176 lines
6.1 KiB
Python
176 lines
6.1 KiB
Python
import os
|
|
import polars as pl
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from torch.utils.data import TensorDataset, DataLoader
|
|
from concurrent.futures import ProcessPoolExecutor
|
|
from tqdm import tqdm
|
|
from sklearn.model_selection import train_test_split
|
|
|
|
# --- CONFIG ---
|
|
BASE_PATH = "asl_kaggle"
|
|
TARGET_FRAMES = 22
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
# --- DATA LOADING WITH RELATIVE FEATURES ---
|
|
|
|
def load_file_to_memory(path, base_path=BASE_PATH):
|
|
try:
|
|
parquet_path = os.path.join(base_path, path)
|
|
df = pl.read_parquet(parquet_path)
|
|
|
|
# 1. Global Anchor (Nose)
|
|
anchors = (
|
|
df.filter((pl.col("type") == "face") & (pl.col("landmark_index") == 0))
|
|
.select([
|
|
pl.col("frame"),
|
|
pl.col("x").alias("nx"),
|
|
pl.col("y").alias("ny"),
|
|
pl.col("z").alias("nz")
|
|
])
|
|
)
|
|
|
|
# 2. Local Anchors (Wrists)
|
|
# Left: 468, Right: 522
|
|
wrists = (
|
|
df.filter(pl.col("landmark_index").is_in([468, 522]))
|
|
.select([
|
|
pl.col("frame"),
|
|
pl.col("landmark_index"),
|
|
pl.col("x").alias("wx"),
|
|
pl.col("y").alias("wy")
|
|
])
|
|
)
|
|
|
|
processed = df.join(anchors, on="frame", how="left")
|
|
|
|
# Join wrist data to the main frame
|
|
# We use a left join on frame and landmark_index to align wrist coords with their rows
|
|
processed = (
|
|
processed.join(wrists, on=["frame", "landmark_index"], how="left")
|
|
.with_columns([
|
|
# Global (Nose-relative)
|
|
(pl.col("x") - pl.col("nx")).alias("x_g"),
|
|
(pl.col("y") - pl.col("ny")).alias("y_g"),
|
|
(pl.col("z") - pl.col("nz")).alias("z_g"),
|
|
# Local (Wrist-relative - defaults to global if not a hand point)
|
|
(pl.col("x") - pl.col("wx")).fill_null(pl.col("x") - pl.col("nx")).alias("x_l"),
|
|
(pl.col("y") - pl.col("wy")).fill_null(pl.col("y") - pl.col("ny")).alias("y_l"),
|
|
])
|
|
.sort(["frame", "type", "landmark_index"])
|
|
)
|
|
|
|
# We now have 5 channels: (x_g, y_g, z_g, x_l, y_l)
|
|
n_frames = processed["frame"].n_unique()
|
|
# Reshape to (Frames, 543 landmarks, 5 features)
|
|
tensor = processed.select(["x_g", "y_g", "z_g", "x_l", "y_l"]).to_numpy().reshape(n_frames, 543, 5)
|
|
|
|
# Temporal Resampling
|
|
indices = np.linspace(0, n_frames - 1, num=TARGET_FRAMES).round().astype(int)
|
|
return tensor[indices]
|
|
except Exception:
|
|
return np.zeros((TARGET_FRAMES, 543, 5))
|
|
|
|
|
|
# --- DUAL-STREAM MODEL ---
|
|
|
|
class ASLClassifier(nn.Module):
|
|
def __init__(self, num_classes):
|
|
super().__init__()
|
|
# 543 landmarks * 5 features per landmark = 2715
|
|
self.feat_dim = 543 * 5
|
|
|
|
self.conv1 = nn.Conv1d(self.feat_dim, 512, kernel_size=3, padding=1)
|
|
self.bn1 = nn.BatchNorm1d(512)
|
|
self.conv2 = nn.Conv1d(512, 512, kernel_size=3, padding=1)
|
|
self.bn2 = nn.BatchNorm1d(512)
