AttributeError: 'SegformerFeatureExtractor' object has no attribute 'reduce_labels' still has no clear guide around

[deanAirre]

System Info

Python 3.11.10, transformers 4.47.0

Who can help?

@stevhliu

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Trying to train by using
`from transformers import AutoFeatureExtractor

feature_extractor = AutoFeatureExtractor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")`

as feature extractor and keep getting AttributeError: 'SegformerFeatureExtractor' object has no attribute 'reduce_labels' still has no clear guide around

found this that said to repair the docs but I still haven't found the solution to do it by reading links and docs surrounding the links. Is it still a feature or should I move to other feature extractor?

Expected behavior

``AttributeError: 'SegformerFeatureExtractor' object has no attribute 'reduce_labels' ` solution should be

feature_extractor = AutoFeatureExtractor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512", do_reduce_labels=True)
according to the link, but the problem persists.

Edit2:
Complete error message since by the time I wrote this I already try running it again for another chance. Here's the complete error code

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[158], line 1
----> 1 trainer.train()
      2 trainer.push_to_hub()

File c:\Users\Lenovo\miniconda3\envs\pretrain-huggingface\Lib\site-packages\transformers\trainer.py:2155, in Trainer.train(self, resume_from_checkpoint, trial, ignore_keys_for_eval, **kwargs)
   2152 try:
   2153     # Disable progress bars when uploading models during checkpoints to avoid polluting stdout
   2154     hf_hub_utils.disable_progress_bars()
-> 2155     return inner_training_loop(
   2156         args=args,
   2157         resume_from_checkpoint=resume_from_checkpoint,
   2158         trial=trial,
   2159         ignore_keys_for_eval=ignore_keys_for_eval,
   2160     )
   2161 finally:
   2162     hf_hub_utils.enable_progress_bars()

File c:\Users\Lenovo\miniconda3\envs\pretrain-huggingface\Lib\site-packages\transformers\trainer.py:2589, in Trainer._inner_training_loop(self, batch_size, args, resume_from_checkpoint, trial, ignore_keys_for_eval)
   2587     self.state.epoch = epoch + (step + 1 + steps_skipped) / steps_in_epoch
   2588     self.control = self.callback_handler.on_step_end(args, self.state, self.control)
-> 2589     self._maybe_log_save_evaluate(
   2590         tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval, start_time
   2591     )
   2592 else:
   2593     self.control = self.callback_handler.on_substep_end(args, self.state, self.control)

File c:\Users\Lenovo\miniconda3\envs\pretrain-huggingface\Lib\site-packages\transformers\trainer.py:3047, in Trainer._maybe_log_save_evaluate(self, tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval, start_time)
   3045 metrics = None
   3046 if self.control.should_evaluate:
-> 3047     metrics = self._evaluate(trial, ignore_keys_for_eval)
   3048     is_new_best_metric = self._determine_best_metric(metrics=metrics, trial=trial)
   3050     if self.args.save_strategy == SaveStrategy.BEST:

File c:\Users\Lenovo\miniconda3\envs\pretrain-huggingface\Lib\site-packages\transformers\trainer.py:3001, in Trainer._evaluate(self, trial, ignore_keys_for_eval, skip_scheduler)
   3000 def _evaluate(self, trial, ignore_keys_for_eval, skip_scheduler=False):
-> 3001     metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
   3002     self._report_to_hp_search(trial, self.state.global_step, metrics)
   3004     # Run delayed LR scheduler now that metrics are populated

File c:\Users\Lenovo\miniconda3\envs\pretrain-huggingface\Lib\site-packages\transformers\trainer.py:4051, in Trainer.evaluate(self, eval_dataset, ignore_keys, metric_key_prefix)
   4048 start_time = time.time()
   4050 eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
-> 4051 output = eval_loop(
   4052     eval_dataloader,
   4053     description="Evaluation",
   4054     # No point gathering the predictions if there are no metrics, otherwise we defer to
   4055     # self.args.prediction_loss_only
   4056     prediction_loss_only=True if self.compute_metrics is None else None,
   4057     ignore_keys=ignore_keys,
   4058     metric_key_prefix=metric_key_prefix,
   4059 )
   4061 total_batch_size = self.args.eval_batch_size * self.args.world_size
   4062 if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:

File c:\Users\Lenovo\miniconda3\envs\pretrain-huggingface\Lib\site-packages\transformers\trainer.py:4340, in Trainer.evaluation_loop(self, dataloader, description, prediction_loss_only, ignore_keys, metric_key_prefix)
   4338     eval_set_kwargs["losses"] = all_losses if "loss" in args.include_for_metrics else None
   4339     eval_set_kwargs["inputs"] = all_inputs if "inputs" in args.include_for_metrics else None
-> 4340     metrics = self.compute_metrics(
   4341         EvalPrediction(predictions=all_preds, label_ids=all_labels, **eval_set_kwargs)
   4342     )
   4343 elif metrics is None:
   4344     metrics = {}

Cell In[156], line 27, in compute_metrics(eval_pred)
     19 pred_labels = logits_tensor.detach().cpu().numpy()
     20 # currently using _compute instead of compute
     21 # see this issue for more info: https://github.com/huggingface/evaluate/pull/328#issuecomment-1286866576
     22 metrics = metric._compute(
     23         predictions=pred_labels,
     24         references=labels,
     25         num_labels=num_labels,
     26         ignore_index=0,
---> 27         reduce_labels=feature_extractor.reduce_labels,
     28     )
     30 # add per category metrics as individual key-value pairs
     31 per_category_accuracy = metrics.pop("per_category_accuracy").tolist()

AttributeError: 'SegformerFeatureExtractor' object has no attribute 'reduce_labels'

[NielsRogge]

Hi,

Feature extractors are deprecated, you can replace the code by:

from transformers import AutoImageProcessor

image_processor = AutoImageProcessor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512", do_reduce_labels=True)

it seems that reduce_labels=image_processor.reduce_labels, is passed to the metric._compute method, whereas you should be passing reduce_labels=image_processor.do_reduce_labels.

Where did you get this code from? I'll make sure it gets updated

[deanAirre]

Hi Niels, thanks for the reply,

Yes it works now like this.

from transformers import SegformerImageProcessor

feature_extractor = SegformerImageProcessor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512", reduce_labels=True)

I got the code with do_reduce_labels from this link, quote:

One additional thing to keep in mind is that one can initialize SegformerImageProcessor with do_reduce_labels set to True or False. In some datasets (like ADE20k), the 0 index is used in the annotated segmentation maps for background. However, ADE20k doesn’t include the “background” class in its 150 labels. Therefore, do_reduce_labels is used to reduce all labels by 1, and to make sure no loss is computed for the background class (i.e. it replaces 0 in the annotated maps by 255, which is the ignore_index of the loss function used by

Thanks in advance, amazing library btw,
Sean.