lombardata/drone-DinoVdeau-large-2024_08_02-batch-size64_epochs100_freeze

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

• Explained variance: 0.3915
• Loss: 0.3204
• MAE: 0.1180
• MSE: 0.0351
• R2: 0.3909
• RMSE: 0.1874

---

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	2371	782	785	3938
Acropore_digitised	1693	580	579	2852
Acropore_sub_massive	353	99	97	549
Acropore_tabular	1112	420	410	1942
Algae	13150	4386	4405	21941
Dead_coral	6824	2242	2250	11316
Millepore	1543	611	631	2785
No_acropore_encrusting	2799	1044	1041	4884
No_acropore_massive	6578	2216	2170	10964
No_acropore_sub_massive	5252	1802	1793	8847
Rock	13532	4529	4529	22590
Rubble	12641	4222	4231	21094
Sand	13315	4438	4438	22191

---

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• Number of Epochs: 100
• 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	Explained Variance	Validation Loss	MAE	MSE	R2	RMSE	Learning Rate
1	0.265	0.348	0.135	0.043	0.26	0.207	0.001
2	0.284	0.342	0.134	0.041	0.279	0.203	0.001
3	0.303	0.339	0.13	0.04	0.301	0.201	0.001
4	0.287	0.344	0.14	0.042	0.267	0.205	0.001
5	0.319	0.335	0.126	0.04	0.315	0.199	0.001
6	0.324	0.335	0.124	0.039	0.321	0.198	0.001
7	0.319	0.335	0.129	0.039	0.315	0.198	0.001
8	0.325	0.334	0.128	0.039	0.324	0.198	0.001
9	0.335	0.333	0.127	0.039	0.332	0.197	0.001
10	0.33	0.334	0.13	0.039	0.327	0.197	0.001
11	0.335	0.332	0.126	0.039	0.334	0.197	0.001
12	0.336	0.332	0.127	0.038	0.334	0.196	0.001
13	0.332	0.333	0.127	0.039	0.327	0.198	0.001
14	0.334	0.333	0.127	0.039	0.33	0.197	0.001
15	0.339	0.331	0.125	0.038	0.338	0.196	0.001
16	0.34	0.332	0.129	0.038	0.339	0.196	0.001
17	0.341	0.331	0.127	0.038	0.34	0.196	0.001
18	0.335	0.332	0.126	0.039	0.332	0.197	0.001
19	0.331	0.334	0.125	0.04	0.322	0.199	0.001
20	0.331	0.333	0.125	0.039	0.328	0.197	0.001
21	0.336	0.332	0.127	0.038	0.336	0.196	0.001
22	0.339	0.331	0.125	0.038	0.337	0.196	0.001
23	0.335	0.333	0.127	0.039	0.331	0.197	0.001
24	0.351	0.329	0.124	0.038	0.35	0.194	0.0001
25	0.353	0.329	0.123	0.037	0.352	0.193	0.0001
26	0.356	0.327	0.123	0.037	0.355	0.193	0.0001
27	0.358	0.328	0.123	0.037	0.357	0.192	0.0001
28	0.36	0.327	0.122	0.037	0.359	0.192	0.0001
29	0.357	0.328	0.123	0.037	0.356	0.193	0.0001
30	0.361	0.326	0.123	0.037	0.361	0.192	0.0001
31	0.362	0.326	0.122	0.037	0.362	0.192	0.0001
32	0.362	0.326	0.122	0.037	0.362	0.191	0.0001
33	0.364	0.326	0.122	0.037	0.363	0.191	0.0001
34	0.363	0.326	0.122	0.037	0.363	0.191	0.0001
35	0.363	0.326	0.123	0.037	0.363	0.191	0.0001
36	0.364	0.326	0.123	0.037	0.364	0.191	0.0001
37	0.37	0.325	0.122	0.036	0.37	0.19	0.0001
38	0.362	0.326	0.123	0.037	0.362	0.191	0.0001
39	0.37	0.325	0.121	0.036	0.369	0.19	0.0001
40	0.368	0.325	0.122	0.036	0.368	0.191	0.0001
41	0.369	0.325	0.121	0.036	0.369	0.19	0.0001
42	0.369	0.325	0.122	0.036	0.369	0.19	0.0001
43	0.373	0.324	0.12	0.036	0.373	0.19	0.0001
44	0.374	0.324	0.121	0.036	0.374	0.19	0.0001
45	0.373	0.324	0.121	0.036	0.372	0.19	0.0001
46	0.375	0.324	0.12	0.036	0.374	0.19	0.0001
47	0.376	0.323	0.119	0.036	0.375	0.189	0.0001
48	0.376	0.324	0.12	0.036	0.375	0.189	0.0001
49	0.376	0.324	0.121	0.036	0.376	0.189	0.0001
50	0.375	0.324	0.12	0.036	0.374	0.19	0.0001
51	0.373	0.324	0.12	0.036	0.371	0.19	0.0001
52	0.377	0.323	0.12	0.036	0.376	0.189	0.0001
53	0.376	0.324	0.121	0.036	0.375	0.189	0.0001
54	0.375	0.323	0.121	0.036	0.375	0.189	0.0001
55	0.378	0.323	0.119	0.036	0.377	0.189	0.0001
56	0.377	0.323	0.119	0.036	0.376	0.189	0.0001
57	0.377	0.323	0.119	0.036	0.377	0.189	0.0001
58	0.38	0.323	0.12	0.036	0.38	0.189	0.0001
59	0.375	0.323	0.119	0.036	0.375	0.189	0.0001
60	0.379	0.323	0.119	0.036	0.378	0.189	0.0001
61	0.377	0.323	0.12	0.036	0.376	0.189	0.0001
62	0.379	0.323	0.119	0.036	0.378	0.189	0.0001
63	0.377	0.323	0.12	0.036	0.377	0.189	0.0001
64	0.379	0.323	0.119	0.036	0.378	0.189	0.0001
65	0.38	0.322	0.12	0.036	0.379	0.188	1e-05
66	0.381	0.322	0.119	0.035	0.381	0.188	1e-05
67	0.381	0.322	0.118	0.035	0.381	0.188	1e-05
68	0.382	0.322	0.119	0.035	0.382	0.188	1e-05
69	0.382	0.322	0.119	0.035	0.382	0.188	1e-05
70	0.382	0.322	0.119	0.035	0.382	0.188	1e-05
71	0.381	0.322	0.119	0.035	0.381	0.188	1e-05
72	0.383	0.322	0.118	0.035	0.382	0.188	1e-05
73	0.382	0.322	0.118	0.035	0.382	0.188	1e-05
74	0.382	0.322	0.119	0.035	0.381	0.188	1e-05
75	0.382	0.322	0.119	0.035	0.381	0.188	1e-05
76	0.381	0.322	0.119	0.035	0.381	0.188	1e-05
77	0.381	0.322	0.117	0.035	0.381	0.188	1.0000000000000002e-06
78	0.382	0.322	0.118	0.035	0.381	0.188	1.0000000000000002e-06
79	0.381	0.322	0.119	0.035	0.381	0.188	1.0000000000000002e-06
80	0.382	0.322	0.118	0.035	0.382	0.188	1.0000000000000002e-06
81	0.382	0.322	0.118	0.035	0.382	0.188	1.0000000000000002e-06
82	0.383	0.322	0.119	0.035	0.382	0.188	1.0000000000000002e-06

---

CO2 Emissions

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

• Emissions: 0.25264438000117284 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: 4.41.1
• Pytorch: 2.3.0+cu121
• Datasets: 2.19.1
• Tokenizers: 0.19.1