Model description

drone-DinoVdeau-from-binary is a model built on top of drone-DinoVdeau-from-binary-large-2024_11_14-batch-size16_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	1272	394	391	2057
Acropore_digitised	624	223	217	1064
Acropore_tabular	344	144	125	613
Algae	5236	1580	1617	8433
Dead_coral	2251	650	655	3556
Millepore	233	96	97	426
No_acropore_encrusting	802	266	285	1353
No_acropore_massive	2381	826	822	4029
No_acropore_sub_massive	2020	625	651	3296
Rock	6151	2004	2004	10159
Rubble	5170	1648	1627	8445
Sand	6121	2019	1978	10118

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

Number of Epochs: 62.0
Learning Rate: 0.001
Train Batch Size: 16
Eval Batch Size: 16
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	0.43063807487487793	0.1621	0.2210	1.0069	0.001
2	0.4245865046977997	0.1547	0.2179	1.3119	0.001
3	0.422325998544693	0.1554	0.2158	1.0982	0.001
4	0.41912660002708435	0.1552	0.2142	1.0414	0.001
5	0.41713497042655945	0.1541	0.2123	1.0698	0.001
6	0.42093637585639954	0.1520	0.2140	1.1959	0.001
7	0.4166290760040283	0.1530	0.2126	1.1709	0.001
8	0.41946443915367126	0.1556	0.2143	0.9712	0.001
9	0.41668570041656494	0.1524	0.2121	1.1432	0.001
10	0.4186115860939026	0.1535	0.2139	0.9121	0.001
11	0.41557687520980835	0.1536	0.2114	0.9950	0.001
12	0.41883811354637146	0.1555	0.2139	1.0106	0.001
13	0.41630858182907104	0.1553	0.2121	1.1482	0.001
14	0.4193180799484253	0.1546	0.2138	1.2111	0.001
15	0.416218638420105	0.1542	0.2121	1.2043	0.001
16	0.41389620304107666	0.1528	0.2102	1.0828	0.001
17	0.4171081781387329	0.1564	0.2118	1.0006	0.001
18	0.4146382212638855	0.1507	0.2107	1.0514	0.001
19	0.41857486963272095	0.1532	0.2114	0.9575	0.001
20	0.41434723138809204	0.1513	0.2108	1.1648	0.001
21	0.4195358157157898	0.1533	0.2123	1.2950	0.001
22	0.4339658319950104	0.1524	0.2110	inf	0.001
23	0.43265336751937866	0.1517	0.2085	nan	0.0001
24	0.4384593963623047	0.1493	0.2092	nan	0.0001
25	0.4271779954433441	0.1490	0.2074	inf	0.0001
26	0.41048941016197205	0.1480	0.2075	1.1903	0.0001
27	0.4096038341522217	0.1494	0.2067	0.9915	0.0001
28	0.4104350507259369	0.1493	0.2075	0.9669	0.0001
29	0.40966179966926575	0.1469	0.2069	1.0433	0.0001
30	0.4094092547893524	0.1490	0.2065	0.9082	0.0001
31	0.40909385681152344	0.1470	0.2065	1.0120	0.0001
32	0.4084269404411316	0.1483	0.2060	0.9708	0.0001
33	0.40824124217033386	0.1474	0.2057	0.9317	0.0001
34	0.40851354598999023	0.1481	0.2061	0.9619	0.0001
35	0.4072923958301544	0.1466	0.2054	1.0523	0.0001
36	0.40741708874702454	0.1460	0.2052	1.0622	0.0001
37	0.40657544136047363	0.1456	0.2047	1.0201	0.0001
38	0.406360387802124	0.1459	0.2045	1.0557	0.0001
39	0.4077896773815155	0.1469	0.2056	1.0055	0.0001
40	0.4068063199520111	0.1464	0.2049	0.9849	0.0001
41	0.40890073776245117	0.1489	0.2063	0.8999	0.0001
42	0.4068816602230072	0.1463	0.2049	1.0617	0.0001
43	0.40578988194465637	0.1450	0.2041	1.0520	0.0001
44	0.4070681035518646	0.1475	0.2050	1.0054	0.0001
45	0.40669572353363037	0.1440	0.2047	1.1386	0.0001
46	0.40670666098594666	0.1457	0.2047	1.0253	0.0001
47	0.4062415659427643	0.1473	0.2043	1.0430	0.0001
48	0.4064981937408447	0.1457	0.2048	1.1041	0.0001
49	0.40709760785102844	0.1463	0.2052	1.0702	0.0001
50	0.40644556283950806	0.1479	0.2042	0.8917	1e-05
51	0.40579161047935486	0.1437	0.2041	0.9960	1e-05
52	0.40528106689453125	0.1446	0.2037	1.0567	1e-05
53	0.4056229293346405	0.1462	0.2039	1.0205	1e-05
54	0.4058997631072998	0.1441	0.2041	0.9905	1e-05
55	0.4060685932636261	0.1471	0.2041	0.9379	1e-05
56	0.40592971444129944	0.1454	0.2041	0.9696	1e-05
57	0.4058408737182617	0.1460	0.2041	1.0591	1e-05
58	0.4063320457935333	0.1460	0.2043	0.9276	1e-05
59	0.4056239724159241	0.1453	0.2038	0.9794	1.0000000000000002e-06
60	0.40571752190589905	0.1446	0.2040	1.0349	1.0000000000000002e-06
61	0.4058452248573303	0.1449	0.2041	0.9860	1.0000000000000002e-06
62	0.4054276943206787	0.1446	0.2037	0.9528	1.0000000000000002e-06

Framework Versions

Transformers: 4.41.0
Pytorch: 2.5.0+cu124
Datasets: 3.0.2
Tokenizers: 0.19.1

lombardata
/

drone-DinoVdeau-from-binary-large-2024_11_14-batch-size16_freeze_probs