def __getitem__(self, idx):
             return self.X[idx], self.masks[idx], self.y[idx]
 
+    # DataLoaders
     batch_size = 128 if device.type == 'cuda' else 64
 
     train_loader = DataLoader(
         ASLSequenceDataset(X_train, masks_train, y_train),
-        batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True
+        batch_size=batch_size,
-    )
+        shuffle=True,
-    test_loader = DataLoader(
+        num_workers=4,
-        ASLSequenceDataset(X_test, masks_test, y_test),
+        pin_memory=True if device.type == 'cuda' else False
-        batch_size=batch_size * 2, shuffle=False, num_workers=4, pin_memory=True
     )
 
-    # Model
+    test_loader = DataLoader(
+        ASLSequenceDataset(X_test, masks_test, y_test),
+        batch_size=batch_size * 2,
+        shuffle=False,
+        num_workers=4,
+        pin_memory=True if device.type == 'cuda' else False
+    )
+
+    # Initialize model
+    print("\nInitializing model...")
     model = TransformerASL(
         input_dim=X.shape[2],
         num_classes=len(le.classes_),
@@ -341,35 +406,41 @@ def main():
     ).to(device)
 
     total_params = sum(p.numel() for p in model.parameters())
-    print(f"\nModel parameters: {total_params:,}")
+    print(f"Model parameters: {total_params:,}")
 
+    # Training setup
     criterion = nn.CrossEntropyLoss(label_smoothing=0.05)
     optimizer = optim.AdamW(model.parameters(), lr=5e-4, weight_decay=1e-4)
-    scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10)
+    scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10, T_mult=2)
 
-    # Training
+    # Training loop
     best_acc = 0.0
     patience = 15
     wait = 0
-    epochs = 70
+    epochs = 60
 
-    print("\nStarting training...")
+    print("\n" + "=" * 60)
+    print("STARTING TRAINING")
     print("=" * 60)
 
     for epoch in range(epochs):
+        # Train
         model.train()
         total_loss = 0
         correct = total = 0
 
-        for x, mask, yb in tqdm(train_loader, desc=f"Epoch {epoch + 1}/{epochs}"):
+        for x, mask, yb in tqdm(train_loader, desc=f"Epoch {epoch + 1}/{epochs}", leave=False):
             x, mask, yb = x.to(device), mask.to(device), yb.to(device)
+
+            # Invert mask: True for padding positions
             key_mask = ~mask
 
             optimizer.zero_grad(set_to_none=True)
             logits = model(x, key_padding_mask=key_mask)
             loss = criterion(logits, yb)
             loss.backward()
-            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=0.8)
+
+            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
             optimizer.step()
 
             total_loss += loss.item()
@@ -378,7 +449,7 @@ def main():
 
         train_acc = correct / total * 100
 
-        # Eval
+        # Evaluate
         model.eval()
         correct = total = 0
         with torch.no_grad():
@@ -391,11 +462,13 @@ def main():
 
         test_acc = correct / total * 100
 
+        # Print progress
         print(f"[{epoch + 1:2d}/{epochs}] Loss: {total_loss / len(train_loader):.4f} | "
-              f"Train: {train_acc:.2f}% | Test: {test_acc:.2f}%")
+              f"Train: {train_acc:.2f}% | Test: {test_acc:.2f}%", end="")
 
         scheduler.step()
 
+        # Save best model
         if test_acc > best_acc:
             best_acc = test_acc
             wait = 0
@@ -406,19 +479,25 @@ def main():
                 'acc': best_acc,
                 'epoch': epoch,
                 'input_dim': X.shape[2],
-                'num_classes': len(le.classes_)
+                'num_classes': len(le.classes_),
+                'd_model': 256,
+                'nhead': 8,
+                'num_layers': 4
             }, "best_asl_transformer.pth")
-            print(f"  → New best: {best_acc:.2f}%")
+            print(f" → New best: {best_acc:.2f}% ✓")
         else:
             wait += 1
+            print()
+
             if wait >= patience:
-                print("Early stopping")
+                print(f"\nEarly stopping triggered at epoch {epoch + 1}")
                 break
 
