写在前面

这是可以说的吗🫣🫣🫣

Efficient Training

调超参的时候可以先试试用Imagenet的100类试试，调的差不多了再上全部。

SNN Transformer Training

Transformer的常用Base lr是1e-4，adamw下是1e-4到6e-4之间，lamb可能需要再精调。adamw的前期收敛性太强了建议学习率linear warmup，用log怕炸

C-Optim

stdconv -> adaptive_clip clip_grad=0.02
正常的linear就torch自带的clip clip_grad=1 or 5

FasterViT?

Spike-driven Transformer V2/V3

检测与分割：除了backbone都是整数不norm

卷积算子貌似和整数（不带norm）很适配、attention貌似很和小数（带norm）适配

Spike-driven Transformer V1

About Finetuning SDT-V1 (Contributed by Qian S., 2025-04-28)

v1没有提供微调的代码，微调的方法：.yml文件不需要改（可能需要改一个epochs）

main.py找到create_model:

model = create_model(
        args.model,
        T=args.time_steps,
        pretrained=args.pretrained,
        drop_rate=args.drop,
        drop_path_rate=args.drop_path,
        drop_block_rate=args.drop_block,
        num_heads=args.num_heads,
        num_classes=args.num_classes,
        pooling_stat=args.pooling_stat,
        img_size_h=args.img_size,
        img_size_w=args.img_size,
        patch_size=args.patch_size,
        embed_dims=args.dim,
        mlp_ratios=args.mlp_ratio,
        in_channels=args.in_channels,
        qkv_bias=False,
        depths=args.layer,
        sr_ratios=1,
        spike_mode=args.spike_mode,
        dvs_mode=args.dvs_mode,
        TET=args.TET,
    )

把args.pretrained注释掉，加下面两行代码：

1 2	pretrained=True, checkpoint_path=your/path,

改后代码如下：

model = create_model(
        args.model,
        T=args.time_steps,
        pretrained=True,
        checkpoint_path='/data/users/sunq/S-Transformer/weights/cifar100_4_384.pth.tar',
        drop_rate=args.drop,
        drop_path_rate=args.drop_path,
        drop_block_rate=args.drop_block,
        num_heads=args.num_heads,
        num_classes=args.num_classes,
        pooling_stat=args.pooling_stat,
        img_size_h=args.img_size,
        img_size_w=args.img_size,
        patch_size=args.patch_size,
        embed_dims=args.dim,
        mlp_ratios=args.mlp_ratio,
        in_channels=args.in_channels,
        qkv_bias=False,
        depths=args.layer,
        sr_ratios=1,
        spike_mode=args.spike_mode,
        dvs_mode=args.dvs_mode,
        TET=args.TET,
    )

(Updated by Qian S., 2025-06-04)
当引入了新的可学习参数时，这种方法就不适用了，因为这个函数的默认strict=True。因此，可以这样做：

model = create_model(
        args.model,
        T=args.time_steps,
        # pretrained=args.pretrained,
        pretrained=False,
        drop_rate=args.drop,
        drop_path_rate=args.drop_path,
        drop_block_rate=args.drop_block,
        num_heads=args.num_heads,
        num_classes=args.num_classes,
        pooling_stat=args.pooling_stat,
        img_size_h=args.img_size,
        img_size_w=args.img_size,
        patch_size=args.patch_size,
        embed_dims=args.dim,
        mlp_ratios=args.mlp_ratio,
        in_channels=args.in_channels,
        qkv_bias=False,
        depths=args.layer,
        sr_ratios=1,
        spike_mode=args.spike_mode,
        dvs_mode=args.dvs_mode,
        TET=args.TET,
    )
    load_checkpoint(model, '/data/users/sunq/S-Transformer/weights/cifar10_4_384.pth.tar', strict=False)