dinov2-large-linearhead-2024_03_06-with_data_aug_batch-size32_epochs93_freeze

DinoVd'eau is a fine-tuned version of facebook/dinov2-large on the multilabel_complete_dataset dataset. It achieves the following results on the evaluation set:

• Loss: 0.0947
• F1 Micro: 0.8528
• F1 Macro: 0.8273
• Roc Auc: 0.9054
• Accuracy: 0.5468
• Learning Rate: 0.0000

Model description

DinoVd'eau is a model built on top of dinov2 model for underwater multilabel image classification.The classification head is a combination of linear, ReLU, batch normalization, and dropout layers.

• Developed by: lombardata, credits to César Leblanc

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

	train	val	test	Total
Acropore_branched	804	202	200	1206
Acropore_digitised	465	108	101	674
Acropore_tabular	964	276	267	1507
Algae_assembly	2172	692	698	3562
Algae_limestone	1327	439	441	2207
Algae_sodding	2079	676	671	3426
Dead_coral	1126	358	355	1839
Fish	874	243	242	1359
Human_object	407	135	136	678
Living_coral	1765	580	571	2916
Millepore	350	119	102	571
No_acropore_encrusting	411	142	129	682
No_acropore_foliaceous	212	34	39	285
No_acropore_massive	921	317	310	1548
No_acropore_sub_massive	1205	362	363	1930
Rock	3736	1218	1217	6171
Sand	3594	1202	1194	5990
Scrap	2121	724	741	3586
Sea_cucumber	781	254	265	1300
Sea_urchins	189	60	72	321
Sponge	226	75	88	389
Syringodium_isoetifolium	1171	386	392	1949
Thalassodendron_ciliatum	783	261	260	1304
Useless	587	195	195	977

Training procedure

Data Augmentation

Data were augmented using the following transformations :

• training transformations : Sequential( (0): PreProcess() (1): Resize(output_size=(518, 518), p=1.0, p_batch=1.0, same_on_batch=True, size=(518, 518), side=short, resample=bilinear, align_corners=True, antialias=False) (2): RandomHorizontalFlip(p=0.25, p_batch=1.0, same_on_batch=False) (3): RandomVerticalFlip(p=0.25, p_batch=1.0, same_on_batch=False) (4): ColorJiggle(brightness=0.0, contrast=0.0, saturation=0.0, hue=0.0, p=0.25, p_batch=1.0, same_on_batch=False) (5): RandomPerspective(distortion_scale=0.5, p=0.25, p_batch=1.0, same_on_batch=False, align_corners=False, resample=bilinear) (6): Normalize(p=1.0, p_batch=1.0, same_on_batch=True, mean=tensor([0.4850, 0.4560, 0.4060]), std=tensor([0.2290, 0.2240, 0.2250])) )
• validation transformations : Sequential( (0): PreProcess() (1): Resize(output_size=(518, 518), p=1.0, p_batch=1.0, same_on_batch=True, size=(518, 518), side=short, resample=bilinear, align_corners=True, antialias=False) (2): Normalize(p=1.0, p_batch=1.0, same_on_batch=True, mean=tensor([0.4850, 0.4560, 0.4060]), std=tensor([0.2290, 0.2240, 0.2250])) )

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: ReduceLROnPlateau with a patience of 5 epochs and a factor of 0.1
• freeze_encoder: True
• num_epochs: 93

