Reparameterization need finetuning?

[Yinjie-ZHENG]

Thanks for your great work! I have noticed that reparameterization mentioned how to replace matmul with conv1x1. However, I wonder why we need "Reparameterized Training".

[wondervictor]

Hi @Yinjie-ZHENG, sorry for the late reply and please check docs/reparameterize for more details:

1. we provide the script to re-prameterize in one command.
2. the re-parameterized training is not required and you can use your custom embeddings for downstream embeddings with the re-parameterized models (zero-shot). If your datasets differ from the generic scenes, you can further re-parameterized training for higher performance.

[Yinjie-ZHENG]

Hi @Yinjie-ZHENG, sorry for the late reply and please check docs/reparameterize for more details:

1. we provide the script to re-prameterize in one command.
2. the re-parameterized training is not required and you can use your custom embeddings for downstream embeddings with the re-parameterized models (zero-shot). If your datasets differ from the generic scenes, you can further re-parameterized training for higher performance.

Thanks for your reply. But it is hard to find out how you re-parameterize the matmul to 1x1 conv, I mean, once you use 1x1 Conv, how do you set the weight of 1x1 Conv to keep the result right(without finetuning)

[jacksonthall22]

If your datasets differ from the generic scenes, you can further re-parameterized training for higher performance.

The docs here are pretty unclear at step 3. Could you please clarify the docs or elaborate on how to "Prepare the model config"? There are no real steps or instructions there, so I'm a little lost with how to deal with this code. Thanks in advance!

[JubSteven]

Same question. The docs are a little ambiguous about exactly how to "re-parameterized training for higher performance." Does the subsequent steps follow the same procedure as normal finetuning?