When to use the freeze_patch_emb method about van-classification

[exiawsh]

There are also questions about the design of layer_scale.

VAN-Classification/models/van.py

Line 96 in d94f156

layer_scale_init_value = 1e-2

After the training, its operation belongs to a static process. What is its original design intention? Why not use a structure similar to SENet to dynamically determine the weight?

[MenghaoGuo]

We adopte the same strategy with this paper.

In practice, we found we can achieve better results with dynamic weights, but it introduced some additional computational and memory overhead

[exiawsh]

We adopte the same strategy with this paper.

In practice, we found we can achieve better results with dynamic weights, but it introduced some additional computational and memory overhead

thx