-    print("=" * 60)
+    print("\n" + "=" * 60)
-    print(f"\n✓ Training complete!")
+    print(f"✓ Training complete!")
     print(f"✓ Best test accuracy: {best_acc:.2f}%")
     print(f"✓ Model saved: best_asl_transformer.pth")
+    print("=" * 60)
 
 
 if __name__ == "__main__":

valid_files.json

@@ -0,0 +1,626 @@
+[
+  {
+    "path": "train_landmark_files/16069/100015657.parquet",
+    "sign": "cloud",
+    "hand": "left",
+    "total_frames": 105,
+    "valid_frames": 28,
+    "nan_ratio": 0.7333333333333333
+  },
+  {
+    "path": "train_landmark_files/32319/1000278229.parquet",
+    "sign": "lips",
+    "hand": "left",
+    "total_frames": 57,
+    "valid_frames": 36,
+    "nan_ratio": 0.3684210526315789
+  },
+  {
+    "path": "train_landmark_files/36257/1000536928.parquet",
+    "sign": "apple",
+    "hand": "left",
+    "total_frames": 13,
+    "valid_frames": 10,
+    "nan_ratio": 0.23076923076923078
+  },
+  {
+    "path": "train_landmark_files/22343/1000638205.parquet",
+    "sign": "puzzle",
+    "hand": "left",
+    "total_frames": 19,
+    "valid_frames": 11,
+    "nan_ratio": 0.42105263157894735
+  },
+  {
+    "path": "train_landmark_files/27610/1000697904.parquet",
+    "sign": "there",
+    "hand": "left",
+    "total_frames": 43,
+    "valid_frames": 15,
+    "nan_ratio": 0.6511627906976745
+  },
+  {
+    "path": "train_landmark_files/61333/1000909322.parquet",
+    "sign": "shirt",
+    "hand": "left",
+    "total_frames": 22,
+    "valid_frames": 22,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/27610/1000956928.parquet",
+    "sign": "owl",
+    "hand": "left",
+    "total_frames": 100,
+    "valid_frames": 98,
+    "nan_ratio": 0.02
+  },
+  {
+    "path": "train_landmark_files/22343/1001223069.parquet",
+    "sign": "not",
+    "hand": "left",
+    "total_frames": 18,
+    "valid_frames": 6,
+    "nan_ratio": 0.6666666666666666
+  },
+  {
+    "path": "train_landmark_files/32319/1001258102.parquet",
+    "sign": "zipper",
+    "hand": "left",
+    "total_frames": 13,
+    "valid_frames": 8,
+    "nan_ratio": 0.38461538461538464
+  },
+  {
+    "path": "train_landmark_files/55372/1001471195.parquet",
+    "sign": "cheek",
+    "hand": "left",
+    "total_frames": 17,
+    "valid_frames": 17,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/36257/1001560021.parquet",
+    "sign": "shoe",
+    "hand": "left",
+    "total_frames": 63,
+    "valid_frames": 63,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/34503/1001685690.parquet",
+    "sign": "empty",
+    "hand": "left",
+    "total_frames": 8,
+    "valid_frames": 8,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/61333/1001819372.parquet",
+    "sign": "balloon",
+    "hand": "left",
+    "total_frames": 27,
+    "valid_frames": 16,
+    "nan_ratio": 0.4074074074074074
+  },
+  {
+    "path": "train_landmark_files/36257/1001899025.parquet",
+    "sign": "same",
+    "hand": "left",
+    "total_frames": 44,
+    "valid_frames": 22,
+    "nan_ratio": 0.5
+  },
+  {
+    "path": "train_landmark_files/28656/1001919956.parquet",
+    "sign": "orange",
+    "hand": "left",
+    "total_frames": 53,
+    "valid_frames": 48,
+    "nan_ratio": 0.09433962264150944
+  },
+  {
+    "path": "train_landmark_files/32319/1001958254.parquet",
+    "sign": "go",
+    "hand": "left",
+    "total_frames": 24,
+    "valid_frames": 23,
+    "nan_ratio": 0.041666666666666664
+  },
+  {
+    "path": "train_landmark_files/61333/1002052130.parquet",
+    "sign": "TV",
+    "hand": "left",
+    "total_frames": 116,
+    "valid_frames": 109,
+    "nan_ratio": 0.0603448275862069
+  },
+  {
+    "path": "train_landmark_files/16069/1002113535.parquet",
+    "sign": "another",
+    "hand": "left",
+    "total_frames": 6,
+    "valid_frames": 3,
+    "nan_ratio": 0.5
+  },
+  {
+    "path": "train_landmark_files/55372/1002129762.parquet",
+    "sign": "giraffe",
