Missing labels in loss kwargs causing issue for electra model / Issues from PR#130

When I try use an electra checkpoint for prediction, the recently merged code in the PR #130 causes an issue.

  File "chebai/result/prediction.py", line 171, in <module>
    CLI(MainPredictor, as_positional=False)
  File "/home/staff/a/akhedekar/miniconda3/envs/gnn/lib/python3.10/site-packages/jsonargparse/_cli.py", line 23, in CLI
    return auto_cli(*args, _stacklevel=3, **kwargs)
  File "/home/staff/a/akhedekar/miniconda3/envs/gnn/lib/python3.10/site-packages/jsonargparse/_cli.py", line 102, in auto_cli
    return _run_component(components, init)
  File "/home/staff/a/akhedekar/miniconda3/envs/gnn/lib/python3.10/site-packages/jsonargparse/_cli.py", line 210, in _run_component
    return component(**cfg)
  File "chebai/result/prediction.py", line 152, in predict_from_file
    predictor.predict_from_file(
  File "chebai/result/prediction.py", line 89, in predict_from_file
    preds: torch.Tensor = self.predict_smiles(smiles=smiles_strings)
  File "/home/staff/a/akhedekar/miniconda3/envs/gnn/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "chebai/result/prediction.py", line 128, in predict_smiles
    self._model.predict_step(batch, batch_idx, dm_hparams=self._dm_hparams)
  File "/home/staff/a/akhedekar/python-chebai/chebai/models/base.py", line 247, in predict_step
    pr, _ = self._get_prediction_and_labels(data, labels, model_output)
  File "/home/staff/a/akhedekar/python-chebai/chebai/models/electra.py", line 330, in _get_prediction_and_labels
    d = d * (~missing_labels).int().to(
RuntimeError: The size of tensor a (1528) must match the size of tensor b (2) at non-singleton dimension 1

Additional Information

shape of missing labels is [2, 2]

with following values. [[False, False], [False, False]]

the missing labels are added using collator with new logic (#130). Relevant part of code below:

            missing_labels = [
                d.get("missing_labels", [False for _ in y[0]]) for d in data
            ]

Also, the pass_loss_kwargs for electra was set to True instead of accepting value from the user or method parameter. Is there any particular reason for it with new changes introduced ?
https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-9b4cba0f4cb875af21676f1998e61cd7b25a1e65d7b971a00ce88d17df880706R244

Also in electra config, the model type is set to regression, shouldn't it be the original classification version by default.

https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-286e6c43d041243386413f98ba9f93eca188d678d39b19a40450c57df5ab9979R3

Also, I wonder whether the issue #53 was also addressed in that PR. As I see files and tests related to tox changed.

Default beta value for BCE is set to 1000

Is this was the intended change ?

https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-00639a90075191fb4a6de89627b1181de9a0d44451fbf239c0a96e9df6750280R2-R3

I wonder why the following line as deleted ?
https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-5a15c77150d45ed1d4be890d0dc4808bfc54d5df4a9a998ddedf9ab583d7c3c3L276

The reason for this question is that I am concerned that deletion and the below error might be related.
https://wandb.ai/chebai/chebai/runs/m74gxlm8/logs?nw=nwuseraditya0by0

SLURM auto-requeueing enabled. Setting signal handlers.
Sanity Checking: |          | 0/? [00:00<?, ?it/s]
Loading validation data... (datamodule.current_epoch=N/A)
Loading splits from data/chebi_v241/ChEBI50/processed/splits.csv...
Sanity Checking DataLoader 0:   0%|          | 0/2 [00:00<?, ?it/s]Computing loss-weights based on v241 dataset (beta=1000.0)
Epoch 0: 100%|██████████| 5017/5017 [02:25<00:00, 34.43it/s, v_num=0153, train_loss_step=nan.0, val_loss_step=nan.0, val_loss_epoch=nan.0, val_macro-f1=0.000, val_micro-f1=0.000, train_loss_epoch=nan.0, train_macro-f1=2.62e-5, train_micro-f1=0.000748]
Loading train data... (datamodule.current_epoch=N/A)
Starting epoch 0
                                                                          
