i7p9h9/myna-lab-task

fillme

Hardware requirements

• GPU with 11Gb memory is necessary
• 64Gb RAM memory at least
• 120Gb free space is required (ssd type partition is recommended)

---

Start with docker

• git clone git@github.com:i7p9h9/myna-lab-task.git
• cd mina-lab-task/
• prepare docker environment:

./local/prepare_for_docker.sh

---

To run training process

• edit path.sh
  • DATASETS_DIR - path to folder with trainig data, next structure expected:

  .
  ├── test-example
  ├── test-example.csv
  ├── train.csv
  └── train

  • PROCESSED_DIR - empty folder where processed dataset and augmentation will be saved, ssd partition type highly recommended
  • RESULT_DIR - folder where weights for neural network will be saved
• start training script:

./train.sh -j X

where 'X' is num cpu threads, 6-12 cores recomended

• wait... :)
• result will be saved in RESULT_DIR/final.torch and RESULT_DIR/final-half.torch

---

To inference process

• run:

./eval.sh -s csv_result_file -m path_to_model_file -d path_to_folder_with_wav

for instance:

./eval.sh -s result.csv -m exps/exp1/final-half.torch -d /media/ssd/myna-labs/numbers2/test-example/

---

Validation results

1. For supervised training process validation on 500 labeled files showed CER: 0.0040
2. For fixmatch (semi-supervised) training process validation on 500 labeled files showed CER: 0.0017