311 lines
9.6 KiB
Python
311 lines
9.6 KiB
Python
# ===============================
|
|
# IMPORTS
|
|
# ===============================
|
|
import os
|
|
import json
|
|
import math
|
|
import numpy as np
|
|
import pandas as pd
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
import torch.optim as optim
|
|
from torch.utils.data import Dataset, DataLoader
|
|
from sklearn.model_selection import train_test_split
|
|
from sklearn.preprocessing import StandardScaler
|
|
from multiprocessing import Pool, cpu_count
|
|
from functools import partial
|
|
from tqdm import tqdm
|
|
|
|
# ===============================
|
|
# DEVICE
|
|
# ===============================
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
print(f"Using device: {device}")
|
|
if device.type == "cuda":
|
|
print("GPU:", torch.cuda.get_device_name(0))
|
|
torch.backends.cudnn.benchmark = True
|
|
|
|
# ===============================
|
|
# DATA LOADING
|
|
# ===============================
|
|
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):
|
|
try:
|
|
df = pd.read_parquet(path)
|
|
hand = df[df["type"].isin(["left_hand", "right_hand"])]
|
|
if len(hand) == 0:
|
|
return None
|
|
|
|
frames = sorted(hand['frame'].unique())
|
|
landmarks_seq = []
|
|
|
|
for frame in frames:
|
|
lm_frame = hand[hand['frame'] == frame]
|
|
lm_list = []
|
|
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])
|
|
else:
|
|
lm_list.append([lm['x'].values[0], lm['y'].values[0], lm['z'].values[0]])
|
|
landmarks_seq.append(lm_list)
|
|
|
|
return np.array(landmarks_seq, dtype=np.float32) # (T, 21, 3)
|
|
except:
|
|
return None
|
|
|
|
def get_features_sequence(landmarks_seq, max_frames=96):
|
|
if landmarks_seq is None or len(landmarks_seq) == 0:
|
|
return None
|
|
|
|
# Center on wrist (landmark 0)
|
|
landmarks_seq = landmarks_seq - landmarks_seq[:, 0:1, :]
|
|
|
|
# Rough scale normalization (using index finger length as reference)
|
|
scale = np.linalg.norm(landmarks_seq[:, 8] - landmarks_seq[:, 5], axis=1, keepdims=True)
|
|
scale = np.maximum(scale, 1e-6)
|
|
landmarks_seq /= scale
|
|
|
|
# Flatten → (T, 63)
|
|
seq = landmarks_seq.reshape(landmarks_seq.shape[0], -1)
|
|
|
|
# Pad / truncate
|
|
if len(seq) < max_frames:
|
|
pad = np.zeros((max_frames - len(seq), seq.shape[1]), dtype=np.float32)
|
|
seq = np.concatenate([seq, pad], axis=0)
|
|
else:
|
|
seq = seq[:max_frames]
|
|
|
|
return seq.astype(np.float32)
|
|
|
|
def process_row(row, base_path, max_frames=96):
|
|
path = os.path.join(base_path, row['path'])
|
|
if not os.path.exists(path):
|
|
return None, None
|
|
lm = extract_hand_landmarks_from_parquet(path)
|
|
feat = get_features_sequence(lm, max_frames)
|
|
if feat is None:
|
|
return None, None
|
|
return feat, row['sign']
|
|
|
|
# ===============================
|
|
# LOAD & PROCESS (with progress)
|
|
# ===============================
|
|
base_path = "asl_kaggle" # ← change if needed
|
|
train_df, sign_to_idx = load_kaggle_asl_data(base_path)
|
|
|
|
print("Processing videos...")
|
|
rows = [row for _, row in train_df.iterrows()]
|
|
|
|
with Pool(cpu_count()) as pool:
|
|
results = list(tqdm(pool.imap(
|
|
partial(process_row, base_path=base_path, max_frames=96),
|
|
rows
|
|
), total=len(rows)))
|
|
|
|
X, y = [], []
|
|
for feat, sign in results:
|
|
if feat is not None:
|
|
X.append(feat)
|
|
y.append(sign)
|
|
|
|
X = np.stack(X) # (N, T, 63)
|
|
print(f"Loaded {len(X)} valid samples | shape: {X.shape}")