Training results

Training Loss	Epoch	Step	Validation Loss	F1 Micro	F1 Macro	Roc Auc	Accuracy	Rate
No log	1.0	274	0.1280	0.7998	0.7183	0.8765	0.4798	0.001
0.1493	2.0	548	0.1210	0.8095	0.7528	0.8793	0.5080	0.001
0.1493	3.0	822	0.1156	0.8119	0.7609	0.8667	0.5237	0.001
0.1156	4.0	1096	0.1167	0.8240	0.7881	0.9000	0.5003	0.001
0.1156	5.0	1370	0.1073	0.8367	0.7939	0.8968	0.5254	0.001
0.1069	6.0	1644	0.1082	0.8319	0.7918	0.8911	0.5341	0.001
0.1069	7.0	1918	0.1065	0.8367	0.7811	0.8963	0.5390	0.001
0.1026	8.0	2192	0.1071	0.8364	0.8001	0.8959	0.5351	0.001
0.1026	9.0	2466	0.1101	0.8346	0.7925	0.9113	0.5111	0.001
0.1	10.0	2740	0.1074	0.8345	0.7808	0.8973	0.5320	0.001
0.0964	11.0	3014	0.1079	0.8375	0.7967	0.8985	0.5292	0.001
0.0964	12.0	3288	0.1070	0.8353	0.7908	0.8951	0.5341	0.001
0.0949	13.0	3562	0.1060	0.8371	0.7926	0.8987	0.5296	0.001
0.0949	14.0	3836	0.1035	0.8443	0.7987	0.9007	0.5438	0.001
0.0926	15.0	4110	0.1099	0.8363	0.8004	0.9060	0.5118	0.001
0.0926	16.0	4384	0.1086	0.8327	0.7991	0.8886	0.5355	0.001
0.0911	17.0	4658	0.1084	0.8333	0.7967	0.8952	0.5209	0.001
0.0911	18.0	4932	0.1083	0.8358	0.7976	0.8968	0.5344	0.001
0.0902	19.0	5206	0.1129	0.8301	0.7799	0.8829	0.5233	0.001
0.0902	20.0	5480	0.1033	0.8464	0.8107	0.9065	0.5400	0.001
0.0896	21.0	5754	0.1091	0.8375	0.8014	0.9018	0.5233	0.001
0.0881	22.0	6028	0.1040	0.8412	0.7987	0.8995	0.5383	0.001
0.0881	23.0	6302	0.1090	0.8385	0.7908	0.9012	0.5278	0.001
0.0874	24.0	6576	0.1078	0.8338	0.7961	0.8917	0.5313	0.001
0.0874	25.0	6850	0.1054	0.8455	0.8077	0.9023	0.5501	0.001
0.0864	26.0	7124	0.1085	0.8346	0.7913	0.8860	0.5348	0.001
0.0864	27.0	7398	0.0994	0.8486	0.8134	0.9040	0.5487	0.0001
0.0793	28.0	7672	0.0989	0.8495	0.8123	0.9039	0.5532	0.0001
0.0793	29.0	7946	0.0986	0.8485	0.8107	0.9028	0.5511	0.0001
0.0751	30.0	8220	0.0986	0.8510	0.8188	0.9080	0.5501	0.0001
0.0751	31.0	8494	0.0990	0.8488	0.8139	0.9034	0.5539	0.0001
0.0753	32.0	8768	0.0983	0.8510	0.8181	0.9048	0.5505	0.0001
0.0748	33.0	9042	0.0987	0.8494	0.8110	0.9018	0.5539	0.0001
0.0748	34.0	9316	0.0980	0.8501	0.8117	0.9045	0.5515	0.0001
0.0748	35.0	9590	0.0981	0.8502	0.8133	0.9064	0.5505	0.0001
0.0748	36.0	9864	0.0984	0.8507	0.8130	0.9045	0.5536	0.0001
0.0745	37.0	10138	0.0983	0.8507	0.8145	0.9067	0.5484	0.0001
0.0745	38.0	10412	0.0986	0.8486	0.8107	0.9011	0.5546	0.0001
0.0749	39.0	10686	0.0986	0.8491	0.8140	0.9029	0.5508	0.0001
0.0749	40.0	10960	0.0982	0.8487	0.8114	0.9002	0.5553	0.0001
0.0752	41.0	11234	0.0976	0.8505	0.8131	0.9058	0.5508	1e-05
0.0734	42.0	11508	0.0977	0.8500	0.8128	0.9046	0.5515	1e-05
0.0734	43.0	11782	0.0975	0.8498	0.8118	0.9053	0.5515	1e-05
0.0736	44.0	12056	0.0976	0.8495	0.8118	0.9046	0.5522	1e-05
0.0736	45.0	12330	0.0975	0.8503	0.8119	0.9053	0.5508	1e-05
0.0731	46.0	12604	0.0976	0.8498	0.8119	0.9046	0.5511	1e-05
0.0731	47.0	12878	0.0975	0.8500	0.8115	0.9046	0.5518	1e-05
0.0736	48.0	13152	0.0975	0.8505	0.8141	0.9052	0.5511	1e-05
0.0736	49.0	13426	0.0975	0.8504	0.8144	0.9053	0.5518	1e-05
0.073	50.0	13700	0.0975	0.8502	0.8138	0.9052	0.5518	0.0000
0.073	51.0	13974	0.0975	0.8499	0.8123	0.9049	0.5515	0.0000
0.0732	52.0	14248	0.0975	0.8500	0.8119	0.9049	0.5515	0.0000
0.0732	53.0	14522	0.0975	0.8499	0.8118	0.9047	0.5522	0.0000

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.0+cu118
• Datasets 2.14.5
• Tokenizers 0.15.0