+    "hand": "left",
+    "total_frames": 23,
+    "valid_frames": 14,
+    "nan_ratio": 0.391304347826087
+  },
+  {
+    "path": "train_landmark_files/29302/1002284514.parquet",
+    "sign": "can",
+    "hand": "left",
+    "total_frames": 41,
+    "valid_frames": 8,
+    "nan_ratio": 0.8048780487804879
+  },
+  {
+    "path": "train_landmark_files/55372/100230619.parquet",
+    "sign": "say",
+    "hand": "left",
+    "total_frames": 20,
+    "valid_frames": 19,
+    "nan_ratio": 0.05
+  },
+  {
+    "path": "train_landmark_files/55372/1002734054.parquet",
+    "sign": "that",
+    "hand": "left",
+    "total_frames": 6,
+    "valid_frames": 6,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/22343/1002776784.parquet",
+    "sign": "black",
+    "hand": "left",
+    "total_frames": 15,
+    "valid_frames": 9,
+    "nan_ratio": 0.4
+  },
+  {
+    "path": "train_landmark_files/37055/1003007869.parquet",
+    "sign": "moon",
+    "hand": "left",
+    "total_frames": 62,
+    "valid_frames": 4,
+    "nan_ratio": 0.9354838709677419
+  },
+  {
+    "path": "train_landmark_files/61333/1003093029.parquet",
+    "sign": "pizza",
+    "hand": "left",
+    "total_frames": 23,
+    "valid_frames": 18,
+    "nan_ratio": 0.21739130434782608
+  },
+  {
+    "path": "train_landmark_files/37055/1003109377.parquet",
+    "sign": "shhh",
+    "hand": "left",
+    "total_frames": 23,
+    "valid_frames": 20,
+    "nan_ratio": 0.13043478260869565
+  },
+  {
+    "path": "train_landmark_files/36257/1003335907.parquet",
+    "sign": "now",
+    "hand": "left",
+    "total_frames": 14,
+    "valid_frames": 8,
+    "nan_ratio": 0.42857142857142855
+  },
+  {
+    "path": "train_landmark_files/22343/1003347075.parquet",
+    "sign": "TV",
+    "hand": "left",
+    "total_frames": 89,
+    "valid_frames": 69,
+    "nan_ratio": 0.2247191011235955
+  },
+  {
+    "path": "train_landmark_files/55372/1003483016.parquet",
+    "sign": "jump",
+    "hand": "left",
+    "total_frames": 6,
+    "valid_frames": 4,
+    "nan_ratio": 0.3333333333333333
+  },
+  {
+    "path": "train_landmark_files/34503/1003483410.parquet",
+    "sign": "sleep",
+    "hand": "left",
+    "total_frames": 8,
+    "valid_frames": 8,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/37055/100393298.parquet",
+    "sign": "uncle",
+    "hand": "left",
+    "total_frames": 60,
+    "valid_frames": 38,
+    "nan_ratio": 0.36666666666666664
+  },
+  {
+    "path": "train_landmark_files/32319/1004200659.parquet",
+    "sign": "dryer",
+    "hand": "left",
+    "total_frames": 59,
+    "valid_frames": 33,
+    "nan_ratio": 0.4406779661016949
+  },
+  {
+    "path": "train_landmark_files/16069/10042041.parquet",
+    "sign": "green",
+    "hand": "left",
+    "total_frames": 105,
+    "valid_frames": 6,
+    "nan_ratio": 0.9428571428571428
+  },
+  {
+    "path": "train_landmark_files/16069/1004211348.parquet",
+    "sign": "bug",
+    "hand": "left",
+    "total_frames": 163,
+    "valid_frames": 97,
+    "nan_ratio": 0.4049079754601227
+  },
+  {
+    "path": "train_landmark_files/29302/1004285193.parquet",
+    "sign": "brother",
+    "hand": "left",
+    "total_frames": 20,
+    "valid_frames": 4,
+    "nan_ratio": 0.8
+  },
+  {
+    "path": "train_landmark_files/22343/1004302418.parquet",
+    "sign": "sad",
+    "hand": "left",
+    "total_frames": 34,
+    "valid_frames": 15,
+    "nan_ratio": 0.5588235294117647
+  },
+  {
+    "path": "train_landmark_files/16069/100438640.parquet",
+    "sign": "penny",
+    "hand": "left",
+    "total_frames": 75,
+    "valid_frames": 23,
+    "nan_ratio": 0.6933333333333334
+  },
+  {
+    "path": "train_landmark_files/27610/1004456612.parquet",
+    "sign": "mitten",
+    "hand": "left",
+    "total_frames": 43,
+    "valid_frames": 26,
+    "nan_ratio": 0.3953488372093023
+  },
+  {
+    "path": "train_landmark_files/55372/1004789884.parquet",
