训练一个图像分类器

tp new my-cnn --method pytorch/image-classification
tp dataset push my-images v1 ./images
tp submit my-cnn --profile h100:1 --train-dataset alice/my-images@v1

train/scratch/0001.png      这张图的标签是 scratch
train/dent/0002.png
train/ok/0003.png
val/scratch/0101.png        验证集，同样的类名
val/ok/0102.png

tp new my-cnn --method pytorch/image-classification

input_pipeline: imagefolder   # 读 <data_dir>/{train,val}/<class>/*
num_classes: 3
arch: simple_cnn              # 或任何 torchvision.models 的构造器，比如 resnet18
epochs: 5
batch_size: 64                # 每进程；全局 batch 是它乘卡数
learning_rate: 1.0e-3
optimizer: adamw              # adamw | sgd
lr_scheduler: cosine          # constant | step | cosine
amp: true

tp submit my-cnn --profile h100:1 -s learning_rate=5e-4 -s epochs=10

tp submit my-cnn --profile h100:1     # 一张卡
tp submit my-cnn --profile h100:8     # 单节点八张

def build_model(config, in_channels):
    """返回任意 nn.Module。"""

def build_datasets(config, data_dir, *, may_download=True):
    """返回 (train_set, val_set, in_channels)。"""