Model description

DinoVdrone is a model built on top of DinoVdrone-large-2025_02_03_31850-bs32_freeze_probs model for underwater multilabel image classification.The classification head is a combination of linear, ReLU, batch normalization, and dropout layers.

The source code for training the model can be found in this Git repository.

Developed by: lombardata, credits to César Leblanc and Victor Illien

Intended uses & limitations

You can use the raw model for classify diverse marine species, encompassing coral morphotypes classes taken from the Global Coral Reef Monitoring Network (GCRMN), habitats classes and seagrass species.

Training and evaluation data

Details on the estimated number of images for each class are given in the following table:

Class	train	test	val	Total
Acropore_branched	575	108	99	782
Acropore_digitised	558	122	107	787
Acropore_tabular	325	119	108	552
Algae	2354	778	776	3908
Atra/Leucospilota	417	79	58	554
Dead_coral	1778	485	503	2766
Fish	1391	361	352	2104
Millepore	591	196	178	965
No_acropore_encrusting	460	212	211	883
No_acropore_massive	1778	604	592	2974
No_acropore_sub_massive	1563	439	443	2445
Rock	2381	793	781	3955
Rubble	2363	784	784	3931
Sand	2401	802	801	4004
Sea_cucumber	1116	313	287	1716
Sea_urchins	158	64	89	311

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

Number of Epochs: 37.0
Learning Rate: 0.001
Train Batch Size: 32
Eval Batch Size: 32
Optimizer: Adam
LR Scheduler Type: ReduceLROnPlateau with a patience of 5 epochs and a factor of 0.1
Freeze Encoder: Yes
Data Augmentation: Yes

Data Augmentation

Data were augmented using the following transformations :

Train Transforms

PreProcess: No additional parameters
Resize: probability=1.00
RandomHorizontalFlip: probability=0.25
RandomVerticalFlip: probability=0.25
ColorJiggle: probability=0.25
RandomPerspective: probability=0.25
Normalize: probability=1.00

Val Transforms

PreProcess: No additional parameters
Resize: probability=1.00
Normalize: probability=1.00

Training results

Epoch	Validation Loss	MAE	RMSE	KL div	Learning Rate
1	0.5543646216392517	0.2278	0.2605	1.7912	0.001
2	0.48171326518058777	0.1602	0.2007	1.0345	0.001
3	0.4615386724472046	0.1370	0.1801	0.6064	0.001
4	0.4632064700126648	0.1391	0.1837	0.6577	0.001
5	0.4579373300075531	0.1363	0.1769	0.6678	0.001
6	0.4571229815483093	0.1330	0.1766	0.7896	0.001
7	0.4586440622806549	0.1307	0.1773	0.6493	0.001
8	0.45786425471305847	0.1319	0.1772	0.9475	0.001
9	0.4551210105419159	0.1306	0.1746	0.7271	0.001
10	0.45817646384239197	0.1316	0.1774	0.6882	0.001
11	0.46834859251976013	0.1372	0.1842	0.3715	0.001
12	0.4578668475151062	0.1316	0.1764	0.5271	0.001
13	0.4558842182159424	0.1301	0.1756	0.9168	0.001
14	0.4555540680885315	0.1292	0.1749	0.8827	0.001
15	0.45217740535736084	0.1262	0.1717	0.7009	0.001
16	0.45556434988975525	0.1286	0.1753	1.0038	0.001
17	0.458648681640625	0.1343	0.1775	0.2600	0.001
18	0.567169725894928	0.1638	0.2369	2.0548	0.001
19	0.45287612080574036	0.1287	0.1727	0.7115	0.001
20	0.45518893003463745	0.1285	0.1746	0.9694	0.001
21	0.45299893617630005	0.1282	0.1724	0.7789	0.001
22	0.4502638280391693	0.1261	0.1700	0.7369	0.0001
23	0.453466534614563	0.1280	0.1716	0.5027	0.0001
24	0.4502425491809845	0.1264	0.1697	0.5968	0.0001
25	0.45040303468704224	0.1267	0.1699	0.6215	0.0001
26	0.4509589374065399	0.1260	0.1704	0.6568	0.0001
27	0.4497845768928528	0.1262	0.1693	0.5748	0.0001
28	0.45060041546821594	0.1256	0.1701	0.7001	0.0001
29	0.4504892826080322	0.1263	0.1699	0.5840	0.0001
30	0.45060065388679504	0.1252	0.1703	0.8101	0.0001
31	0.45080825686454773	0.1249	0.1701	0.7416	0.0001
32	0.4501984417438507	0.1254	0.1697	0.6402	0.0001
33	0.4510658085346222	0.1250	0.1710	0.8411	0.0001
34	0.45148056745529175	0.1259	0.1711	0.7204	1e-05
35	0.4502483904361725	0.1247	0.1698	0.7355	1e-05
36	0.4508889615535736	0.1261	0.1703	0.4990	1e-05
37	0.44998663663864136	0.1260	0.1696	0.5451	1e-05

Framework Versions

Transformers: 4.48.0
Pytorch: 2.5.1+cu124
Datasets: 3.0.2
Tokenizers: 0.21.0

groderg
/

DinoVdrone-large-2025_02_03_31850-bs32_freeze_probs