Traceback (most recent call last):
  File "/home/staff/a/akhedekar/miniconda3/envs/gnn/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/home/staff/a/akhedekar/miniconda3/envs/gnn/lib/python3.10/runpy.py", line 86, in _run_code
 ...
 ...
 ...
  File "/home/staff/a/akhedekar/miniconda3/envs/gnn/lib/python3.10/site-packages/lightning/pytorch/callbacks/model_checkpoint.py", line 383, in _save_topk_checkpoint
    raise MisconfigurationException(m)
lightning.fabric.utilities.exceptions.MisconfigurationException: `ModelCheckpoint(monitor='val_roc-auc')` could not find the monitored key in the returned metrics: ['train_loss', 'train_loss_step', 'val_loss', 'val_loss_epoch', 'val_macro-f1', 'val_micro-f1', 'train_loss_epoch', 'train_macro-f1', 'train_micro-f1', 'epoch', 'step']. HINT: Did you call `log('val_roc-auc', value)` in the `LightningModule`?

@schnamo

Oh dear, it seems like we missed some important things in PR #130.

Also, the pass_loss_kwargs for electra was set to True instead of accepting value from the user or method parameter. Is there any particular reason for it with new changes introduced ?
https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-9b4cba0f4cb875af21676f1998e61cd7b25a1e65d7b971a00ce88d17df880706R244

That does not make sense. True is already the default and I don't see a reason for overriding the user input. We should remove that line.

Also in electra config, the model type is set to regression, shouldn't it be the original classification version by default.

https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-286e6c43d041243386413f98ba9f93eca188d678d39b19a40450c57df5ab9979R3

You are right, it should be classification.

Also, I wonder whether the issue #53 was also addressed in that PR. As I see files and tests related to tox changed.

Probably, that is a question for @schnamo

Default beta value for BCE is set to 1000

Is this was the intended change ?

https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-00639a90075191fb4a6de89627b1181de9a0d44451fbf239c0a96e9df6750280R2-R3

I would say not intended since beta is supposed to be a value between 0 and 1, we should ask @schnamo about this (and also about the other BCE configs).

I wonder why the following line as deleted ? https://github.com/ChEB-AI/python-chebai/pull/130/files#diff-5a15c77150d45ed1d4be890d0dc4808bfc54d5df4a9a998ddedf9ab583d7c3c3L276

I don't know why it got removed, we should re-add it (although I'm not sure if that solves the problem you mention).

For the loss configs, see also #139

pass_loss_kwargs always set to True removed
electra config set to classification
loss configs tidied up
BCE loss beta parameter set to 0.99 in accordance with Chebifier
Re-added accidentally removed line

@schnamo

Thanks. @schnamo Can you confirm that the toxicity-related issues are out of date (#53) and your version of the toxicity datasets is functional?

Seems to work fine now!

Navigation Menu

Repository navigation

Missing labels in loss kwargs causing issue for electra model / Issues from PR#130 #140

Additional Information

Default beta value for BCE is set to 1000

Default beta value for BCE is set to 1000

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Missing labels in loss kwargs causing issue for electra model / Issues from PR#130 #140

Description

Additional Information

Activity

aditya0by0 commented on Dec 18, 2025

aditya0by0 commented on Dec 18, 2025

aditya0by0 commented on Dec 18, 2025

aditya0by0 commented on Jan 9, 2026

Default beta value for BCE is set to 1000

aditya0by0 commented on Jan 11, 2026

sfluegel05 commented on Jan 19, 2026

Default beta value for BCE is set to 1000

sfluegel05 commented on Jan 19, 2026

schnamo commented on Jan 20, 2026

sfluegel05 commented on Jan 20, 2026

schnamo commented on Jan 26, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions