|
--- |
|
license: apache-2.0 |
|
base_model: facebook/wav2vec2-base |
|
tags: |
|
- audio-classification |
|
- generated_from_trainer |
|
metrics: |
|
- accuracy |
|
model-index: |
|
- name: wav2vec2-base_down_on |
|
results: [] |
|
--- |
|
|
|
<!-- This model card has been generated automatically according to the information the Trainer had access to. You |
|
should probably proofread and complete it, then remove this comment. --> |
|
|
|
# wav2vec2-base_down_on |
|
|
|
This model is a fine-tuned version of [facebook/wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) on the MatsRooth/down_on dataset. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 0.1385 |
|
- Accuracy: 0.9962 |
|
|
|
MatsRooth/down_on is the part of [superb ks](https://huggingface.co/datasets/superb) |
|
with the labels *down* and *on*. |
|
Superb ks is in turn derived from (Speech Commands dataset v1.0)[https://www.tensorflow.org/datasets/catalog/speech_commands]. |
|
Train/validation/test splits are as in superb ks. |
|
|
|
## Intended uses |
|
|
|
MatsRooth/down_on and this model exercise methodology for creating an audio classification dataset from |
|
local directory structures and audio files, and check whether fine tuning wav2vec2 classification with two labels works |
|
well. |
|
|
|
## Training procedure |
|
Training used 'sbatch' on a cluster and the program [run_audio_classification.py](https://github.com/huggingface/transformers). |
|
'down_on.sub' is below, start it with 'sbatch down_on.sub'. |
|
|
|
''' |
|
#!/bin/bash |
|
#SBATCH -J down_on # Job name |
|
#SBATCH -o down_on_%j.out # Name of stdout output log file (%j expands to jobID) |
|
#SBATCH -e down_on_%j.err # Name of stderr output log file (%j expands to jobID) |
|
#SBATCH -N 1 # Total number of nodes requested |
|
#SBATCH -n 1 # Total number of cores requested |
|
#SBATCH --mem=5000 # Total amount of (real) memory requested (per node) |
|
#SBATCH -t 10:00:00 # Time limit (hh:mm:ss) |
|
#SBATCH --partition=gpu # Request partition for resource allocation |
|
#SBATCH --gres=gpu:1 # Specify a list of generic consumable resources (per node) |
|
|
|
cd ~/ac_h |
|
/home/mr249/env/hugh/bin/python run_audio_classification.py \ |
|
--model_name_or_path facebook/wav2vec2-base \ |
|
--dataset_name MatsRooth/down_on \ |
|
--output_dir wav2vec2-base_down_on \ |
|
--overwrite_output_dir \ |
|
--remove_unused_columns False \ |
|
--do_train \ |
|
--do_eval \ |
|
--fp16 \ |
|
--learning_rate 3e-5 \ |
|
--max_length_seconds 1 \ |
|
--attention_mask False \ |
|
--warmup_ratio 0.1 \ |
|
--num_train_epochs 5 \ |
|
--per_device_train_batch_size 32 \ |
|
--gradient_accumulation_steps 4 \ |
|
--per_device_eval_batch_size 32 \ |
|
--dataloader_num_workers 1 \ |
|
--logging_strategy steps \ |
|
--logging_steps 10 \ |
|
--evaluation_strategy epoch \ |
|
--save_strategy epoch \ |
|
--load_best_model_at_end True \ |
|
--metric_for_best_model accuracy \ |
|
--save_total_limit 3 \ |
|
--seed 0 |
|
''' |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 3e-05 |
|
- train_batch_size: 32 |
|
- eval_batch_size: 32 |
|
- seed: 0 |
|
- gradient_accumulation_steps: 4 |
|
- total_train_batch_size: 128 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- lr_scheduler_warmup_ratio: 0.1 |
|
- num_epochs: 5.0 |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Accuracy | |
|
|:-------------:|:-----:|:----:|:---------------:|:--------:| |
|
| 0.6089 | 1.0 | 29 | 0.1385 | 0.9962 | |
|
| 0.1297 | 2.0 | 58 | 0.0513 | 0.9962 | |
|
| 0.0835 | 3.0 | 87 | 0.0389 | 0.9885 | |
|
| 0.058 | 4.0 | 116 | 0.0302 | 0.9923 | |
|
| 0.0481 | 5.0 | 145 | 0.0245 | 0.9942 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.31.0.dev0 |
|
- Pytorch 2.0.1+cu117 |
|
- Datasets 2.13.1 |
|
- Tokenizers 0.13.3 |
|
|
