drone-DinoVdeau-produttoria-binary is a fine-tuned version of drone-DinoVdeau-produttoria-binary-large-2024_11_04-batch-size64_freeze. It achieves the following results on the test set:

Loss: 0.2920
F1 Micro: 0.8387
F1 Macro: 0.6349
Accuracy: 0.2681

Class	F1 per class
Acropore_branched	0.8103
Acropore_digitised	0.4910
Acropore_tabular	0.3804
Algae	0.9090
Dead_coral	0.7329
Fish	0.6809
Millepore	0.2665
No_acropore_encrusting	0.5949
No_acropore_massive	0.7760
No_acropore_sub_massive	0.6423
Rock	0.9539
Rubble	0.9035
Sand	0.9058

Model description

drone-DinoVdeau-produttoria-binary is a model built on top of drone-DinoVdeau-produttoria-binary-large-2024_11_04-batch-size64_freeze 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	test	val	Total
Acropore_branched	1483	532	519	2534
Acropore_digitised	1085	379	354	1818
Acropore_tabular	486	173	181	840
Algae	9163	3119	3145	15427
Dead_coral	3710	1265	1254	6229
Fish	1456	523	515	2494
Millepore	731	273	297	1301
No_acropore_encrusting	1980	747	745	3472
No_acropore_massive	4438	1625	1617	7680
No_acropore_sub_massive	3036	1122	1091	5249
Rock	10216	3437	3446	17099
Rubble	9345	3093	3120	15558
Sand	9240	3135	3129	15504

