Spaces:
Sleeping
Sleeping
File size: 9,699 Bytes
0f049b3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 | # -------------------------------------------------------------
# This script handles the full training loop:
# 1️⃣ Train and validate a model for multiple epochs
# 2️⃣ Track accuracy, loss, learning rate, etc.
# 3️⃣ Save the best model (highest val accuracy)
# 4️⃣ Automatically stop early if val accuracy stops improving
# 5️⃣ Save training history and graphs
# -------------------------------------------------------------
import json, time
from pathlib import Path
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
from typing import Tuple, Dict, List
from torch.utils.tensorboard import SummaryWriter
from tqdm.auto import tqdm
# -------------------------------------------------------------
# HELPER: Compute batch accuracy
# -------------------------------------------------------------
def accuracy_from_logits(logits: torch.Tensor, y: torch.Tensor) -> float:
preds = torch.argmax(logits, dim=1)
return (preds == y).float().mean().item()
# -------------------------------------------------------------
# HELPER: Format seconds to mm:ss (for clean epoch timing)
# -------------------------------------------------------------
def _fmt_time(s: float) -> str:
m, s = divmod(int(s), 60)
return f"{m:02d}:{s:02d}"
# -------------------------------------------------------------
# HELPER: Get all target labels from a dataset
# -------------------------------------------------------------
def _get_targets(ds) -> torch.Tensor:
if hasattr(ds, "targets"):
return torch.as_tensor(ds.targets)
return torch.as_tensor([t for _, t in ds.samples])
# -------------------------------------------------------------
# HELPER: Count how many samples per class
# -------------------------------------------------------------
def _class_counts(ds) -> List[str]:
if not hasattr(ds, "classes"): return []
y = _get_targets(ds)
counts = torch.bincount(y, minlength=len(ds.classes)).tolist()
return [f"{cls}({cnt})" for cls, cnt in zip(ds.classes, counts)]
# -------------------------------------------------------------
# HELPER: Count total and trainable parameters
# -------------------------------------------------------------
def _num_params(model) -> Dict[str, int]:
total = sum(p.numel() for p in model.parameters())
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
return {"total": total, "trainable": trainable}
# -------------------------------------------------------------
# HELPER: Show current learning rate(s)
# -------------------------------------------------------------
def _current_lrs(optimizer) -> str:
lrs = sorted({pg["lr"] for pg in optimizer.param_groups})
if len(lrs) == 1: return f"{lrs[0]:.2e}"
return ", ".join(f"{lr:.2e}" for lr in lrs)
# -------------------------------------------------------------
# FUNCTION: Run one full epoch (Train OR Validation)
# -------------------------------------------------------------
def run_epoch(model,loader,criterion,optimizer,device,train: bool,epoch: int,epochs: int,) -> Tuple[float, float]:
model.train() if train else model.eval()
epoch_loss = 0.0
epoch_acc = 0.0
total = 0
# tqdm creates a nice progress bar in terminal
bar = tqdm(
loader,
desc=f"[{epoch:02d}/{epochs}] {'train' if train else 'val '}",
leave=False,
dynamic_ncols=True,
)
for xb, yb in bar:
xb = xb.to(device, non_blocking=True)
yb = yb.to(device, non_blocking=True)
if train:
optimizer.zero_grad()
with torch.set_grad_enabled(train):
out = model(xb) # Forward pass
loss = criterion(out, yb) # Compute loss
if train:
loss.backward() # Backpropagation
optimizer.step() # Update weights
# Update running totals
bsz = xb.size(0)
epoch_loss += loss.item() * bsz
epoch_acc += accuracy_from_logits(out, yb) * bsz
total += bsz
# Update progress bar text
bar.set_postfix(loss=f"{epoch_loss/max(total,1):.4f}",
acc=f"{epoch_acc/max(total,1):.4f}")
# Return average loss and accuracy
return epoch_loss / total, epoch_acc / total
# -------------------------------------------------------------
# FUNCTION: Save training curves (loss + accuracy)
# -------------------------------------------------------------
def _save_final_figure(history, best_val_acc, outpath="reports/curves_final.png", title="ResNet (Fine-tune)"):
