Using the pre-trained weights (ImageNet-1K) for encoder-only initialization

[Nikolai10]

Hello @deep-diver, @sayakpaul,

great job implementing Segformer in TensorFlow!

I would like to train the model from scratch, using the on the Imagenet-1K pre-trained encoder weights (https://github.com/NVlabs/SegFormer -> section Training), as described in https://arxiv.org/pdf/2105.15203.pdf (section 4.1, implementational details).

Could you please tell me briefly if this is possible and what steps are required for this?
This would be very helpful!

Thanks,
Nikolai

[sayakpaul]

We cannot provide these details, sorry. I suggest you follow the instructions and contact the SegFormer authors directly for this.