About the version of forward_type

[aifeixingdelv]

Thank you very much for your contribution. I have some question, really looking forward to your reply

1. I want to know if the latest forward_type version is "xv". Can I use it stability?
2. Can you describe the difference of "xv1a", "xv2a", "xv3a"? Which will you recommend？
3. Which initialize will you recommend to get the best outputs from initialize option "v0"、"v1" or "v2"

[MzeroMiko]

1. xv shows less stable training in base scale.
2. the difference between xv1a xv2a and xv3a is the way how dts is generated.
3. in my ablations, v0 v1 v2 shows a little difference. For details you can see the arxiv paper.

[aifeixingdelv]

1. xv shows less stable training in base scale.

2. the difference between xv1a xv2a and xv3a is the way how dts is generated.

3. in my ablations, v0 v1 v2 shows a little difference. For details you can see the arxiv paper.

4. xv shows less stable training in base scale.

5. the difference between xv1a xv2a and xv3a is the way how dts is generated.

6. in my ablations, v0 v1 v2 shows a little difference. For details you can see the arxiv paper.

Thank you for your rapid reply!