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

Number of Epochs: 93.0
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	Validation Loss	Accuracy	F1 Macro	F1 Micro	Learning Rate
1	0.33133190870285034	0.2465	0.8232	0.6028	0.001
2	0.32479074597358704	0.2574	0.8257	0.5804	0.001
3	0.3155537247657776	0.2504	0.8294	0.5895	0.001
4	0.31338146328926086	0.2637	0.8299	0.5890	0.001
5	0.31261566281318665	0.2551	0.8303	0.5945	0.001
6	0.3106238543987274	0.2551	0.8277	0.5928	0.001
7	0.3110489547252655	0.2553	0.8345	0.6182	0.001
8	0.3142373561859131	0.2566	0.8315	0.6097	0.001
9	0.3147599399089813	0.2340	0.8318	0.6138	0.001
10	0.3089006543159485	0.2634	0.8340	0.6072	0.001
11	0.3095955550670624	0.2587	0.8332	0.5998	0.001
12	0.30922698974609375	0.2525	0.8331	0.6163	0.001
13	0.3101893365383148	0.2491	0.8316	0.6016	0.001
14	0.30971381068229675	0.2639	0.8288	0.5952	0.001
15	0.30827441811561584	0.2462	0.8338	0.6153	0.001
16	0.3080792725086212	0.2634	0.8348	0.6102	0.001
17	0.3121964633464813	0.2342	0.8313	0.6059	0.001
18	0.3083283007144928	0.2566	0.8293	0.5870	0.001
19	0.3106550872325897	0.2504	0.8363	0.6266	0.001
20	0.306607723236084	0.2634	0.8337	0.6034	0.001
21	0.3118128776550293	0.2592	0.8271	0.5862	0.001
22	0.30495893955230713	0.2665	0.8329	0.5988	0.001
23	0.30966898798942566	0.2553	0.8340	0.6189	0.001
24	0.3067616820335388	0.2587	0.8318	0.6039	0.001
25	N/A	0.0000	0.0000	0.0000	0.001
26	0.30948761105537415	0.2684	0.8310	0.5952	0.001
27	0.30627554655075073	0.2517	0.8346	0.6156	0.001
28	0.31095221638679504	0.2559	0.8264	0.6011	0.001
29	0.3020830452442169	0.2712	0.8345	0.6083	0.0001
30	0.30176734924316406	0.2744	0.8351	0.6046	0.0001
31	0.3010377883911133	0.2751	0.8376	0.6161	0.0001
32	0.3006710708141327	0.2767	0.8372	0.6172	0.0001
33	0.30090999603271484	0.2710	0.8375	0.6230	0.0001
34	0.29988643527030945	0.2746	0.8379	0.6223	0.0001
35	0.2998106777667999	0.2762	0.8377	0.6148	0.0001
36	0.3001900315284729	0.2783	0.8370	0.6114	0.0001
37	0.299510657787323	0.2770	0.8370	0.6122	0.0001
38	0.2991965413093567	0.2772	0.8377	0.6142	0.0001
39	0.2990879416465759	0.2725	0.8383	0.6138	0.0001
40	0.2994089722633362	0.2759	0.8373	0.6189	0.0001
41	0.29888105392456055	0.2770	0.8396	0.6285	0.0001
42	0.29893118143081665	0.2725	0.8384	0.6237	0.0001
43	0.2987787425518036	0.2762	0.8379	0.6187	0.0001
44	0.2980958819389343	0.2764	0.8382	0.6218	0.0001
45	0.29851531982421875	0.2668	0.8385	0.6295	0.0001
46	0.297795832157135	0.2770	0.8388	0.6196	0.0001
47	0.29809537529945374	0.2741	0.8384	0.6250	0.0001
48	0.29759711027145386	0.2809	0.8387	0.6230	0.0001
49	0.2976347804069519	0.2791	0.8386	0.6252	0.0001
50	N/A	0.0000	0.0000	0.0000	0.0001
51	0.2974720299243927	0.2757	0.8399	0.6319	0.0001
52	0.2965853810310364	0.2733	0.8394	0.6275	0.0001
53	0.2971484959125519	0.2801	0.8385	0.6244	0.0001
54	0.2965540885925293	0.2744	0.8392	0.6268	0.0001
55	0.2974574863910675	0.2762	0.8382	0.6246	0.0001
56	0.2963787019252777	0.2793	0.8392	0.6285	0.0001
57	0.29714566469192505	0.2764	0.8391	0.6324	0.0001
58	0.2968769073486328	0.2738	0.8397	0.6347	0.0001
59	0.29710978269577026	0.2749	0.8404	0.6320	0.0001
60	0.2966577410697937	0.2723	0.8377	0.6251	0.0001
61	0.29699769616127014	0.2757	0.8383	0.6253	0.0001
62	0.29695647954940796	0.2731	0.8385	0.6263	0.0001
63	0.2960326671600342	0.2736	0.8394	0.6280	1e-05
64	0.29593583941459656	0.2723	0.8396	0.6326	1e-05
65	0.29625141620635986	0.2712	0.8407	0.6350	1e-05
66	0.2958442270755768	0.2738	0.8390	0.6261	1e-05
67	0.2959185838699341	0.2767	0.8395	0.6301	1e-05
68	0.2957761883735657	0.2723	0.8398	0.6392	1e-05
69	0.29576200246810913	0.2720	0.8395	0.6347	1e-05
70	0.295784056186676	0.2751	0.8395	0.6321	1e-05
71	0.29544582962989807	0.2757	0.8398	0.6346	1e-05
72	0.29554447531700134	0.2736	0.8398	0.6381	1e-05
73	0.29577216506004333	0.2746	0.8397	0.6336	1e-05
74	0.2956525683403015	0.2751	0.8401	0.6362	1e-05
75	N/A	0.0000	0.0000	0.0000	1e-05
76	0.2954920828342438	0.2731	0.8396	0.6358	1e-05
77	0.29537785053253174	0.2762	0.8396	0.6331	1e-05
78	0.29544633626937866	0.2731	0.8398	0.6383	1e-05
79	0.29545679688453674	0.2738	0.8402	0.6394	1e-05
80	0.2953413426876068	0.2699	0.8403	0.6379	1e-05
81	0.29545214772224426	0.2770	0.8403	0.6335	1e-05
82	0.2955772876739502	0.2733	0.8408	0.6396	1e-05
83	0.2952810227870941	0.2744	0.8392	0.6358	1e-05
84	0.29596611857414246	0.2728	0.8396	0.6398	1e-05
85	0.29555922746658325	0.2751	0.8392	0.6343	1e-05
86	0.29549410939216614	0.2736	0.8402	0.6371	1e-05
87	0.29544368386268616	0.2749	0.8404	0.6396	1e-05
88	0.29540950059890747	0.2746	0.8405	0.6406	1e-05
89	0.295379102230072	0.2731	0.8400	0.6393	1e-05
90	0.29556041955947876	0.2749	0.8402	0.6371	1.0000000000000002e-06
91	0.29565492272377014	0.2710	0.8401	0.6422	1.0000000000000002e-06
92	0.2954220473766327	0.2738	0.8404	0.6401	1.0000000000000002e-06
93	0.29537004232406616	0.2746	0.8398	0.6368	1.0000000000000002e-06

Framework Versions

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

lombardata
/

drone-DinoVdeau-produttoria-binary-large-2024_11_04-batch-size64_freeze