+    "sign": "brown",
+    "hand": "left",
+    "total_frames": 18,
+    "valid_frames": 11,
+    "nan_ratio": 0.3888888888888889
+  },
+  {
+    "path": "train_landmark_files/27610/100498908.parquet",
+    "sign": "drink",
+    "hand": "left",
+    "total_frames": 11,
+    "valid_frames": 11,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/16069/1005009451.parquet",
+    "sign": "stay",
+    "hand": "left",
+    "total_frames": 25,
+    "valid_frames": 14,
+    "nan_ratio": 0.44
+  },
+  {
+    "path": "train_landmark_files/55372/1005052722.parquet",
+    "sign": "tooth",
+    "hand": "left",
+    "total_frames": 20,
+    "valid_frames": 20,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/34503/1005088445.parquet",
+    "sign": "awake",
+    "hand": "left",
+    "total_frames": 8,
+    "valid_frames": 8,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/32319/100515039.parquet",
+    "sign": "hot",
+    "hand": "left",
+    "total_frames": 30,
+    "valid_frames": 17,
+    "nan_ratio": 0.43333333333333335
+  },
+  {
+    "path": "train_landmark_files/36257/1005189210.parquet",
+    "sign": "like",
+    "hand": "left",
+    "total_frames": 7,
+    "valid_frames": 7,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/16069/1005223850.parquet",
+    "sign": "where",
+    "hand": "left",
+    "total_frames": 15,
+    "valid_frames": 7,
+    "nan_ratio": 0.5333333333333333
+  },
+  {
+    "path": "train_landmark_files/36257/1005422610.parquet",
+    "sign": "potty",
+    "hand": "left",
+    "total_frames": 16,
+    "valid_frames": 16,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/61333/1005476876.parquet",
+    "sign": "down",
+    "hand": "left",
+    "total_frames": 9,
+    "valid_frames": 8,
+    "nan_ratio": 0.1111111111111111
+  },
+  {
+    "path": "train_landmark_files/16069/1005492440.parquet",
+    "sign": "old",
+    "hand": "left",
+    "total_frames": 76,
+    "valid_frames": 15,
+    "nan_ratio": 0.8026315789473685
+  },
+  {
+    "path": "train_landmark_files/37055/1005734995.parquet",
+    "sign": "no",
+    "hand": "left",
+    "total_frames": 6,
+    "valid_frames": 4,
+    "nan_ratio": 0.3333333333333333
+  },
+  {
+    "path": "train_landmark_files/61333/1005826772.parquet",
+    "sign": "head",
+    "hand": "left",
+    "total_frames": 26,
+    "valid_frames": 4,
+    "nan_ratio": 0.8461538461538461
+  },
+  {
+    "path": "train_landmark_files/22343/100585188.parquet",
+    "sign": "down",
+    "hand": "left",
+    "total_frames": 71,
+    "valid_frames": 23,
+    "nan_ratio": 0.676056338028169
+  },
+  {
+    "path": "train_landmark_files/32319/1005865446.parquet",
+    "sign": "orange",
+    "hand": "left",
+    "total_frames": 58,
+    "valid_frames": 36,
+    "nan_ratio": 0.3793103448275862
+  },
+  {
+    "path": "train_landmark_files/36257/1005866711.parquet",
+    "sign": "food",
+    "hand": "left",
+    "total_frames": 52,
+    "valid_frames": 47,
+    "nan_ratio": 0.09615384615384616
+  },
+  {
+    "path": "train_landmark_files/16069/1006058791.parquet",
+    "sign": "apple",
+    "hand": "left",
+    "total_frames": 14,
+    "valid_frames": 9,
+    "nan_ratio": 0.35714285714285715
+  },
+  {
+    "path": "train_landmark_files/37055/1006223282.parquet",
+    "sign": "pretty",
+    "hand": "left",
+    "total_frames": 27,
+    "valid_frames": 21,
+    "nan_ratio": 0.2222222222222222
+  },
+  {
+    "path": "train_landmark_files/55372/1006402771.parquet",
+    "sign": "nuts",
+    "hand": "left",
+    "total_frames": 18,
+    "valid_frames": 17,
+    "nan_ratio": 0.05555555555555555
+  },
+  {
+    "path": "train_landmark_files/16069/1006443357.parquet",
+    "sign": "animal",
+    "hand": "left",
+    "total_frames": 33,
+    "valid_frames": 23,
+    "nan_ratio": 0.30303030303030304
+  },
+  {
+    "path": "train_landmark_files/34503/1006554061.parquet",
+    "sign": "say",
