hand and face again

This commit is contained in:
2026-01-18 14:43:14 -06:00
parent 716428ec0b
commit 9256050292
3 changed files with 551 additions and 647 deletions

View File

@@ -35,188 +35,222 @@ else:
print("=" * 60)
# ===============================
# 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 LOADING - HANDLES PARTIAL NaN
# ENHANCED DATA EXTRACTION
# ===============================
def load_kaggle_asl_data(base_path):
train_df = pd.read_csv(os.path.join(base_path, "train.csv"))
with open(os.path.join(base_path, "sign_to_prediction_index_map.json")) as f:
sign_to_idx = json.load(f)
return train_df, sign_to_idx
def extract_hand_landmarks_from_parquet(path):
"""Extract hand landmarks - ONLY uses frames with valid (non-NaN) data"""
def extract_multi_landmarks(path, min_valid_frames=5):
"""
Extract both hands + selected face landmarks with modality flags
Returns: dict with 'landmarks', 'left_hand_valid', 'right_hand_valid', 'face_valid'
"""
try:
df = pd.read_parquet(path)
seq = []
left_valid_frames = []
right_valid_frames = []
face_valid_frames = []
# Get hand data
left = df[df["type"] == "left_hand"]
right = df[df["type"] == "right_hand"]
if len(left) == 0 and len(right) == 0:
all_types = df["type"].unique()
if "left_hand" in all_types or "right_hand" in all_types or "face" in all_types:
frames = sorted(df["frame"].unique())
else:
return None
# 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:
if frames is None or len(frames) < min_valid_frames:
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]
frame_df = df[df["frame"] == frame]
frame_points = np.full((TOTAL_POINTS_PER_FRAME, 3), np.nan, dtype=np.float32)
# Count how many valid landmarks this frame has
valid_count = lm_frame[['x', 'y', 'z']].notna().all(axis=1).sum()
pos = 0
left_valid = False
right_valid = False
face_valid = False
# Skip frames with too few valid landmarks
if valid_count < 10:
continue
# Left hand
left = frame_df[frame_df["type"] == "left_hand"]
if len(left) >= 15:
valid_count = 0
for i in range(21):
row = left[left["landmark_index"] == i]
if len(row) > 0 and row[['x', 'y', 'z']].notna().all().all():
frame_points[pos] = row[['x', 'y', 'z']].values[0]
valid_count += 1
pos += 1
left_valid = (valid_count >= 15)
else:
pos += 21
# Extract landmarks for this frame
frame_landmarks = []
valid_landmarks_in_frame = 0
# Right hand
right = frame_df[frame_df["type"] == "right_hand"]
if len(right) >= 15:
valid_count = 0
for i in range(21):
row = right[right["landmark_index"] == i]
if len(row) > 0 and row[['x', 'y', 'z']].notna().all().all():
frame_points[pos] = row[['x', 'y', 'z']].values[0]
valid_count += 1
pos += 1
right_valid = (valid_count >= 15)
else:
pos += 21
for i in range(21):
lm = lm_frame[lm_frame['landmark_index'] == i]
# Face
face = frame_df[frame_df["type"] == "face"]
if len(face) > 0:
valid_count = 0
for idx in IMPORTANT_FACE_INDICES:
row = face[face["landmark_index"] == idx]
if len(row) > 0 and row[['x', 'y', 'z']].notna().all().all():
frame_points[pos] = row[['x', 'y', 'z']].values[0]
valid_count += 1
pos += 1
face_valid = (valid_count >= len(IMPORTANT_FACE_INDICES) * 0.5)
if len(lm) == 0:
frame_landmarks.append([0.0, 0.0, 0.0])
else:
x = float(lm['x'].iloc[0])
y = float(lm['y'].iloc[0])
z = float(lm['z'].iloc[0])
valid_ratio = 1 - np.isnan(frame_points).mean()
if valid_ratio >= 0.40:
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)
# Check if valid
if pd.notna(x) and pd.notna(y) and pd.notna(z):
frame_landmarks.append([x, y, z])
valid_landmarks_in_frame += 1
else:
frame_landmarks.append([0.0, 0.0, 0.0])
# Only add frame if it has enough valid landmarks
if valid_landmarks_in_frame >= 10:
landmarks_seq.append(frame_landmarks)
# Need at least 3 valid frames
if len(landmarks_seq) < 3:
if len(seq) < min_valid_frames:
return None
return np.array(landmarks_seq, dtype=np.float32)
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:
except Exception:
return None
def get_features_sequence(landmarks_seq, max_frames=100):
"""Extract features from landmark sequence"""
if landmarks_seq is None or len(landmarks_seq) == 0:
return None, None
# Center on wrist (landmark 0)
wrist = landmarks_seq[:, 0:1, :].copy()
landmarks_seq = landmarks_seq - wrist
# Scale normalization using wrist to middle finger MCP (landmark 9)
scale = np.linalg.norm(landmarks_seq[:, 9, :] - np.zeros(3), axis=1, keepdims=True)
scale = np.maximum(scale, 1e-6) # Avoid division by zero
landmarks_seq = landmarks_seq / scale[:, np.newaxis, :]
# Clean up any remaining NaN/Inf
landmarks_seq = np.nan_to_num(landmarks_seq, nan=0.0, posinf=0.0, neginf=0.0)
# Clip extreme values
landmarks_seq = np.clip(landmarks_seq, -10, 10)
# 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) # (T, 5)
# Temporal deltas (motion)
deltas = np.zeros_like(landmarks_seq)
