lombardata/drone-DinoVdeau-large-2024_09_16-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: 0.3677
• Loss: 0.3353
• MAE: 0.1229
• MSE: 0.0346
• R2: 0.3673
• RMSE: 0.1861

---

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	1956	650	653	3259
Acropore_digitised	1717	577	575	2869
Acropore_tabular	1105	375	388	1868
Algae	11085	3678	3680	18443
Dead_coral	5888	1953	1958	9799
Fish	3453	1157	1157	5767
Millepore	1779	666	698	3143
No_acropore_encrusting	2726	966	988	4680
No_acropore_massive	6486	2188	2138	10812
No_acropore_sub_massive	5026	1772	1769	8567
Rock	11176	3725	3725	18626
Rubble	10689	3563	3563	17815
Sand	11168	3723	3723	18614
Sea_cucumber	2751	1065	1129	4945
Sea_urchins	651	274	269	1194

---

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	0.254	0.361	0.147	0.043	0.244	0.207	0.001
2	0.297	0.35	0.134	0.04	0.294	0.201	0.001
3	0.301	0.349	0.134	0.04	0.299	0.2	0.001
4	0.303	0.349	0.132	0.04	0.298	0.2	0.001
5	0.309	0.348	0.138	0.04	0.304	0.199	0.001
6	0.319	0.346	0.136	0.039	0.316	0.197	0.001
7	0.316	0.345	0.133	0.039	0.313	0.197	0.001
8	0.322	0.344	0.132	0.038	0.32	0.196	0.001
9	0.328	0.343	0.131	0.038	0.327	0.195	0.001
10	0.331	0.343	0.134	0.038	0.327	0.195	0.001
11	0.327	0.344	0.135	0.039	0.322	0.196	0.001
12	0.335	0.342	0.129	0.038	0.332	0.195	0.001
13	0.333	0.342	0.132	0.038	0.33	0.195	0.001
14	0.327	0.343	0.131	0.038	0.325	0.196	0.001
15	0.333	0.344	0.135	0.038	0.328	0.196	0.001
16	0.331	0.342	0.131	0.038	0.329	0.195	0.001
17	0.332	0.342	0.131	0.038	0.331	0.195	0.001
18	0.326	0.346	0.135	0.039	0.319	0.198	0.001
19	0.343	0.34	0.13	0.037	0.343	0.193	0.0001
20	0.344	0.34	0.128	0.037	0.343	0.193	0.0001
21	0.348	0.339	0.129	0.037	0.348	0.192	0.0001
22	0.349	0.338	0.128	0.037	0.348	0.192	0.0001
23	0.349	0.338	0.129	0.037	0.348	0.192	0.0001
24	0.351	0.338	0.128	0.037	0.35	0.191	0.0001
25	0	0	0	0	0	0	0.0001
26	0.354	0.337	0.128	0.036	0.353	0.191	0.0001
27	0.356	0.337	0.127	0.036	0.355	0.19	0.0001
28	0.356	0.337	0.129	0.036	0.354	0.191	0.0001
29	0.358	0.337	0.127	0.036	0.357	0.19	0.0001
30	0.358	0.337	0.127	0.036	0.357	0.19	0.0001
31	0.357	0.336	0.126	0.036	0.357	0.19	0.0001
32	0.36	0.336	0.127	0.036	0.359	0.19	0.0001
33	0.36	0.336	0.126	0.036	0.359	0.19	0.0001
34	0.361	0.336	0.126	0.036	0.36	0.19	0.0001
35	0.361	0.336	0.127	0.036	0.36	0.19	0.0001
36	0.362	0.336	0.127	0.036	0.361	0.19	0.0001
37	0.364	0.335	0.126	0.036	0.363	0.189	0.0001
38	0.363	0.335	0.125	0.036	0.362	0.189	0.0001
39	0.363	0.336	0.127	0.036	0.362	0.189	0.0001
40	0.363	0.335	0.126	0.036	0.362	0.189	0.0001
41	0.365	0.335	0.126	0.036	0.363	0.189	0.0001
42	0.364	0.335	0.125	0.036	0.362	0.189	0.0001
43	0.364	0.335	0.124	0.036	0.363	0.189	0.0001
44	0.365	0.335	0.125	0.036	0.364	0.189	1e-05
45	0.367	0.335	0.126	0.036	0.366	0.189	1e-05
46	0.367	0.335	0.125	0.036	0.366	0.189	1e-05
47	0.368	0.335	0.125	0.036	0.366	0.189	1e-05
48	0.368	0.335	0.126	0.036	0.366	0.189	1e-05
49	0.368	0.335	0.125	0.036	0.366	0.189	1e-05
50	0	0	0	0	0	0	1e-05
51	0.369	0.334	0.125	0.036	0.368	0.188	1e-05
52	0.368	0.334	0.124	0.036	0.367	0.188	1e-05
53	0.369	0.334	0.125	0.035	0.368	0.188	1e-05
54	0.369	0.334	0.125	0.035	0.368	0.188	1e-05
55	0.368	0.334	0.124	0.036	0.367	0.189	1e-05
56	0.369	0.334	0.125	0.035	0.369	0.188	1e-05
57	0.369	0.334	0.125	0.035	0.369	0.188	1e-05
58	0.369	0.334	0.125	0.035	0.368	0.188	1e-05
59	0.37	0.334	0.124	0.035	0.37	0.188	1e-05
60	0.371	0.334	0.125	0.035	0.37	0.188	1e-05
61	0.37	0.334	0.125	0.035	0.369	0.188	1e-05
62	0.371	0.334	0.124	0.035	0.369	0.188	1e-05
63	0.369	0.334	0.125	0.035	0.368	0.188	1e-05
64	0.37	0.334	0.125	0.035	0.369	0.188	1e-05
65	0.369	0.334	0.124	0.035	0.368	0.188	1e-05
66	0.371	0.334	0.124	0.035	0.369	0.188	1e-05
67	0.371	0.334	0.124	0.035	0.37	0.188	1.0000000000000002e-06
68	0.37	0.334	0.125	0.035	0.369	0.188	1.0000000000000002e-06
69	0.371	0.334	0.126	0.035	0.369	0.188	1.0000000000000002e-06
70	0.371	0.334	0.124	0.035	0.37	0.188	1.0000000000000002e-06
71	0.371	0.334	0.125	0.035	0.369	0.188	1.0000000000000002e-06
72	0.37	0.334	0.124	0.035	0.37	0.188	1.0000000000000002e-06
73	0.371	0.334	0.125	0.035	0.369	0.188	1.0000000000000002e-06
74	0.371	0.334	0.125	0.035	0.37	0.188	1.0000000000000002e-07
75	0	0	0	0	0	0	1.0000000000000002e-07
76	0.37	0.334	0.124	0.035	0.37	0.188	1.0000000000000002e-07
77	0.37	0.334	0.124	0.035	0.369	0.188	1.0000000000000002e-07

---

CO2 Emissions

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

• Emissions: 0.13788314685965944 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