|
|
|
|
# Global normalization (very important!)
|
|
scaler = StandardScaler()
|
|
X_reshaped = X.reshape(-1, X.shape[-1])
|
|
X_reshaped = scaler.fit_transform(X_reshaped)
|
|
X = X_reshaped.reshape(X.shape)
|
|
|
|
# ===============================
|
|
# LABELS
|
|
# ===============================
|
|
from sklearn.preprocessing import LabelEncoder
|
|
le = LabelEncoder()
|
|
y = le.fit_transform(y)
|
|
num_classes = len(le.classes_)
|
|
print(f"Classes: {num_classes}")
|
|
|
|
# ===============================
|
|
# SPLIT
|
|
# ===============================
|
|
X_train, X_test, y_train, y_test = train_test_split(
|
|
X, y, test_size=0.15, stratify=y, random_state=42
|
|
)
|
|
|
|
# ===============================
|
|
# DATASET + DATALOADER
|
|
# ===============================
|
|
class ASLSequenceDataset(Dataset):
|
|
def __init__(self, X, y):
|
|
self.X = torch.from_numpy(X).float()
|
|
self.y = torch.from_numpy(y).long()
|
|
|
|
def __len__(self):
|
|
return len(self.X)
|
|
|
|
def __getitem__(self, idx):
|
|
return self.X[idx], self.y[idx]
|
|
|
|
train_loader = DataLoader(ASLSequenceDataset(X_train, y_train),
|
|
batch_size=64, shuffle=True, num_workers=4, pin_memory=True)
|
|
test_loader = DataLoader(ASLSequenceDataset(X_test, y_test),
|
|
batch_size=96, shuffle=False, num_workers=4, pin_memory=True)
|
|
|
|
# ===============================
|
|
# MODEL
|
|
# ===============================
|
|
class PositionalEncoding(nn.Module):
|
|
def __init__(self, d_model, max_len=128):
|
|
super().__init__()
|
|
pe = torch.zeros(max_len, d_model)
|
|
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
|
|
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
|
|
pe[:, 0::2] = torch.sin(position * div_term)
|
|
pe[:, 1::2] = torch.cos(position * div_term)
|
|
self.register_buffer('pe', pe.unsqueeze(0))
|
|
|
|
def forward(self, x):
|
|
return x + self.pe[:, :x.size(1)]
|
|
|
|
class TransformerASL(nn.Module):
|
|
def __init__(self, input_dim=63, num_classes=250, d_model=192, nhead=6, num_layers=4):
|
|
super().__init__()
|
|
self.proj = nn.Linear(input_dim, d_model)
|
|
self.norm_in = nn.LayerNorm(d_model)
|
|
|
|
self.pos = PositionalEncoding(d_model)
|
|
|
|
encoder_layer = nn.TransformerEncoderLayer(
|
|
d_model=d_model,
|
|
nhead=nhead,
|
|
dim_feedforward=d_model*4,
|
|
dropout=0.15,
|
|
activation='gelu',
|
|
batch_first=True,
|
|
norm_first=True
|
|
)
|
|
self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
|
|
|
|
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 = self.proj(x)
|
|
x = self.norm_in(x)
|
|
x = self.pos(x)
|
|
|
|
x = self.encoder(x, src_key_padding_mask=key_padding_mask)
|
|
x = x.mean(dim=1) # global average pooling
|
|
x = self.head(x)
|
|
return x
|
|
|
|
model = TransformerASL(input_dim=63, num_classes=num_classes).to(device)
|
|
print(f"Model parameters: {sum(p.numel() for p in model.parameters()):,}")
|
|
|
|
# ===============================
|
|
# TRAINING SETUP
|
|
# ===============================
|
|
criterion = nn.CrossEntropyLoss(label_smoothing=0.05)
|
|
optimizer = optim.AdamW(model.parameters(), lr=8e-4, weight_decay=1e-4, betas=(0.9, 0.98))
|
|
scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=15, T_mult=2)
|
|
|
|
# ===============================
|
|
# TRAIN / EVAL
|
|
# ===============================
|
|
def create_padding_mask(seq_len, max_len):
|
|
# True = ignore this position
|
|
return torch.arange(max_len, device=device)[None, :] >= seq_len[:, None]
|
|
|
|
def train_epoch():
|
|
model.train()
|
|
total_loss = 0
|
|
correct = 0
|
|
total = 0
|
|
|
|
for x, y in tqdm(train_loader, desc="Train"):
|
|
x, y = x.to(device), y.to(device)
|
|
|
|
# Very simple length heuristic (can be improved later)
|
|
real_lengths = (x.abs().sum(dim=2) > 1e-6).sum(dim=1)
|
|
mask = create_padding_mask(real_lengths, x.size(1))
|
|
|
|
optimizer.zero_grad(set_to_none=True)
|
|
logits = model(x, key_padding_mask=mask)
|
|
|
|
loss = criterion(logits, y)
|
|
loss.backward()
|
|
|
|
# STRONG clipping — very important for landmarks
|
|
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=0.8)
|
|
|
|
optimizer.step()
|
|
|
|
total_loss += loss.item()
|
|
correct += (logits.argmax(dim=-1) == y).sum().item()
|
|
total += y.size(0)
|
|
|
|
# Debug exploding gradients
|
|
if torch.isnan(loss) or grad_norm > 50:
|
|
print(f"WARNING - NaN or huge grad! norm={grad_norm:.2f}")
|
|
|
|
return total_loss / len(train_loader), correct / total * 100
|
|
|
|
@torch.no_grad()
|
|
def evaluate():
|
|
model.eval()
|
|
correct = 0
|
|
total = 0
|
|
for x, y in test_loader:
|
|
x, y = x.to(device), y.to(device)
|
|
real_lengths = (x.abs().sum(dim=2) > 1e-6).sum(dim=1)
|
|
mask = create_padding_mask(real_lengths, x.size(1))
|
|
|
|
logits = model(x, key_padding_mask=mask)
|
|
correct += (logits.argmax(dim=-1) == y).sum().item()
|
|
total += y.size(0)
|
|
return correct / total * 100
|
|
|
|
# ===============================
|
|
# TRAINING LOOP
|
|
# ===============================
|
|
best_acc = 0
|
|
patience = 18
|
|
wait = 0
|
|
epochs = 80
|
|
|
|
for epoch in range(epochs):
|
|
loss, train_acc = train_epoch()
|
|
test_acc = evaluate()
|
|
|
|
print(f"[{epoch+1:2d}/{epochs}] loss: {loss:.4f} | train: {train_acc:.2f}% | test: {test_acc:.2f}%")
|
|
|
|
scheduler.step()
|
|
|
|
if test_acc > best_acc:
|
|
best_acc = test_acc
|
|
wait = 0
|
|
torch.save({
|
|
'model': model.state_dict(),
|
|
'optimizer': optimizer.state_dict(),
|
|
'scaler': scaler,
|
|
'label_encoder_classes': le.classes_
|
|
}, "best_asl_transformer.pth")
|
|
print("→ Saved new best model")
|
|
else:
|
|
wait += 1
|
|
if wait >= patience:
|
|
print("Early stopping triggered")
|
|
break
|
|
|
|
print(f"\nBest test accuracy achieved: {best_acc:.2f}%") |