Closed Mark-Laohu closed 2 years ago
Hi, we visualize the final self.position
as the learned anchor points, which have self.num_position
points and each of them has two dimensions (x, y). You can also create them by the nn.Parameter(torch.FloatTensor(self.num_position, 2))
instead of the nn.Embedding(self.num_position, 2)
, which may be easier to understand.
谢谢回答,learned points生成我大概明白了,就是我想问的是learned points在训练后能可视化出来吗,paper中给的是初始化得到的图还是训练后得到的可视化图啊?
The figure in paper is the final learned anchor points after training. You can visualize the model.transformer.position.weight
in the checkpoint we provide.
好的,非常感谢!
This issue is not active for a long time and it will be closed in 5 days. Feel free to re-open it if you have further concerns.
你好,请问learned anchor points更新后的可视化图Figure 2(b)是怎么可视化出来的,不是被embedding了吗