【开源实习】audio_spectrogram_transformer模型微调

[Bireflection]

audio_spectrogram_transformer 模型微调报告

任务

• 任务编号：#IAUOSX
• 任务链接：【开源实习】audio_spectrogram_transformer模型微调
• 实现了audio_spectrogram_transformer 在esc50数据集上的微调
• 模型：MIT/ast-finetuned-audioset-10-10-0.4593
• 数据集: ashraq/esc50

结果对比：

Mindnlp+D910B

Epoch	Training Loss	Eval Loss	Accuracy
1	3.0928	2.2305	0.8150
2	1.4845	0.9815	0.8950
3	0.5733	0.4876	0.9250
4	0.2061	0.3770	0.9125
5	0.0742	0.3207	0.9300
6	0.0443	0.1821	0.9600
7	0.0178	0.2144	0.9575
8	0.0111	0.2155	0.9575
9	0.0094	0.2167	0.9575
10	0.0087	0.2174	0.9575

---

Pytorch+3090

Epoch	Training Loss	Validation Loss	Accuracy
1	1.2550	0.3934	0.8825
2	0.4477	0.3656	0.8925
3	0.3289	0.2777	0.9200
4	0.2200	0.3645	0.9175
5	0.1679	0.2345	0.9350
6	0.1140	0.1877	0.9575
7	0.0925	0.1641	0.9575
8	0.0648	0.1810	0.9475
9	0.0593	0.1285	0.9550
10	0.0269	0.1222	0.9575