Train 1 Report: CNN Baseline v1

1. Experiment Overview

2. Architecture (Reproducible Specification)

2.1 Overview

单帧 2D CNN，输入一张气象空间快照，经 6 层 ResBlock 逐步降采样后全局平均池化，最后由全连接层回归 6 个目标变量。架构为自定义设计。

Input (B, 42, 450, 449)
    |
    +-- Stem -----------------------------------------------------------------
    |    Conv2d(42, 64, kernel=7x7, stride=2, padding=3, bias=False)
    |    BatchNorm2d(64)
    |    ReLU(inplace=True)
    |    Output: (B, 64, 225, 225)
    |
    +-- Layer1: ResBlock(64 -> 64, stride=1)   -> (B, 64, 225, 225)
    +-- Layer2: ResBlock(64 -> 128, stride=2)  -> (B, 128, 113, 113)
    +-- Layer3: ResBlock(128 -> 256, stride=2) -> (B, 256, 57, 57)
    +-- Layer4: ResBlock(256 -> 256, stride=2) -> (B, 256, 29, 29)
    +-- Layer5: ResBlock(256 -> 512, stride=2) -> (B, 512, 15, 15)
    +-- Layer6: ResBlock(512 -> 512, stride=2) -> (B, 512, 8, 8)
    |
    +-- AdaptiveAvgPool2d(1) ------------------> (B, 512, 1, 1)
    |
    +-- Head -----------------------------------------------------------------
         Flatten                              -> (B, 512)
         Linear(512, 128)
         ReLU(inplace=True)
         Dropout(p=0.3)
         Linear(128, 6)                       -> (B, 6)

2.2 ResBlock Detail

每个 ResBlock 包含 2 个 3x3 卷积 + BatchNorm，以及一条 skip connection：

Input x (in_ch, H, W)
    |
    +-- Main path:
    |    Conv2d(in_ch, out_ch, 3x3, stride=stride, padding=1, bias=False)
    |    BatchNorm2d(out_ch)
    |    ReLU(inplace=True)
    |    Conv2d(out_ch, out_ch, 3x3, stride=1, padding=1, bias=False)
    |    BatchNorm2d(out_ch)
    |
    +-- Shortcut:
    |    if stride != 1 or in_ch != out_ch:
    |        Conv2d(in_ch, out_ch, 1x1, stride=stride, bias=False)
    |        BatchNorm2d(out_ch)
    |    else:
    |        Identity()
    |
    +-- Output = ReLU(main + shortcut)

2.3 Channel Progression

2.4 Design Choices

3. Optimization Configuration

3.1 Optimizer

3.2 Learning Rate Schedule

LR trajectory: lr(t) = 0.5 * lr_init * (1 + cos(pi * t / T_max))

3.3 Loss Function

criterion = nn.MSELoss()

• 对 6 个标准化后的目标变量统一计算 MSE，无 per-variable 加权
• 无独立分类 loss（AUC 通过 APCP 预测值的大小做 ranking score 计算，不参与训练）

3.4 Gradient Clipping

torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)

4. Regularization

4.1 Current Methods

4.2 Coverage Analysis

问题: 99.4% 的参数（卷积层）仅靠 weight decay (1e-4) 和 BN 的轻微正则效果。对 11.3M 参数 / ~20K 样本 (~565:1 参数/样本比) 而言不足以防止过拟合。

4.3 Not Used

Data augmentation, Spatial dropout / DropBlock, Label smoothing, Stochastic depth, Early stopping, Mixup / CutMix

5. Data Pipeline

5.1 Dataset

5.2 Data Split

5.3 Normalization

• Method: Per-channel z-score
• Computation: 从训练集随机采样 1000 个样本，计算每通道的空间均值/标准差
• Input: x_norm = (x - mean) / (std + 1e-7), mean/std shape (42, 1, 1)
• Target: y_norm = (y - mean) / (std + 1e-7), mean/std shape (6,)
• Persistence: 保存为 norm_stats.pt; 推理时从 checkpoint 加载

5.4 NaN Handling

5.5 DataLoader Config

6. Training Process

6.1 Per-Epoch Loop

for epoch in range(50):
    1. train_one_epoch():
       - model.train()
       - for batch in train_loader (1611 batches):
           - forward -> MSELoss -> NaN check -> backward -> clip_grad -> step
           - print every 200 batches
    2. validate():
       - model.eval() + no_grad()
       - collect all preds/targets/binary
       - compute val_loss (MSE)
    3. compute_metrics():
       - denormalize preds & targets
       - per-variable RMSE (filter NaN per variable)
       - conditional APCP RMSE (true > 2mm only)
       - AUC via roc_auc_score
    4. scheduler.step() (cosine)
    5. save latest.pt; if val_loss improved: save best.pt
    6. log CSV + update training_curves.png

6.2 Earlier Failed Runs

7. Results

7.1 Best Validation (Epoch 11) -- best.pt

7.2 Final Epoch (Epoch 49) -- latest.pt

7.3 Best Per-Variable Across All Epochs

7.4 Training Curves

Left (Loss): train 0.89->0.37, val 0.96->0.62 (epoch 11 后停滞), 过拟合明显。 Right (RMSE): RH ~15 主导; TMP ~3.5; 风速 ~2-3.4; APCP ~4.0; epoch 8-10 后全部收敛。

8. Evaluation Scope and Limitations

独立测试集评估与 persistence 基线对比尚未完成。第 11 轮保存点由验证损失选出，验证结果不代表独立测试结果。

训练记录显示过拟合、降雨 AUC 退化，以及数据中的 NaN batch 被跳过。统一 MSE 对不同变量的影响需要进一步分析。

10. Reproduction

10.1 Command

# On HPC (Tufts Pax cluster)
module load class/default
module load cs137/2026spring

cd <PROJECT_ROOT>
sbatch scripts/train.slurm  # or:
python train.py --model cnn_baseline --epochs 50 --batch_size 8 --lr 0.001 \
    --weight_decay 0.0001 --scheduler cosine --base_channels 64 --num_workers 2

10.2 Key Files