asr-christiian

Working Repo von Christian

Week 1: 08.10 - 15.10

• Proof of concept: get finetuning script from Peter working => improves performance on mini train set
• prepare code for hyperparameter optimization => not seemless integration with ray tune (industry standard for hyperparameter optimization)
• prepare presentation on hyperparameter optimization

Week 2&3: 16.10 - 31.11

• Set up High-Perfomance cluster
• Submit evaluation job for baseline whisper model
• prepare code for hyperparameter optimization (from Week 1) => Collator Function had to be adapted as Ray Tune datasets created a numpy object which could not be processed from HF tokenizer.pad class

Next steps:

• Evaluate Baseline model (WhisperX e.g.) on test set to directly compare performance with fine-tuned model
• Define hyperparameters to optimize
• Run first dummy-finetuning on cluster