You can deploy this model and start earning money today!

lombardata/
bd_ortho-DinoVdeau-large-2024_11_27-batch-size64_freeze_probs

Name: bd_ortho-DinoVdeau-large-2024_11_27-batch-size64_freeze_probs
Availability: InStock

tensorboard

safetensors

dinov2

multilabel-image-classification

multilabel

generated_from_trainer

eng

cc0-1.0

Views

No views yet

bd_ortho-DinoVdeau is a fine-tuned version of bd_ortho-DinoVdeau-large-2024_11_27-batch-size64_freeze_probs. It achieves the following results on the test set:

Loss: 0.4551
RMSE: 0.0866
MAE: 0.0630
KL Divergence: 0.1147

Model description

bd_ortho-DinoVdeau is a model built on top of bd_ortho-DinoVdeau-large-2024_11_27-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	8074	2667	2658	13399
Acropore_digitised	3730	829	823	5382
Acropore_tabular	125	23	40	188
Algae	14027	4662	4660	23349
Dead_coral	11369	3364	3369	18102
Millepore	2	1	1	4
No_acropore_encrusting	0	0	0	0
No_acropore_massive	3265	423	463	4151
No_acropore_sub_massive	10241	2911	2924	16076
Rock	14090	4694	4693	23477
Rubble	12455	3915	3883	20253
Sand	12848	4098	4079	21025

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

Number of Epochs: 62.0
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	0.46336060762405396	0.0760	0.1018	0.0696	0.001
2	0.45933997631073	0.0716	0.0952	0.0038	0.001
3	0.457367479801178	0.0670	0.0918	0.0583	0.001
4	0.459468811750412	0.0713	0.0955	-0.0650	0.001
5	0.45927393436431885	0.0702	0.0954	-0.0835	0.001
6	0.46080395579338074	0.0728	0.0977	-0.0705	0.001
7	0.4581476151943207	0.0683	0.0927	-0.0044	0.001
8	0.4573117196559906	0.0680	0.0916	0.0799	0.001
9	0.45939013361930847	0.0706	0.0947	0.0233	0.001
10	0.45772281289100647	0.0675	0.0918	0.0885	0.001
11	0.45641985535621643	0.0662	0.0898	0.1296	0.001
12	0.45718902349472046	0.0677	0.0913	-0.0061	0.001
13	0.4622880220413208	0.0747	0.1002	-0.2060	0.001
14	0.45775285363197327	0.0678	0.0925	-0.0371	0.001
15	0.4575214684009552	0.0667	0.0917	0.0458	0.001
16	0.4578736424446106	0.0680	0.0926	0.0151	0.001
17	0.4592094421386719	0.0702	0.0949	-0.0679	0.001
18	0.45573291182518005	0.0651	0.0887	0.0421	0.0001
19	0.4555513262748718	0.0647	0.0885	0.0468	0.0001
20	0.45553284883499146	0.0648	0.0884	0.0405	0.0001
21	0.4555487334728241	0.0650	0.0884	0.0475	0.0001
22	0.45551028847694397	0.0646	0.0883	0.0570	0.0001
23	0.45505577325820923	0.0641	0.0874	0.0887	0.0001
24	0.4552234709262848	0.0642	0.0878	0.0555	0.0001
25	0.45521080493927	0.0645	0.0878	0.0238	0.0001
26	0.4557025730609894	0.0646	0.0885	0.0409	0.0001
27	0.4550967216491699	0.0639	0.0876	0.0548	0.0001
28	0.45512688159942627	0.0642	0.0876	0.0273	0.0001
29	0.45477041602134705	0.0634	0.0869	0.0744	0.0001
30	0.4549327790737152	0.0636	0.0873	0.0492	0.0001
31	0.4547973871231079	0.0632	0.0869	0.0688	0.0001
32	0.454988956451416	0.0639	0.0874	0.0271	0.0001
33	0.455375999212265	0.0647	0.0882	-0.0174	0.0001
34	0.45461305975914	0.0628	0.0866	0.1094	0.0001
35	0.45498156547546387	0.0639	0.0874	0.0571	0.0001
36	0.4548388123512268	0.0629	0.0869	0.1453	0.0001
37	0.45526784658432007	0.0645	0.0881	-0.0152	0.0001
38	0.45479556918144226	0.0636	0.0870	0.0490	0.0001
39	0.454780250787735	0.0631	0.0870	0.0726	0.0001
40	0.45476558804512024	0.0632	0.0870	0.0637	0.0001
41	0.45470812916755676	0.0634	0.0869	0.0390	1e-05
42	0.4543863534927368	0.0628	0.0862	0.1115	1e-05
43	0.4545557498931885	0.0632	0.0866	0.0533	1e-05
44	0.45448434352874756	0.0625	0.0864	0.1350	1e-05
45	0.4550137519836426	0.0642	0.0874	0.0044	1e-05
46	0.4545902609825134	0.0632	0.0867	0.0389	1e-05
47	0.4544997215270996	0.0630	0.0866	0.0370	1e-05
48	0.4546374976634979	0.0634	0.0868	0.0194	1e-05
49	0.45436596870422363	0.0627	0.0862	0.0667	1.0000000000000002e-06
50	0.45450592041015625	0.0631	0.0865	0.0548	1.0000000000000002e-06
51	0.4544804096221924	0.0629	0.0865	0.0428	1.0000000000000002e-06
52	0.45421910285949707	0.0623	0.0859	0.1236	1.0000000000000002e-06
53	0.4542272686958313	0.0625	0.0859	0.0887	1.0000000000000002e-06
54	0.4543103575706482	0.0624	0.0862	0.0917	1.0000000000000002e-06
55	0.45456644892692566	0.0631	0.0865	0.0774	1.0000000000000002e-06
56	0.45458319783210754	0.0633	0.0866	0.0473	1.0000000000000002e-06
57	0.4548773169517517	0.0639	0.0871	-0.0046	1.0000000000000002e-06
58	0.45440155267715454	0.0627	0.0864	0.0553	1.0000000000000002e-06
59	0.45448538661003113	0.0631	0.0865	0.0368	1.0000000000000002e-07
60	0.4544091522693634	0.0629	0.0863	0.0471	1.0000000000000002e-07
61	0.4542348086833954	0.0624	0.0860	0.0928	1.0000000000000002e-07
62	0.4545469284057617	0.0632	0.0866	0.0286	1.0000000000000002e-07

Framework Versions

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

lombardata/bd_ortho-DinoVdeau-large-2024_11_27-batch-size64_freeze_probs