Path(outpath).parent.mkdir(parents=True, exist_ok=True)
xs = range(1, len(history["train_loss"]) + 1)
fig, axs = plt.subplots(1, 2, figsize=(11, 4.5))
# ---- Loss subplot ----
axs[0].plot(xs, history["train_loss"], label="train")
axs[0].plot(xs, history["val_loss"], label="val")
axs[0].set_title("Loss"); axs[0].set_xlabel("Epoch"); axs[0].set_ylabel("Loss"); axs[0].legend()
# ---- Accuracy subplot ----
axs[1].plot(xs, history["train_acc"], label="train")
axs[1].plot(xs, history["val_acc"], label="val")
axs[1].set_title("Accuracy"); axs[1].set_xlabel("Epoch"); axs[1].set_ylabel("Acc"); axs[1].legend()
fig.suptitle(f"{title} | best val acc = {best_val_acc:.4f}")
fig.tight_layout(rect=[0, 0, 1, 0.95])
plt.savefig(outpath, dpi=200)
plt.close(fig)
# Save training history as JSON
with open("reports/history.json", "w") as f:
json.dump(history, f, indent=2)
# -------------------------------------------------------------
# MAIN FUNCTION: train_model
# -------------------------------------------------------------
def train_model(model, loaders, epochs, lr, device, patience=3):
# 1️⃣ Define loss, optimizer, and learning rate scheduler
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=1e-4)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, mode="max", factor=0.5, patience=1
)
writer = SummaryWriter("runs/recycle") # For TensorBoard visualization
# Print training setup summary
pstats = _num_params(model)
classes_line = ", ".join(_class_counts(loaders["train"].dataset)) or "(classes unavailable)"
print("\n" + "=" * 68)
print(" TRAINING START ")
print("-" * 68)
print(f" epochs: {epochs} | lr: {lr:.2e} | patience: {patience} | bs: {loaders['train'].batch_size}")
print(f" train batches: {len(loaders['train'])} | val batches: {len(loaders['val'])}")
print(f" params: {pstats['trainable']:,} trainable / {pstats['total']:,} total")
print(f" classes: {classes_line}")
print("=" * 68 + "\n")
# 2️⃣ Initialize trackers
best_val_acc, best_state, bad = 0.0, None, 0
history = {"train_loss": [], "val_loss": [], "train_acc": [], "val_acc": []}
saved_final = False
# 3️⃣ Training loop over all epochs
for ep in range(1, epochs + 1):
ep_t0 = time.time()
# Run training + validation
tr_loss, tr_acc = run_epoch(model, loaders["train"], criterion, optimizer, device, True, ep, epochs)
va_loss, va_acc = run_epoch(model, loaders["val"], criterion, optimizer, device, False, ep, epochs)
# Update scheduler based on validation accuracy
scheduler.step(va_acc)
# Store metrics in history
history["train_loss"].append(tr_loss); history["val_loss"].append(va_loss)
history["train_acc"].append(tr_acc); history["val_acc"].append(va_acc)
# Log to TensorBoard
writer.add_scalar("loss/train", tr_loss, ep)
writer.add_scalar("loss/val", va_loss, ep)
writer.add_scalar("acc/train", tr_acc, ep)
writer.add_scalar("acc/val", va_acc, ep)
writer.flush()
# Check for improvement
improved = va_acc > best_val_acc
if improved:
best_val_acc, best_state, bad = va_acc, model.state_dict(), 0
Path("models").mkdir(parents=True, exist_ok=True)
torch.save(best_state, "models/resnet18_best.pt")
else:
bad += 1 # No improvement counter
# Print summary for this epoch
print("\n" + "-" * 68)
print(f" EPOCH {ep:02d}/{epochs:02d} | time { _fmt_time(time.time()-ep_t0) } | lr { _current_lrs(optimizer) }")
print("-" * 68)
print(f" Train • loss {tr_loss:.4f} | acc {tr_acc:.4f}")
print(f" Val • loss {va_loss:.4f} | acc {va_acc:.4f} | " f"{'NEW BEST ✓' if improved else f'best {best_val_acc:.4f}'}")
if improved:
print(" Saved • models/resnet18_best.pt")
# Early stopping
if bad >= patience:
print("-" * 68)
print(f" EARLY STOP • no val acc improvement for {patience} epoch(s)")
print("-" * 68)
_save_final_figure(history, best_val_acc, outpath="reports/curves_final.png", title="ResNet (Fine-tune)")
saved_final = True
break
# 4️⃣ Load the best model weights
if best_state is not None:
model.load_state_dict(best_state)
# 5️⃣ Save final results if not done during early stop
if not saved_final:
_save_final_figure(history, best_val_acc, outpath="reports/curves_final.png", title="ResNet (Fine-tune)")
writer.close()
# 6️⃣ Summary printout
print("\n" + "=" * 68)
print(f" TRAINING DONE • best val acc = {best_val_acc:.4f}")
print(" Saved final curves: reports/curves_final.png")
print("=" * 68 + "\n")
return model, best_val_acc
|