if len(landmarks_seq) > 1:
deltas[1:] = landmarks_seq[1:] - landmarks_seq[:-1]
# Flatten each component separately, then concatenate
landmarks_flat = landmarks_seq.reshape(landmarks_seq.shape[0], -1) # (T, 63)
deltas_flat = deltas.reshape(deltas.shape[0], -1) # (T, 63)
# curl_features is already (T, 5)
# Combine: 63 + 63 + 5 = 131 features per frame
seq = np.concatenate([
landmarks_flat,
deltas_flat,
curl_features
], axis=1)
# 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 attention mask (True for valid positions)
valid_mask = np.zeros(max_frames, dtype=bool)
valid_mask[:min(T, max_frames)] = True
return seq.astype(np.float32), valid_mask
def process_row(row, base_path, max_frames=100):
"""Process a single row - worker function for multiprocessing"""
path = os.path.join(base_path, row["path"])
if not os.path.exists(path):
def get_features_sequence(landmarks_data, max_frames=100):
"""
Enhanced feature extraction with separate modality processing
"""
if landmarks_data is None:
return None, None, None
landmarks_3d = landmarks_data['landmarks']
if len(landmarks_3d) == 0:
return None, None, None
T, N, _ = landmarks_3d.shape
# Separate modalities for independent normalization
left_hand = landmarks_3d[:, :21, :]
right_hand = landmarks_3d[:, 21:42, :]
face = landmarks_3d[:, 42:, :]
# Independent centering per modality
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'])
]:
# Center on modality-specific mean
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)
# Combine all modalities
features = np.concatenate(features_list, axis=1)
# Create modality availability mask (which frames have which modalities)
modality_mask = np.stack([
landmarks_data['left_hand_valid'],
landmarks_data['right_hand_valid'],
landmarks_data['face_valid']
], axis=1).astype(np.float32) # (T, 3)
# 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
def process_row(row_data, base_path, max_frames=100):
"""Process a single row - expects tuple of (path, sign)"""
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:
# Extract landmarks
lm = extract_hand_landmarks_from_parquet(path)
if lm is None:
return None, None, None
lm_data = extract_multi_landmarks(path)
if lm_data is None:
return None, None, None, None
# Get features
feat, mask = get_features_sequence(lm, max_frames)
feat, frame_mask, modality_mask = get_features_sequence(lm_data, max_frames)
if feat is None:
return None, None, None
return None, None, None, None
# Final safety check
if np.isnan(feat).any() or np.isinf(feat).any():
return None, None, None
return feat, frame_mask, modality_mask, sign
return feat, mask, row["sign"]
except Exception as e:
return None, None, None
except Exception:
return None, None, None, None
# ===============================
# TRANSFORMER MODEL
# ENHANCED MODEL WITH MODALITY AWARENESS
# ===============================
class PositionalEncoding(nn.Module):
def __init__(self, d_model, max_len=128):
@@ -232,16 +266,19 @@ class PositionalEncoding(nn.Module):
return x + self.pe[:, :x.size(1)]
class TransformerASL(nn.Module):
def __init__(self, input_dim, num_classes, d_model=256, nhead=8, num_layers=4):
class ModalityAwareTransformer(nn.Module):
def __init__(self, input_dim, num_classes, d_model=384, nhead=8, num_layers=5):
super().__init__()
# Input projection
# Main projection
self.proj = nn.Linear(input_dim, d_model)
self.norm_in = nn.LayerNorm(d_model)
self.pos = PositionalEncoding(d_model, max_len=128)
# Transformer encoder
# 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,
@@ -253,30 +290,45 @@ 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),
nn.Linear(d_model, num_classes)
)
def forward(self, x, key_padding_mask=None):
# x: (batch, seq_len, input_dim)
# key_padding_mask: (batch, seq_len) - True for padding positions
def forward(self, x, modality_mask=None, key_padding_mask=None):
# Project features
x = self.proj(x)
# Add modality information if available
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)
# Global average pooling over valid positions
x = x.mean(dim=1)
# Weighted average (giving more weight to frames with more valid modalities)
if modality_mask is not None:
weights = modality_mask.sum(dim=-1, keepdim=True) + 1e-6 # (B, T, 1)
weights = weights / weights.sum(dim=1, keepdim=True)
x = (x * weights).sum(dim=1)
else:
x = x.mean(dim=1)
return self.head(x)
def load_kaggle_asl_data(base_path):
"""Load training metadata"""
train_path = os.path.join(base_path, "train.csv")
train_df = pd.read_csv(train_path)
return train_df, None
# ===============================
# MAIN TRAINING
# MAIN
# ===============================
def main():
base_path = "asl_kaggle"
@@ -284,163 +336,142 @@ def main():
MIN_SAMPLES_PER_CLASS = 5
print("\nLoading metadata...")
