Hi! I was running into an error using the compute_metrics function for fine-tuning a wav2vec 2.0 model in the way that this blog post recommends. When trying to fine-tune (and evaluate) with a language model (following this blog post ), I originally ran into an error with using processor.batch_deco…

[image] fauxneticien: make Hey @fauxneticien , Good question! I think you should do the following: def compute_metrics(pred): pred_logits = pred.predictions pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id if type(processor).__name__ == "Wav2Vec2Process…

Code review: compute_metrics for WER with Wav2Vec2ProcessorWithLM

🤗Transformers

patrickvonplaten April 19, 2022, 12:11pm 4

You can always access the tokenizer from the processor

Topic		Replies	Views
Compute_metrics do not find tokenizer (whisper finetuning) 🤗Transformers	1	327	March 6, 2024
How is compute_metrics working internally? Beginners	2	827	August 30, 2021
Input of compute_metrics in ASR model Beginners	2	1330	April 19, 2021
Different wer using Trainer.evaluate() and a loop with metric.add_batch Beginners	0	570	February 8, 2022
Eliminating PAD token from wav2vec2 prediction 🤗Transformers	2	947	June 3, 2022

Code review: compute_metrics for WER with Wav2Vec2ProcessorWithLM

Related topics