+    "hand": "left",
+    "total_frames": 47,
+    "valid_frames": 6,
+    "nan_ratio": 0.8723404255319149
+  },
+  {
+    "path": "train_landmark_files/22343/1006584947.parquet",
+    "sign": "finish",
+    "hand": "left",
+    "total_frames": 26,
+    "valid_frames": 19,
+    "nan_ratio": 0.2692307692307692
+  },
+  {
+    "path": "train_landmark_files/37055/1006625592.parquet",
+    "sign": "beside",
+    "hand": "left",
+    "total_frames": 42,
+    "valid_frames": 4,
+    "nan_ratio": 0.9047619047619048
+  },
+  {
+    "path": "train_landmark_files/28656/1006698125.parquet",
+    "sign": "noisy",
+    "hand": "left",
+    "total_frames": 311,
+    "valid_frames": 273,
+    "nan_ratio": 0.12218649517684887
+  },
+  {
+    "path": "train_landmark_files/22343/1006778422.parquet",
+    "sign": "happy",
+    "hand": "left",
+    "total_frames": 36,
+    "valid_frames": 3,
+    "nan_ratio": 0.9166666666666666
+  },
+  {
+    "path": "train_landmark_files/34503/1006899334.parquet",
+    "sign": "stuck",
+    "hand": "left",
+    "total_frames": 225,
+    "valid_frames": 219,
+    "nan_ratio": 0.02666666666666667
+  },
+  {
+    "path": "train_landmark_files/27610/1006973536.parquet",
+    "sign": "bug",
+    "hand": "left",
+    "total_frames": 42,
+    "valid_frames": 41,
+    "nan_ratio": 0.023809523809523808
+  },
+  {
+    "path": "train_landmark_files/36257/1007054441.parquet",
+    "sign": "say",
+    "hand": "left",
+    "total_frames": 48,
+    "valid_frames": 48,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/16069/1007073382.parquet",
+    "sign": "sticky",
+    "hand": "left",
+    "total_frames": 76,
+    "valid_frames": 51,
+    "nan_ratio": 0.32894736842105265
+  },
+  {
+    "path": "train_landmark_files/16069/1007127288.parquet",
+    "sign": "owie",
+    "hand": "left",
+    "total_frames": 105,
+    "valid_frames": 47,
+    "nan_ratio": 0.5523809523809524
+  },
+  {
+    "path": "train_landmark_files/61333/1007285542.parquet",
+    "sign": "high",
+    "hand": "left",
+    "total_frames": 24,
+    "valid_frames": 21,
+    "nan_ratio": 0.125
+  },
+  {
+    "path": "train_landmark_files/55372/1007289801.parquet",
+    "sign": "fine",
+    "hand": "left",
+    "total_frames": 116,
+    "valid_frames": 43,
+    "nan_ratio": 0.6293103448275862
+  },
+  {
+    "path": "train_landmark_files/61333/1007294943.parquet",
+    "sign": "shirt",
+    "hand": "left",
+    "total_frames": 13,
+    "valid_frames": 11,
+    "nan_ratio": 0.15384615384615385
+  },
+  {
+    "path": "train_landmark_files/32319/1007313467.parquet",
+    "sign": "finish",
+    "hand": "left",
+    "total_frames": 19,
+    "valid_frames": 16,
+    "nan_ratio": 0.15789473684210525
+  },
+  {
+    "path": "train_landmark_files/16069/1007343357.parquet",
+    "sign": "boat",
+    "hand": "left",
+    "total_frames": 15,
+    "valid_frames": 9,
+    "nan_ratio": 0.4
+  },
+  {
+    "path": "train_landmark_files/32319/1007376023.parquet",
+    "sign": "all",
+    "hand": "left",
+    "total_frames": 39,
+    "valid_frames": 13,
+    "nan_ratio": 0.6666666666666666
+  },
+  {
+    "path": "train_landmark_files/36257/1007470312.parquet",
+    "sign": "pencil",
+    "hand": "left",
+    "total_frames": 8,
+    "valid_frames": 8,
+    "nan_ratio": 0.0
+  },
+  {
+    "path": "train_landmark_files/16069/1007589714.parquet",
+    "sign": "sleepy",
+    "hand": "left",
+    "total_frames": 12,
+    "valid_frames": 4,
+    "nan_ratio": 0.6666666666666666
+  },
+  {
+    "path": "train_landmark_files/32319/1007595679.parquet",
+    "sign": "moon",
+    "hand": "left",
+    "total_frames": 21,
+    "valid_frames": 16,
+    "nan_ratio": 0.23809523809523808
+  },
+  {
+    "path": "train_landmark_files/27610/1007714989.parquet",
+    "sign": "puzzle",
+    "hand": "left",
+    "total_frames": 145,
+    "valid_frames": 59,
+    "nan_ratio": 0.593103448275862
+  }
+]