Model description

drone-DinoVdeau-from-probs is a model built on top of drone-DinoVdeau-from-probs-large-2024_11_15-batch-size64_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	1220	363	362	1945
Acropore_digitised	586	195	189	970
Acropore_tabular	308	133	119	560
Algae	4777	1372	1384	7533
Dead_coral	2513	671	693	3877
Millepore	136	55	59	250
No_acropore_encrusting	252	88	93	433
No_acropore_massive	2158	725	726	3609
No_acropore_sub_massive	2036	582	612	3230
Rock	5976	1941	1928	9845
Rubble	4851	1486	1474	7811
Sand	6155	2019	1990	10164

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

Number of Epochs: 79.0
Learning Rate: 0.001
Train Batch Size: 64
Eval Batch Size: 64
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	0.5005590319633484	0.1552	0.1904	0.1025	0.001
2	0.47547808289527893	0.1245	0.1681	0.5180	0.001
3	0.47452571988105774	0.1227	0.1675	0.6862	0.001
4	0.47420722246170044	0.1255	0.1672	0.3212	0.001
5	0.47245556116104126	0.1224	0.1653	0.5072	0.001
6	0.4725925624370575	0.1216	0.1657	0.6710	0.001
7	0.4731809198856354	0.1255	0.1655	0.3162	0.001
8	0.47284314036369324	0.1260	0.1651	0.2719	0.001
9	0.4707973003387451	0.1206	0.1639	0.6393	0.001
10	0.4732784628868103	0.1230	0.1654	0.5359	0.001
11	0.47162503004074097	0.1253	0.1647	0.2479	0.001
12	0.47083696722984314	0.1244	0.1631	0.3119	0.001
13	0.47152063250541687	0.1230	0.1635	0.3694	0.001
14	0.47212228178977966	0.1216	0.1653	0.5592	0.001
15	0.47012239694595337	0.1213	0.1628	0.4936	0.001
16	0.4718552827835083	0.1229	0.1646	0.2820	0.001
17	0.46933484077453613	0.1200	0.1621	0.5294	0.001
18	0.4710436165332794	0.1216	0.1635	0.4093	0.001
19	0.4698491394519806	0.1219	0.1622	0.2918	0.001
20	0.4691685736179352	0.1190	0.1617	0.4772	0.001
21	0.46830564737319946	0.1204	0.1606	0.4336	0.001
22	0.47239789366722107	0.1183	0.1650	0.7962	0.001
23	0.47136834263801575	0.1223	0.1641	0.2854	0.001
24	0.4706868529319763	0.1207	0.1633	0.4206	0.001
25	0.46786901354789734	0.1185	0.1606	0.5436	0.001
26	0.47084224224090576	0.1192	0.1634	0.4964	0.001
27	0.4695045053958893	0.1185	0.1625	0.6399	0.001
28	0.4700873792171478	0.1184	0.1624	0.5737	0.001
29	0.4698559045791626	0.1200	0.1624	0.4459	0.001
30	0.4722815454006195	0.1254	0.1643	0.2726	0.001
31	0.46958214044570923	0.1184	0.1622	0.5308	0.001
32	0.46677276492118835	0.1175	0.1593	0.4200	0.0001
33	0.46626824140548706	0.1177	0.1587	0.3529	0.0001
34	0.46665358543395996	0.1181	0.1592	0.3588	0.0001
35	0.46587392687797546	0.1160	0.1584	0.4813	0.0001
36	0.46578526496887207	0.1173	0.1581	0.3504	0.0001
37	0.4654408395290375	0.1158	0.1578	0.3919	0.0001
38	0.46546319127082825	0.1166	0.1580	0.4058	0.0001
39	0.465843141078949	0.1174	0.1585	0.4118	0.0001
40	0.46561121940612793	0.1170	0.1579	0.3564	0.0001
41	0.4657152593135834	0.1171	0.1582	0.3573	0.0001
42	0.4651602804660797	0.1155	0.1579	0.5042	0.0001
43	0.4651065468788147	0.1157	0.1575	0.4462	0.0001
44	0.46537330746650696	0.1166	0.1579	0.4236	0.0001
45	0.46489208936691284	0.1151	0.1574	0.4510	0.0001
46	0.46484702825546265	0.1157	0.1575	0.4490	0.0001
47	0.4648602306842804	0.1152	0.1574	0.4751	0.0001
48	0.4647873342037201	0.1151	0.1575	0.5305	0.0001
49	0.4647849500179291	0.1154	0.1574	0.4799	0.0001
50	N/A	0.0000	0.0000	0.0000	0.0001
51	0.465638667345047	0.1151	0.1582	0.4879	0.0001
52	0.46429532766342163	0.1155	0.1566	0.4199	0.0001
53	0.46441230177879333	0.1156	0.1569	0.3880	0.0001
54	0.4646008610725403	0.1148	0.1569	0.4229	0.0001
55	0.4644174873828888	0.1159	0.1569	0.4009	0.0001
56	0.464743047952652	0.1164	0.1572	0.3405	0.0001
57	0.4645179808139801	0.1152	0.1569	0.4188	0.0001
58	0.465102881193161	0.1164	0.1576	0.3079	0.0001
59	0.4644688367843628	0.1150	0.1570	0.4339	1e-05
60	0.46417686343193054	0.1150	0.1566	0.3894	1e-05
61	0.4639436900615692	0.1146	0.1563	0.4145	1e-05
62	0.4641311764717102	0.1148	0.1565	0.4064	1e-05
63	0.4643491506576538	0.1149	0.1565	0.3542	1e-05
64	0.46402981877326965	0.1150	0.1564	0.3718	1e-05
65	0.4640822410583496	0.1152	0.1565	0.4128	1e-05
66	0.46441909670829773	0.1145	0.1570	0.4988	1e-05
67	0.46383005380630493	0.1151	0.1562	0.4122	1e-05
68	0.4639807641506195	0.1144	0.1565	0.4579	1e-05
69	0.4637599587440491	0.1143	0.1561	0.4197	1e-05
70	0.46392253041267395	0.1145	0.1563	0.4286	1e-05
71	0.46406444907188416	0.1153	0.1563	0.3542	1e-05
72	0.46417826414108276	0.1147	0.1566	0.4250	1e-05
73	0.4637835919857025	0.1140	0.1561	0.4397	1e-05
74	0.463798850774765	0.1145	0.1563	0.4437	1e-05
75	0.46379053592681885	0.1145	0.1561	0.4049	1e-05
76	0.4639701247215271	0.1141	0.1565	0.4926	1.0000000000000002e-06
77	0.463869571685791	0.1142	0.1562	0.4427	1.0000000000000002e-06
78	0.46388140320777893	0.1145	0.1563	0.4293	1.0000000000000002e-06
79	0.46412238478660583	0.1147	0.1564	0.3765	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-probs-large-2024_11_15-batch-size64_freeze_probs