Ziboiai is a fine-tuned version of Ziboiai-large-2024_10_31-prova_batch-size32_freeze_probs. It achieves the following results on the test set:

Loss: 0.6188
F1 Micro: 0.9261
F1 Macro: 0.8546
Accuracy: 0.1600
RMSE: 0.3415
MAE: 0.3061
R2: -1.5955

Class	F1 per class
Acropore_branched	0.8966
Acropore_digitised	0.6301
Acropore_tabular	1.0000
Algae	1.0000
Dead_coral	0.8395
Fish	0.8861
Millepore	1.0000
No_acropore_encrusting	1.0000
No_acropore_massive	0.0000
No_acropore_sub_massive	0.8571
Rock	1.0000
Rubble	1.0000
Sand	1.0000

Model description

Ziboiai is a model built on top of Ziboiai-large-2024_10_31-prova_batch-size32_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	41	32	33	106
Acropore_digitised	15	14	14	43
Acropore_tabular	5	8	7	20
Algae	50	50	50	150
Dead_coral	25	28	30	83
Fish	34	24	31	89
Millepore	1	0	0	1
No_acropore_encrusting	1	0	0	1
No_acropore_massive	2	5	5	12
No_acropore_sub_massive	27	28	27	82
Rock	45	47	45	137
Rubble	40	45	44	129
Sand	42	46	45	133

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

Number of Epochs: 40.0
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	R2	Learning Rate
1	0.7150455713272095	0.3848940134048462	0.40997111797332764	-20.29086685180664	0.001
2	0.7314126491546631	0.3895121216773987	0.4163060486316681	-21.218204498291016	0.001
3	0.7726277112960815	0.40413352847099304	0.4320966601371765	-24.822391510009766	0.001
4	0.7917326092720032	0.4094983637332916	0.4379725754261017	-26.581586837768555	0.001
5	0.7852649092674255	0.402120441198349	0.43184274435043335	-26.95589256286621	0.001
6	0.7647674679756165	0.3905399441719055	0.42244094610214233	-24.40153694152832	0.001
7	0.7391812205314636	0.376028835773468	0.41028541326522827	-22.557889938354492	0.001
8	0.7115270495414734	0.36385056376457214	0.39825379848480225	-20.067392349243164	0.0001
9	0.6896975040435791	0.35347798466682434	0.3878582715988159	-18.16646385192871	0.0001
10	0.6777035593986511	0.34683120250701904	0.3818005323410034	-16.94469451904297	0.0001
11	0.6701759099960327	0.3423532247543335	0.3779585659503937	-16.037521362304688	0.0001
12	0.663905918598175	0.3388546407222748	0.37438222765922546	-15.605177879333496	0.0001
13	0.656491219997406	0.3345881700515747	0.3702985942363739	-14.805088996887207	0.0001
14	0.6501385569572449	0.33100754022598267	0.3668138384819031	-14.231175422668457	0.0001
15	0.6467865705490112	0.32885220646858215	0.36475783586502075	-14.07986831665039	0.0001
16	0.6471170783042908	0.3288896679878235	0.3650059998035431	-14.255745887756348	0.0001
17	0.6435126662254333	0.3268200755119324	0.36310678720474243	-14.059813499450684	0.0001
18	0.6437923908233643	0.3269612491130829	0.36342939734458923	-14.036934852600098	0.0001
19	0.6399621367454529	0.3249860107898712	0.36136963963508606	-13.81522274017334	0.0001
20	0.6391971707344055	0.3246455192565918	0.3608955144882202	-13.710391998291016	0.0001
21	0.6386714577674866	0.32462170720100403	0.3606450855731964	-13.809860229492188	0.0001
22	0.6388444304466248	0.3243348002433777	0.36056435108184814	-13.849721908569336	0.0001
23	0.6361631155014038	0.3227779269218445	0.35895633697509766	-13.562189102172852	0.0001
24	0.635435163974762	0.3223152160644531	0.35847193002700806	-13.645319938659668	0.0001
25	0.6344550848007202	0.32144099473953247	0.35783687233924866	-13.602314949035645	0.0001
26	0.6348865628242493	0.3211889863014221	0.3580625355243683	-13.630416870117188	0.0001
27	0.6332749724388123	0.32009246945381165	0.3570806384086609	-13.561347007751465	0.0001
28	0.6295092701911926	0.31767499446868896	0.35479238629341125	-13.23308277130127	0.0001
29	0.6285346746444702	0.3173280954360962	0.35434553027153015	-13.162256240844727	0.0001
30	0.6263097524642944	0.31627562642097473	0.3532228171825409	-12.713174819946289	0.0001
31	0.6272528767585754	0.316723495721817	0.35376670956611633	-12.873921394348145	0.0001
32	0.6294133067131042	0.31807586550712585	0.3550169765949249	-12.935453414916992	0.0001
33	0.6299176216125488	0.3185364603996277	0.35538923740386963	-12.93520736694336	0.0001
34	0.6320692300796509	0.3193182349205017	0.35644862055778503	-13.267191886901855	0.0001
35	0.6279481649398804	0.31752488017082214	0.3541102707386017	-12.99951171875	0.0001
36	0.6280075907707214	0.31736499071121216	0.35407301783561707	-13.00741195678711	0.0001
37	0.6303659081459045	0.3187006115913391	0.35543760657310486	-13.230977058410645	1e-05
38	0.6297122836112976	0.31833118200302124	0.3550592064857483	-12.983016967773438	1e-05
39	0.630845308303833	0.3193325996398926	0.35580796003341675	-13.159842491149902	1e-05
40	0.6291573643684387	0.3182610869407654	0.3547934889793396	-13.069788932800293	1e-05

Framework Versions

Transformers: 4.44.2
Pytorch: 2.4.1+cu121
Datasets: 3.0.0
Tokenizers: 0.19.1

groderg
/

Ziboiai-large-2024_10_31-prova_batch-size32_freeze_probs