gradio_article.md

Dataset

The USK-Coffee dataset, made available at https://comvis.unsyiah.ac.id/usk-coffee/ is multi class image dataset derived from a coffee bean collection that includes 4 classes: peaberry, longberry, defect, and premium.

Training

Fast.ai was used to train this classifier with a Timm ConvNext nano vision learner, without heavy customization. The training was performed on the provided train split, and validation on the val split.

The final fine tuning of the training loop resulted in the following losses.

epoch	train_loss	valid_loss	accuracy	time
0	0.238523	0.383621	0.869375	00:25
1	0.257938	0.293417	0.907500	00:25
2	0.205048	0.412420	0.847500	00:25
3	0.170284	0.308219	0.901875	00:25
4	0.154471	0.308811	0.894375	00:26
5	0.107862	0.480474	0.874375	00:26
6	0.075452	0.506489	0.843125	00:26
7	0.060802	0.317052	0.906875	00:26
8	0.049216	0.242317	0.932500	00:26
9	0.040890	0.233353	0.935625	00:26

Examples

The example images provided in the demo are from the test split in the dataset, which was never made available to the model in the training process.