|
|
|
|
self.pool = nn.MaxPool1d(2)
|
|
self.dropout = nn.Dropout(0.4)
|
|
|
|
self.fc = nn.Sequential(
|
|
nn.Linear(512, 1024),
|
|
nn.ReLU(),
|
|
nn.Dropout(0.2),
|
|
nn.Linear(1024, num_classes)
|
|
)
|
|
|
|
def forward(self, x):
|
|
# x shape: (Batch, 22, 543, 5)
|
|
b, t, l, f = x.shape
|
|
# Flatten landmarks and features into one vector, then transpose for Conv1d
|
|
x = x.view(b, t, -1).transpose(1, 2) # (Batch, 2715, 22)
|
|
|
|
x = F.relu(self.bn1(self.conv1(x)))
|
|
x = self.pool(x)
|
|
|
|
x = F.relu(self.bn2(self.conv2(x)))
|
|
x = self.pool(x)
|
|
|
|
# Global Average Pool across the time dimension
|
|
x = F.adaptive_avg_pool1d(x, 1).squeeze(-1)
|
|
|
|
return self.fc(self.dropout(x))
|
|
|
|
|
|
# --- EXECUTION ---
|
|
|
|
if __name__ == "__main__":
|
|
# 1. Setup Data
|
|
metadata = pl.read_csv(os.path.join(BASE_PATH, "train.csv"))
|
|
unique_signs = sorted(metadata["sign"].unique().to_list())
|
|
sign_to_idx = {sign: i for i, sign in enumerate(unique_signs)}
|
|
labels = [sign_to_idx[s] for s in metadata["sign"].to_list()]
|
|
|
|
paths = metadata["path"].to_list()
|
|
|
|
# 2. Load to RAM (Parallelized)
|
|
print(f"Loading {len(paths)} files into RAM with 5-channel features...")
|
|
with ProcessPoolExecutor() as executor:
|
|
data_list = list(tqdm(executor.map(load_file_to_memory, paths), total=len(paths)))
|
|
|
|
X = torch.tensor(np.array(data_list), dtype=torch.float32)
|
|
y = torch.tensor(labels, dtype=torch.long)
|
|
|
|
# 3. Split
|
|
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.1, stratify=y, random_state=42)
|
|
|
|
train_loader = DataLoader(TensorDataset(X_train, y_train), batch_size=64, shuffle=True)
|
|
val_loader = DataLoader(TensorDataset(X_val, y_val), batch_size=64)
|
|
|
|
# 4. Train
|
|
model = ASLClassifier(len(unique_signs)).to(device)
|
|
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
|
|
criterion = nn.CrossEntropyLoss(label_smoothing=0.1) # Helps prevent over-confidence
|
|
|
|
print(f"Starting training on {device}...")
|
|
for epoch in range(25):
|
|
model.train()
|
|
train_loss = 0
|
|
for batch_x, batch_y in tqdm(train_loader, desc=f"Epoch {epoch + 1}"):
|
|
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
|
|
|
|
optimizer.zero_grad()
|
|
output = model(batch_x)
|
|
loss = criterion(output, batch_y)
|
|
loss.backward()
|
|
optimizer.step()
|
|
train_loss += loss.item()
|
|
|
|
# Validation
|
|
model.eval()
|
|
correct, total = 0, 0
|
|
with torch.no_grad():
|
|
for vx, vy in val_loader:
|
|
vx, vy = vx.to(device), vy.to(device)
|
|
pred = model(vx).argmax(1)
|
|
correct += (pred == vy).sum().item()
|
|
total += vy.size(0)
|
|
|
|
print(f"Epoch {epoch + 1} | Loss: {train_loss / len(train_loader):.4f} | Val Acc: {100 * correct / total:.2f}%")
|
|
|
|
if (epoch + 1) % 5 == 0:
|
|
torch.save(model.state_dict(), f"asl_model_v2_e{epoch + 1}.pth") |