train_df, sign_to_idx = load_kaggle_asl_data(base_path)
print(f"Total sequences: {len(train_df)}")
train_df, _ = load_kaggle_asl_data(base_path)
rows = [row for _, row in train_df.iterrows()]
# Convert to simple tuples for multiprocessing compatibility
rows = [(row['path'], row['sign']) for _, row in train_df.iterrows()]
print("\nProcessing sequences (this will take a few minutes)...")
print("Expected: ~36,000 valid sequences based on diagnostic")
# Process with multiprocessing
print("\nProcessing sequences with BOTH hands + FACE (enhanced)...")
with Pool(cpu_count()) as pool:
results = list(tqdm(
pool.imap(
partial(process_row, base_path=base_path, max_frames=max_frames),
rows,
chunksize=100
chunksize=80
),
total=len(rows),
desc="Extracting landmarks"
desc="Landmarks extraction"
))
# Filter valid results
X_list, masks_list, y_list = [], [], []
for feat, mask, sign in results:
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)
X_list, frame_masks_list, modality_masks_list, y_list = [], [], [], []
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)
print(f"\n✓ Successfully extracted: {len(X_list)} valid sequences")
print(f" Success rate: {len(X_list) / len(train_df) * 100:.1f}%")
if len(X_list) < 100:
print("❌ Too few valid sequences found!")
print(" This shouldn't happen - please share this output for debugging")
if not X_list:
print("No valid sequences extracted!")
return
# Stack into arrays
X = np.stack(X_list)
masks = np.stack(masks_list)
frame_masks = np.stack(frame_masks_list)
modality_masks = np.stack(modality_masks_list)
print(f"\nData shape: {X.shape}")
print(f"Feature dimension: {X.shape[2]}")
print(f"\nExtracted {len(X):,} sequences")
print(f"Feature shape: {X.shape[1:]} (input_dim = {X.shape[2]})")
print(f"Modality mask shape: {modality_masks.shape}")
# Global normalization
print("Normalizing features...")
X = np.clip(X, -10.0, 10.0)
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
# Encode labels
# Labels
le = LabelEncoder()
y = le.fit_transform(y_list)
# Filter classes with too few samples
# Filter rare classes
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)
valid = [k for k, v in counts.items() if v >= MIN_SAMPLES_PER_CLASS]
mask = np.isin(y, valid)
X = X[mask_valid]
masks = masks[mask_valid]
y = y[mask_valid]
X = X[mask]
frame_masks = frame_masks[mask]
modality_masks = modality_masks[mask]
y = y[mask]
# Re-encode after filtering
le = LabelEncoder()
y = le.fit_transform(y)
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])}")
print(f"After filtering: {len(X):,} samples | {len(le.classes_)} classes")
# 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
# Analyze modality usage
print("\nModality statistics:")
print(f" Sequences with left hand: {(modality_masks[:, :, 0].sum(axis=1) > 0).mean() * 100:.1f}%")
print(f" Sequences with right hand: {(modality_masks[:, :, 1].sum(axis=1) > 0).mean() * 100:.1f}%")
print(f" Sequences with face: {(modality_masks[:, :, 2].sum(axis=1) > 0).mean() * 100:.1f}%")
# 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
)
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):
# Dataset
class ASLMultiDataset(Dataset):
def __init__(self, X, frame_masks, modality_masks, y):
self.X = torch.from_numpy(X).float()
self.masks = torch.from_numpy(masks)
self.frame_masks = torch.from_numpy(frame_masks).bool()
self.modality_masks = torch.from_numpy(modality_masks).float()
self.y = torch.from_numpy(y).long()
def __len__(self):
return len(self.X)
def __getitem__(self, idx):
return self.X[idx], self.masks[idx], self.y[idx]
return self.X[idx], self.frame_masks[idx], self.modality_masks[idx], self.y[idx]
# DataLoaders
batch_size = 128 if device.type == 'cuda' else 64
batch_size = 64 if device.type == 'cuda' else 32
train_loader = DataLoader(
ASLSequenceDataset(X_train, masks_train, y_train),
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True if device.type == 'cuda' else False
ASLMultiDataset(X_tr, fm_tr, mm_tr, y_tr),
batch_size=batch_size, shuffle=True,
num_workers=4, pin_memory=device.type == 'cuda'
)
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
ASLMultiDataset(X_te, fm_te, mm_te, y_te),
batch_size=batch_size * 2, shuffle=False,
num_workers=4, pin_memory=device.type == 'cuda'
)
# Initialize model
print("\nInitializing model...")
