KeyError when using Segformer architecture in segmentation_models_pytorch

[essawey]

Description:

I encountered an error while trying to use the "Segformer" architecture in segmentation_models_pytorch as shown below:

import segmentation_models_pytorch as smp
model = smp.create_model(arch="Segformer")

Error:
KeyError: "Wrong architecture type Segformer. Available options are: ['unet', 'unetplusplus', 'manet', 'linknet', 'fpn', 'pspnet', 'deeplabv3', 'deeplabv3plus', 'pan']"

It appears that the "Segformer" architecture is not linked to the library. Based on the error message, only the listed architectures are currently available for use.

Environment:

segmentation_models_pytorch version: 0.3.4
Python version: 3.8

I believe the architecture might not be integrated into the library yet.
Please let me know if there are any updates regarding this,
if there’s another way to use Segformer with this library.

Thanks <3

[qubvel]

Hey! Segformer is not yet in the release, please install from source. The release planned in a week

[Alberto1404]

Hey @qubvel, so great work with this repository!
I was wondering about the posibility to use the segformer with original weights from Nvidia, calling create_model directly.
Moreover, I attempted to finetune the Segformer in a custom dataset for binary segmentation, but it is somehow not taking the ignore_mismatched_sizes = True argument.
Perhaps it is not implemented? Or is it not expected to do that?
Thank you in advance!!

[qubvel]

Hi @Alberto1404, thanks for reporting this! Yes, it's not implemented yet, but it would be great to have this option! The feature should be fairly simple to implement. Let me know if you have the bandwidth to contribute, otherwise, I will take care of it 🤗

[Alberto1404]

Hi @qubvel, I attempted to spend some time during weekend, as the following days I won't be able to do anything. If you don't mind taking care of it, I will be highly appreciated.
Thank you

[qubvel]

Hey @Alberto1404, the PR is up

• Load model with mismatched sizes #1107

[Alberto1404]

Awesome!