Initial commit: handshapes multiclass project

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
2026-01-19 22:27:20 -05:00
commit 816e34cb17
22 changed files with 2820 additions and 0 deletions

137
doc/train_seq.py Normal file
View File

@@ -0,0 +1,137 @@
#!/usr/bin/env python3
# Train BiGRU on (T, F) sequences; reads input_dim from meta.json
import os, json, argparse # stdlib
import numpy as np # arrays
import torch, torch.nn as nn # model/ops
from torch.utils.data import Dataset, DataLoader # data pipeline
def get_device():
"""
Prefer Apple Silicon's MPS if available; fallback to CPU/GPU accordingly.
"""
return torch.device("mps") if torch.backends.mps.is_available() else torch.device("cpu")
class SeqDataset(Dataset):
"""
Simple dataset wrapper with optional light augmentation.
"""
def __init__(self, X, y, augment=False):
self.X = X.astype(np.float32) # ensure float32 features
self.y = y.astype(np.int64) # class indices as int64
self.augment = augment
def __len__(self): return len(self.y) # number of samples
def _augment(self, seq):
# Add tiny Gaussian noise; helpful regularizer for high-D continuous features.
return seq + np.random.normal(0, 0.01, size=seq.shape).astype(np.float32)
def __getitem__(self, i):
xi = self.X[i] # (T, F)
if self.augment: xi = self._augment(xi) # optional noise
return torch.from_numpy(xi).float(), int(self.y[i]) # return (tensor, label)
class SeqGRU(nn.Module):
"""
BiGRU → MLP head classifier.
Uses last time step of GRU outputs (many-to-one).
"""
def __init__(self, input_dim, hidden=128, num_classes=26):
super().__init__()
self.gru = nn.GRU(input_dim, hidden, batch_first=True, bidirectional=True)
self.head = nn.Sequential(
nn.Linear(hidden*2, 128), nn.ReLU(), nn.Dropout(0.2),
nn.Linear(128, num_classes),
)
def forward(self, x):
h,_ = self.gru(x) # (B,T,2H)
return self.head(h[:, -1, :]) # logits (B,C)
def main():
"""
Train loop:
- Load prepared dataset
- Compute global mean/std on train and normalize train/val
- Train BiGRU with AdamW + cosine schedule
- Save best checkpoint by val accuracy (includes mean/std)
"""
ap = argparse.ArgumentParser()
ap.add_argument("--landmarks", default="landmarks_seq32") # dataset folder
ap.add_argument("--epochs", type=int, default=40)
ap.add_argument("--batch", type=int, default=64)
ap.add_argument("--lr", type=float, default=1e-3)
ap.add_argument("--out", default="asl_seq32_gru.pt") # model save path
args = ap.parse_args()
trX = np.load(os.path.join(args.landmarks,"train_X.npy")) # (Ntr, T, F)
trY = np.load(os.path.join(args.landmarks,"train_y.npy")) # (Ntr,)
vaX = np.load(os.path.join(args.landmarks,"val_X.npy")) # (Nva, T, F)
vaY = np.load(os.path.join(args.landmarks,"val_y.npy")) # (Nva,)
classes = json.load(open(os.path.join(args.landmarks,"class_names.json")))
meta = json.load(open(os.path.join(args.landmarks,"meta.json")))
T = int(meta["frames"]) # #frames per clip
input_dim = int(meta.get("input_dim", trX.shape[-1])) # feature dim (safety)
print(f"Loaded: train {trX.shape} val {vaX.shape} classes={classes} input_dim={input_dim}")
# Global normalization (feature-wise) computed on TRAIN ONLY
X_mean = trX.reshape(-1, trX.shape[-1]).mean(axis=0, keepdims=True).astype(np.float32) # (1,F)
X_std = trX.reshape(-1, trX.shape[-1]).std(axis=0, keepdims=True).astype(np.float32) + 1e-6
trXn = (trX - X_mean) / X_std # normalize train
vaXn = (vaX - X_mean) / X_std # normalize val using train stats
tr_ds = SeqDataset(trXn, trY, augment=True) # datasets
va_ds = SeqDataset(vaXn, vaY, augment=False)
tr_dl = DataLoader(tr_ds, batch_size=args.batch, shuffle=True) # loaders
va_dl = DataLoader(va_ds, batch_size=args.batch, shuffle=False)
device = get_device() # target device
model = SeqGRU(input_dim=input_dim, hidden=128, num_classes=len(classes)).to(device)
crit = nn.CrossEntropyLoss() # standard multi-class loss
opt = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=1e-4) # AdamW helps generalization
sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=args.epochs) # smooth LR decay
best_acc, best_state = 0.0, None # track best val acc
for epoch in range(1, args.epochs+1):
model.train()
tot, correct, loss_sum = 0, 0, 0.0
for xb, yb in tr_dl:
xb, yb = xb.to(device), yb.to(device) # move to device
opt.zero_grad(set_to_none=True) # reset grads
logits = model(xb) # forward
loss = crit(logits, yb) # compute loss
loss.backward() # backprop
opt.step() # update weights
loss_sum += loss.item() * yb.size(0) # accumulate loss
correct += (logits.argmax(1)==yb).sum().item() # count train correct
tot += yb.size(0) # sample counter
tr_loss = loss_sum / max(1, tot)
tr_acc = correct / max(1, tot)
model.eval()
vtot, vcorrect = 0, 0
with torch.no_grad():
for xb, yb in va_dl:
xb, yb = xb.to(device), yb.to(device)
logits = model(xb)
vcorrect += (logits.argmax(1)==yb).sum().item()
vtot += yb.size(0)
va_acc = vcorrect / max(1, vtot) # validation accuracy
sch.step() # update LR schedule
print(f"Epoch {epoch:02d}: train_loss={tr_loss:.4f} train_acc={tr_acc:.3f} val_acc={va_acc:.3f}")
if va_acc > best_acc: # save best checkpoint
best_acc = va_acc
best_state = {
"model": model.state_dict(),
"classes": classes,
"frames": T,
"X_mean": torch.from_numpy(X_mean),
"X_std": torch.from_numpy(X_std),
}
torch.save(best_state, args.out)
print(f" ✅ Saved best → {args.out} (val_acc={best_acc:.3f})")
print("Done. Best val_acc:", best_acc)
if __name__ == "__main__":
main()