ast-finetuned-audioset-10-10-0.4593_ft_ESC-50_aug_0-1

This model is a fine-tuned version of MIT/ast-finetuned-audioset-10-10-0.4593 on a subset of ashraq/esc50 dataset. It achieves the following results on the evaluation set:

• Loss: 0.7391
• Accuracy: 0.9286
• Precision: 0.9449
• Recall: 0.9286
• F1: 0.9244

Training and evaluation data

Training and evaluation data were augmented with audiomentations GitHub: iver56/audiomentations library and the following augmentation methods have been performed based on previous experiments Elliott et al.: Tiny transformers for audio classification at the edge:

Gain

• each audio sample is amplified/attenuated by a random factor between 0.5 and 1.5 with a 0.3 probability

Noise

• a random amount of Gaussian noise with a relative amplitude between 0.001 and 0.015 is added to each audio sample with a 0.5 probability

Speed adjust

• duration of each audio sample is extended by a random amount between 0.5 and 1.5 with a 0.3 probability

Pitch shift

• pitch of each audio sample is shifted by a random amount of semitones selected from the closed interval [-4,4] with a 0.3 probability

Time masking

• a random fraction of lenght of each audio sample in the range of (0,0.02] is erased with a 0.3 probability

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10

Training results

Test results

Framework versions

• Transformers 4.27.4
• Pytorch 2.0.0
• Datasets 2.10.1
• Tokenizers 0.13.2