Swin学习笔记

# 从头开始读代码，学习论文中描述的点是如何用代码实现的。

Swin Transformer训练设置：

1. Regular ImageNet-1K training

optimizer: AdamW for 300 epochs

from torch import optim as optim

optimizer = optim.AdamW(parameters, eps=config.TRAIN.OPTIMIZER.EPS, betas=config.TRAIN.OPTIMIZER.BETAS, lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)

cosine decay learning rate scheduler and 20 epochs of linear warm-up.

A batch size of 1024, an initial learning rate of 0.001, and a weight decay of 0.05 are used.

2. Pretraining on ImageNet-22K and fine-tuning on ImageNet-1K.

消融实验因素：

1. 相对位置偏差（B）

\[{\rm{Attention}}(Q,K,V) = {\rm{SoftMax}}(\frac{ {Q{K^T}}}{ {\sqrt { {d_k}} }} + B)V\]

2. shifted windows

下采样结构（downsampling在前三个阶段的最后为PatchMerging&