README.md

metadata

language:
  - eng
license: cc0-1.0
tags:
  - multilabel-image-classification
  - multilabel
  - generated_from_trainer
base_model: >-
  drone-DinoVdeau-produttoria-probabilities-large-2024_11_06-batch-size16_freeze_probs
model-index:
  - name: >-
      drone-DinoVdeau-produttoria-probabilities-large-2024_11_06-batch-size16_freeze_probs
    results: []

drone-DinoVdeau-produttoria-probabilities is a fine-tuned version of drone-DinoVdeau-produttoria-probabilities-large-2024_11_06-batch-size16_freeze_probs. It achieves the following results on the test set:

• Loss: 0.3261
• F1 Micro: 0.8621
• F1 Macro: 0.8264
• Accuracy: 0.1682
• RMSE: 0.2445
• MAE: 0.1621
• R2: 0.4057

Class	F1 per class
Acropore_branched	0.8063
Acropore_digitised	0.7335
Acropore_tabular	0.6247
Algae	0.9859
Dead_coral	0.8424
Fish	0.7464
Millepore	0.6453
No_acropore_encrusting	0.7292
No_acropore_massive	0.8681
No_acropore_sub_massive	0.8092
Rock	0.9925
Rubble	0.9693
Sand	0.9904

---

Model description

drone-DinoVdeau-produttoria-probabilities is a model built on top of drone-DinoVdeau-produttoria-probabilities-large-2024_11_06-batch-size16_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	2028	684	686	3398
Acropore_digitised	2006	735	717	3458
Acropore_tabular	1237	461	451	2149
Algae	11086	3671	3675	18432
Dead_coral	6354	2161	2147	10662
Fish	4032	1430	1430	6892
Millepore	1943	783	772	3498
No_acropore_encrusting	2663	986	957	4606
No_acropore_massive	6897	2375	2375	11647
No_acropore_sub_massive	5416	1988	1958	9362
Rock	11164	3726	3725	18615
Rubble	10687	3570	3572	17829
Sand	11151	3726	3723	18600

---

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• Number of Epochs: 45.0
• Learning Rate: 0.001
• Train Batch Size: 16
• Eval Batch Size: 16
• 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
0	N/A	0.0000	0.0000	0.0000	0.001
1	0.36246591806411743	0.1880	0.2669	0.2744	0.001
2	0.3457428216934204	0.1685	0.2560	0.3367	0.001
3	0.3518487811088562	0.1747	0.2597	0.3157	0.001
4	0.3507988750934601	0.1751	0.2563	0.3345	0.001
5	0.3436409533023834	0.1696	0.2546	0.3371	0.001
6	0.35096481442451477	0.1767	0.2598	0.3175	0.001
7	0.3412320613861084	0.1750	0.2538	0.3471	0.001
8	0.3456409275531769	0.1678	0.2561	0.3435	0.001
9	0.3425351679325104	0.1741	0.2545	0.3409	0.001
10	0.33964109420776367	0.1711	0.2525	0.3583	0.001
11	0.34479108452796936	0.1721	0.2542	0.3498	0.001
12	0.3415849804878235	0.1767	0.2527	0.3577	0.001
13	0.33990854024887085	0.1677	0.2527	0.3523	0.001
14	0.34520208835601807	0.1746	0.2540	0.3443	0.001
15	0.34849879145622253	0.1801	0.2568	0.3333	0.001
16	0.34347954392433167	0.1718	0.2537	0.3473	0.001
17	0.341246634721756	0.1711	0.2508	0.3633	0.0001
18	0.3398562967777252	0.1708	0.2507	0.3649	0.0001
19	0.3332718312740326	0.1675	0.2483	0.3775	0.0001
20	0.333162784576416	0.1688	0.2478	0.3810	0.0001
21	0.3324449062347412	0.1673	0.2476	0.3810	0.0001
22	0.3320053517818451	0.1671	0.2472	0.3836	0.0001
23	0.3301050662994385	0.1658	0.2461	0.3890	0.0001
24	0.3298528492450714	0.1648	0.2458	0.3899	0.0001
25	0.32962867617607117	0.1641	0.2458	0.3903	0.0001
26	0.32889437675476074	0.1632	0.2454	0.3926	0.0001
27	0.33042922616004944	0.1674	0.2461	0.3891	0.0001
28	0.32880541682243347	0.1645	0.2451	0.3955	0.0001
29	0.3293789327144623	0.1656	0.2451	0.3961	0.0001
30	0.33135533332824707	0.1684	0.2464	0.3914	0.0001
31	0.32911789417266846	0.1608	0.2457	0.3904	0.0001
32	0.3289436399936676	0.1631	0.2453	0.3959	0.0001
33	0.3271527588367462	0.1628	0.2444	0.3972	0.0001
34	0.32699429988861084	0.1621	0.2443	0.3976	0.0001
35	0.32638314366340637	0.1615	0.2439	0.3987	0.0001
36	0.3293066918849945	0.1656	0.2455	0.3946	0.0001
37	0.3271186649799347	0.1597	0.2442	0.3996	0.0001
38	0.32695677876472473	0.1613	0.2437	0.4022	0.0001
39	0.33263665437698364	0.1575	0.2438	0.4007	0.0001
40	0.33278176188468933	0.1651	0.2442	0.4003	0.0001
41	0.33069443702697754	0.1627	0.2435	0.4031	0.0001
42	0.3310275375843048	0.1641	0.2436	0.4030	1e-05
43	0.32956016063690186	0.1603	0.2429	0.4052	1e-05
44	0.33022987842559814	0.1625	0.2432	0.4038	1e-05
45	0.3266430199146271	0.1617	0.2430	0.4047	1e-05

---

Framework Versions

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