lombardata/dinov2-large-2024_05_23-drone_batch-size512_epochs50_freeze

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

• Loss: 0.2361
• F1 Micro: 0.7694
• F1 Macro: 0.4048
• Roc Auc: 0.8448
• Accuracy: 0.1449

---

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.

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

Class	train	val	test	Total
Acropore_branched	1575	562	565	2702
Acropore_digitised	1020	356	370	1746
Acropore_sub_massive	198	56	60	314
Acropore_tabular	659	248	238	1145
Algae_assembly	7175	2447	2430	12052
Algae_drawn_up	439	156	156	751
Algae_limestone	4694	1576	1523	7793
Algae_sodding	7151	2460	2467	12078
Bleached_coral	352	162	150	664
Dead_coral	4615	1589	1553	7757
Living_coral	85	37	28	150
Millepore	860	287	313	1460
No_acropore_encrusting	1978	675	667	3320
No_acropore_massive	4539	1613	1585	7737
No_acropore_sub_massive	3696	1245	1252	6193
Rock	10810	3735	3718	18263
Rubble	9948	3429	3403	16780
Sand	10976	3659	3659	18294
Sea_urchins	400	147	135	682
Sponge	207	59	56	322
Thalassodendron_ciliatum	216	74	70	360
Useless	89	30	30	149

---

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• Number of Epochs: 50
• Learning Rate: 0.001
• Train Batch Size: 512
• Eval Batch Size: 512
• 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	Accuracy	F1 Macro	F1 Micro	Learning Rate
1.0	0.5951732397079468	0.012405938580435224	0.5738973203699311	0.40667627783698285	0.001
2.0	0.4730209410190582	0.06975798250966037	0.7307120964254151	0.4367882492755507	0.001
3.0	0.3240152895450592	0.10738255033557047	0.7498981835953409	0.37702459211637257	0.001
4.0	0.2770342230796814	0.11795810453528574	0.7521195160095482	0.3710481900670742	0.001
5.0	0.25879302620887756	0.11958511287370348	0.7507292550220328	0.3714736793659693	0.001
6.0	0.25328728556632996	0.12182224933902787	0.7520252586099456	0.36304822534346387	0.001
7.0	0.25132349133491516	0.11531421598535692	0.7517183920016662	0.3646331511607325	0.001
8.0	0.2507544159889221	0.12283912955053895	0.7576399892988702	0.38940077262215617	0.001
9.0	0.24785615503787994	0.12751677852348994	0.7549859932265752	0.38290945223752887	0.001
10.0	0.2480766475200653	0.12649989831197886	0.7583163191651716	0.37973264961121395	0.001
11.0	0.24667006731033325	0.12426276184665447	0.7600958878849345	0.3964288145693209	0.001
12.0	0.2459569126367569	0.12507626601586333	0.7564640698455339	0.3733203958034908	0.001
13.0	0.245611771941185	0.1297539149888143	0.7581923944769908	0.38618999344086086	0.001
14.0	0.24649737775325775	0.13707545251169412	0.7526021832952525	0.37084554766098704	0.001
15.0	0.24523988366127014	0.1271100264388855	0.7540528606572888	0.37953234953900117	0.001
16.0	0.24370642006397247	0.1293471629042099	0.7597242635642867	0.39042586476441543	0.001
17.0	0.24466517567634583	0.13158429936953428	0.7525727259224682	0.38542350135117487	0.001
17.857142857142858	N/A	N/A	N/A	N/A	0.001
18.0	0.24544650316238403	0.133211307707952	0.7534316217590239	0.35783734462173733	0.001
19.0	0.2440878450870514	0.13239780353874314	0.7568417082268136	0.3694145346248099	0.001
20.0	0.2453632354736328	0.13605857230018303	0.750895096799091	0.3768127127776539	0.001
21.0	0.24377579987049103	0.12487288997356112	0.760243826841616	0.38961590782494593	0.001
22.0	0.24192409217357635	0.13016066707341875	0.7576183975637929	0.3715634230883189	0.001
23.0	0.24348826706409454	0.12649989831197886	0.7628996647313762	0.3880375815747224	0.001
24.0	0.2413305789232254	0.1342281879194631	0.7561114991428027	0.3896884130115941	0.001
25.0	0.24189460277557373	0.1297539149888143	0.7599182173024102	0.38267978517684004	0.001
26.0	0.2437727451324463	0.12670327435428105	0.7593076827294236	0.3971421437602147	0.001
27.0	0.24182096123695374	0.1309741712426276	0.761437908496732	0.38383597863653807	0.001
28.0	0.24316559731960297	0.13341468375025423	0.7498440155769273	0.3792682503180625	0.001
29.0	0.24201267957687378	0.1366687004270897	0.7621594930458399	0.39596794972011545	0.001
30.0	0.2406790852546692	0.14236322961155176	0.7619565217391304	0.38596411111358153	0.001
31.0	0.24222084879875183	0.13280455562334756	0.7611869607298037	0.3928781445591724	0.001
32.0	0.24304261803627014	0.13117754728492984	0.7516135926480015	0.3912203123758987	0.001
33.0	0.24139608442783356	0.13016066707341875	0.758885526453094	0.38844227395152936	0.001
34.0	0.24039919674396515	0.1354484441732764	0.7624706542289075	0.4037409737349212	0.001
35.0	0.24134761095046997	0.12995729103111653	0.7601615858737297	0.3973020120442106	0.001
35.714285714285715	N/A	N/A	N/A	N/A	0.001
36.0	0.24192169308662415	0.13565182021557862	0.7622066694112803	0.38761085480429286	0.001
37.0	0.2399486005306244	0.1342281879194631	0.7598352387357096	0.3992187594370792	0.001
38.0	0.24004822969436646	0.13300793166564978	0.7607364527387098	0.3932700433432016	0.001
39.0	0.24091550707817078	0.13890583689241406	0.7619087275149901	0.4007929579258356	0.001
40.0	0.23991511762142181	0.1354484441732764	0.76	0.39250375468507387	0.001
41.0	0.2422637641429901	0.12487288997356112	0.7639710985018574	0.40608061408264917	0.001
42.0	0.24256455898284912	0.1309741712426276	0.7568840806286871	0.4005098857996497	0.001
43.0	0.23922023177146912	0.13361805979255645	0.7594289817122102	0.4007981173529554	0.001
44.0	0.24184103310108185	0.13036404311572097	0.7576905272903253	0.40641694858015515	0.001
45.0	0.24105145037174225	0.13788895668090298	0.7591085068536151	0.39055068831340933	0.001
46.0	0.23963303864002228	0.13626194834248526	0.7653508320819534	0.4106196361694743	0.001
47.0	0.23957742750644684	0.13260117958104536	0.7575076348829317	0.3967990889217657	0.001
48.0	0.24231907725334167	0.12873703477730322	0.7563947423325684	0.38777990454974365	0.001
49.0	0.23978127539157867	0.13300793166564978	0.7608376348147216	0.40266317126303974	0.001
50.0	0.23673731088638306	0.14236322961155176	0.7652267908369019	0.4087415721658059	0.0001

---

CO2 Emissions

The estimated CO2 emissions for training this model are documented below:

• Emissions: 0.02562211166966913 grams of CO2
• Source: Code Carbon
• Training Type: fine-tuning
• Geographical Location: Brest, France
• Hardware Used: NVIDIA Tesla V100 PCIe 32 Go

---

Framework Versions

• Transformers: 4.41.0
• Pytorch: 2.3.0+cu118
• Datasets: 2.19.1
• Tokenizers: 0.19.1