AttributeError: 'SegformerForSemanticSegmentation' object has no attribute 'load_weights'. Did you mean: 'init_weights'?

[ddbuscombe]

The transformers library now uses SegformerForSemanticSegmentation instead of TFSegformerForSemanticSegmentation. I have updated doodleverse-utils to accommodate this change. I reinstalled the pingmapper and rockmapper envs and it looks like it is pulling the latest version (0.0.43)

So, pingseg needs to be updated, specifically seg_gym_folder imported as from pingseg.seg_gym import seg_gym_folder

I'm happy to try to contribute this change in https://github.com/PINGEcosystem/PINGSeg

[CameronBodine]

By all means, help yourself! Thanks! - Cam

…

On Fri, Mar 6, 2026 at 4:24 PM Daniel Buscombe ***@***.***> wrote: *dbuscombe-usgs* created an issue (PINGEcosystem/RockMapper#8) <#8> The transformers library now uses SegformerForSemanticSegmentation instead of TFSegformerForSemanticSegmentation. I have updated doodleverse-utils to accommodate this change. I reinstalled the pingmapper and rockmapper envs and it looks like it is pulling the latest version (0.0.43) image.png (view on web) <https://github.com/user-attachments/assets/7bf249b7-47cf-49f7-b72e-8ded81b9c628> So, pingseg needs to be updated, specifically seg_gym_folder imported as from pingseg.seg_gym import seg_gym_folder I'm happy to try to contribute this change in https://github.com/PINGEcosystem/PINGSeg — Reply to this email directly, view it on GitHub <#8>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AM5DMX7D2OKUWOWMFSD6GET4PM6WZAVCNFSM6AAAAACWJ6LDF2VHI2DSMVQWIX3LMV43ASLTON2WKOZUGAZTMMRTGM3TKMI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[ddbuscombe]

It would appear that to use custom weights with the new transformers/SegformerForSemanticSegmentation the weights would need to be provided at model compilation, e.g.

from transformers import SegformerForSemanticSegmentation

# Load model with custom classification head weights
model = SegformerForSemanticSegmentation.from_pretrained(
    "./path_to_custom_weights",
    num_labels=10, # Example: new number of labels
    ignore_mismatched_sizes=True # Set to True to overwrite head weights
)

so we'd need to make this change in both doodleverse-utils and pingseg (in fact, throughout the doodleverse, UGH). In doodleverse-utils/model_inputs/segformer, we'd need to pass the model weights directly to the function

I have my fingers crossed that the models are backwards-compatible!! They may not be ...

[ddbuscombe]

The 'model checkpoint' would go, to be replaced with the model weights file -- I have a feeling this will NOT work

[ddbuscombe]

I think the whole doodleverse needs an overhaul, which will be a heavy lift. I think a hackathon is called for ... or just abandon segformer models

[CameronBodine]

I would be up for a hackathon if we could figure out a good time for it. TBH, and I know you won’t like it, but copilot would be of great assistance here. - Cam

…

On Fri, Mar 6, 2026 at 5:42 PM Daniel Buscombe ***@***.***> wrote: *dbuscombe-usgs* left a comment (PINGEcosystem/RockMapper#8) <#8 (comment)> I think the whole doodleverse needs an overhaul, which will be a heavy lift. I think a hackathon is called for ... — Reply to this email directly, view it on GitHub <#8 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AM5DMX3JCKTXARUE3D4DJG34PNH5RAVCNFSM6AAAAACWJ6LDF2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DAMJUGU2TCNJWGA> . You are receiving this because you commented.Message ID: ***@***.***>

[2320sharon]

Hey @dbuscombe-usgs
I'm testing out some environment that use tensorflow and the transforms library.
I'm currently trying out tranformers version 5.3 and tensorflow 2.15 with the latest doodleverse_utils.
I'm getting the error message TF equivalent of the class you tried to import would be "TFSegformerForSemanticSegmentation". If you want to use TensorFlow, please use TF classes instead! but when I try to import TFSegformerForSemanticSegmentation its not available at all. Have you run into this error before?

I just wanted to put this issue on your radar just in case you or someone you knows runs into this issue.

I think I can fix this by downgrading the transformers library to an earlier version that did not have this issue.

[ddbuscombe]

Hi Sharon, yes the last update to doodleverse-utils accommodated this change from TFSegformerForSemanticSegmentation to SegformerForSemanticSegmentation, but is only partially implemented. This is because SegformerForSemanticSegmentation doesn't load weights from h5 file like previous versions.

Downgrading the transformers library to 4.56.2 seems to work for now. I also am using doodleverse_utils==0.0.39 just in case

In the longer term I feel like there are changes to the doodleverse-utils codebase that need to happen so we can use the newer transformers/SegformerForSemanticSegmentation but it is too much of a heavy lift for me at the moment.

[2320sharon]

Thanks for explaining why SegformerForSemanticSegmentation wouldn't load the h5 files.
I did try converting the segformer models in doodleverse_utils to onnx format to make them easier to use but unfortunately there are some unsupported functions that onnx did not support in the model architecture.