scripts/gen_resnet34_notebook.py

#!/usr/bin/env python3
"""Generate ResNet34 notebook."""

import nbformat as nbf

nb = nbf.v4.new_notebook()
nb.metadata = {
    "kernelspec": {"display_name": "Python 3", "language": "python", "name": "python3"},
    "language_info": {"name": "python", "version": "3.12.0"},
}

cells = []
def md(s): cells.append(nbf.v4.new_markdown_cell(s))
def code(s): cells.append(nbf.v4.new_code_cell(s))

md("""\
# ResNet34: Advanced ResNet

Extended ResNet with [3,4,6,3] blocks, full CelebA (40 attributes), data augmentation.
""")

md("""\
## 背景

ResNet34 是 ResNet18 的升级版：从 [2,2,2,2] 扩展到 [3,4,6,3] 个残差块。
除了更深的网络，ResNet34 在此项目中还展示了更完整的训练流程：

- **SGD + Momentum** 替代 Adam（更通用的优化器）
- **CosineAnnealingLR** 学习率调度
- **数据增强**：随机翻转、颜色抖动、旋转
- **Loss 加权**：`pos_weight` 处理属性不平衡
- **Early stopping**：按验证 loss 保存最优模型
""")

md("""\
## 架构对比

```
ResNet18: [BasicBlock×2] → [BasicBlock×2] → [BasicBlock×2] → [BasicBlock×2]
ResNet34: [BasicBlock×3] → [BasicBlock×4] → [BasicBlock×6] → [BasicBlock×3]
```

ResNet34 的每个 `BasicBlock` 结构与 ResNet18 完全相同（2× Conv3×3 + BN + ReLU）。
差异只在于 block 数量。

> 本项目中的 ResNet34 直接从 `cv.resnet18.model` 复用 `ResNet` 类和 `BasicBlock`，
> 仅通过 `num_blocks` 参数实现架构升级。
""")

code("""\
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.data import DataLoader
from torchvision import transforms

from cv.resnet34.data import CelebADataset, CELEBA_ATTR_ORDER, train_transform
from cv.resnet34.model import resnet34
from utils.device import get_device

device = get_device()
print(f"Device: {device}")
""")

code("""\
# 使用 20K 子集控制 notebook 训练时间
# 全量 162K → 改为 20000
SUBSET_SIZE = 20000

train_dataset = CelebADataset(split="train", transform=train_transform())
# 截取前 SUBSET_SIZE 个样本
train_dataset.samples = train_dataset.samples[:SUBSET_SIZE]

val_dataset = CelebADataset(split="val", transform=train_transform())

train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True, num_workers=4, pin_memory=True)
val_loader = DataLoader(val_dataset, batch_size=64, shuffle=False, num_workers=4, pin_memory=True)

print(f"Train: {len(train_dataset):,}  Val: {len(val_dataset):,}")
print(f"Attributes: {len(CELEBA_ATTR_ORDER)}")
""")

code("""\
# 计算 pos_weight（从训练数据中统计正负样本比例）
print("Computing pos_weight...")
pos_counts = torch.zeros(40)
for _, labels in train_loader:
    pos_counts += labels.sum(dim=0)
neg_counts = len(train_dataset) - pos_counts
pos_weight = (neg_counts / pos_counts).clamp(min=1.0)
print(f"pos_weight range: [{pos_weight.min():.2f}, {pos_weight.max():.2f}]")
""")

code("""\
model = resnet34(num_classes=40).to(device)
print(f"Parameters: {sum(p.numel() for p in model.parameters()):,}")
""")

md("""\
## 训练

> ⏱ 预估耗时：**20 epoch × ~70s/epoch ≈ 23 分钟**（20K 子集, M4 Max, batch_size=64）
""")

code("""\
NUM_EPOCHS = 20
LR = 0.1
MOMENTUM = 0.9
WEIGHT_DECAY = 1e-4

criterion = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=MOMENTUM, weight_decay=WEIGHT_DECAY)
scheduler = CosineAnnealingLR(optimizer, T_max=NUM_EPOCHS)

train_loss_hist, val_loss_hist = [], []

for epoch in range(1, NUM_EPOCHS + 1):
    model.train()
    train_loss = 0.0
    for images, labels in train_loader:
        images, labels = images.to(device), labels.to(device)
        optimizer.zero_grad()
        outputs = model(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        train_loss += loss.item()
    scheduler.step()

    model.eval()
    val_loss = 0.0
    correct = total = 0
    with torch.no_grad():
        for images, labels in val_loader:
            images, labels = images.to(device), labels.to(device)
            outputs = model(images)
            loss = criterion(outputs, labels)
            val_loss += loss.item()
            preds = (torch.sigmoid(outputs) > 0.5).float()
            correct += (preds == labels).sum().item()
            total += labels.numel()

    avg_train = train_loss / len(train_loader)
    avg_val = val_loss / len(val_loader)
    acc = correct / total * 100
    train_loss_hist.append(avg_train)
    val_loss_hist.append(avg_val)
    print(f"Epoch [{epoch:2d}/{NUM_EPOCHS}]  Train: {avg_train:.4f}  Val: {avg_val:.4f}  Acc: {acc:.2f}%")
""")

md("""## Loss 曲线""")

code("""\
import matplotlib.pyplot as plt
from utils.device import get_device

plt.figure(figsize=(8, 4))
plt.plot(train_loss_hist, label='train', marker='o')
plt.plot(val_loss_hist, label='val', marker='o')
plt.xlabel("Epoch"); plt.ylabel("Loss"); plt.legend(); plt.grid(True)
plt.title("ResNet34 Training on CelebA (20K subset)"); plt.show()
""")

md("""\
## 思考题

1. SGD + Momentum 和 Adam 各自的优缺点是什么？什么场景下 SGD 更好？
2. `pos_weight` 的作用是什么？不使用时对哪些属性影响最大？
3. 数据增强（翻转、颜色抖动、旋转）为什么会提升泛化能力？
4. 把 `SUBSET_SIZE` 改到 162770（全量），训练 30 epoch 观察效果。
5. 对比 ResNet18 和 ResNet34 在此任务上的表现差异。
""")

nb.cells = cells
out = "cv/resnet34/resnet34.ipynb"
with open(out, "w") as f:
    nbf.write(nb, f)
print(f"Generated {out}")