model = TransformerASL(
# Enhanced model
model = ModalityAwareTransformer(
input_dim=X.shape[2],
num_classes=len(le.classes_),
d_model=256,
d_model=384,
nhead=8,
num_layers=4
num_layers=5
).to(device)
total_params = sum(p.numel() for p in model.parameters())
print(f"Model parameters: {total_params:,}")
print(f"\nModel parameters: {sum(p.numel() for p in model.parameters()):,}")
# 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, T_mult=2)
optimizer = optim.AdamW(model.parameters(), lr=4e-4, weight_decay=1e-4)
scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10)
# Training loop
best_acc = 0.0
epochs = 60
print("\n" + "=" * 60)
print("STARTING TRAINING")
print("=" * 60)
epochs = 70
for epoch in range(epochs):
# Train
model.train()
total_loss = 0
correct = total = 0
total_loss = correct = total = 0
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)
for x, frame_mask, modality_mask, yb in tqdm(train_loader, desc=f"Epoch {epoch + 1}", leave=False):
x = x.to(device)
frame_mask = frame_mask.to(device)
modality_mask = modality_mask.to(device)
yb = yb.to(device)
# Invert mask: True for padding positions
key_mask = ~mask
key_padding_mask = ~frame_mask
optimizer.zero_grad(set_to_none=True)
logits = model(x, key_padding_mask=key_mask)
logits = model(x, modality_mask=modality_mask, key_padding_mask=key_padding_mask)
loss = criterion(logits, yb)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
total_loss += loss.item()
@@ -449,49 +480,35 @@ def main():
train_acc = correct / total * 100
# Evaluate
# Eval
model.eval()
correct = total = 0
with torch.no_grad():
for x, mask, yb in test_loader:
x, mask, yb = x.to(device), mask.to(device), yb.to(device)
key_mask = ~mask
logits = model(x, key_padding_mask=key_mask)
for x, frame_mask, modality_mask, yb in test_loader:
x = x.to(device)
frame_mask = frame_mask.to(device)
modality_mask = modality_mask.to(device)
yb = yb.to(device)
key_padding_mask = ~frame_mask
logits = model(x, modality_mask=modality_mask, key_padding_mask=key_padding_mask)
correct += (logits.argmax(-1) == yb).sum().item()
total += yb.size(0)
test_acc = correct / total * 100
scheduler.step()
# Print progress
print(f"[{epoch + 1:2d}/{epochs}] Loss: {total_loss / len(train_loader):.4f} | "
f"Train: {train_acc:.2f}% | Test: {test_acc:.2f}%", end="")
scheduler.step()
# Save best model
if test_acc > best_acc:
best_acc = test_acc
torch.save({
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'label_encoder_classes': le.classes_,
'acc': best_acc,
'epoch': epoch,
'input_dim': X.shape[2],
'num_classes': len(le.classes_),
'd_model': 256,
'nhead': 8,
'num_layers': 4
}, "best_asl_transformer.pth")
print(f" → New best: {best_acc:.2f}% ✓")
torch.save(model.state_dict(), "best_asl_modality_aware.pth")
print(" → saved")
else:
print()
print("\n" + "=" * 60)
print(f"✓ Training complete!")
print(f"✓ Best test accuracy: {best_acc:.2f}%")
print(f"✓ Model saved: best_asl_transformer.pth")
print("=" * 60)
print(f"\nBest test accuracy: {best_acc:.2f}%")
if __name__ == "__main__":