groderg/SegForCoral-b2-2025_06_03_30567-bs16_refine

SegForCoral-b2-2025_06_03_30567-bs16_refine is a fine-tuned version of nvidia/mit-b2.

---

Model description

SegForCoral-b2-2025_06_03_30567-bs16_refine is a model built on top of nvidia/mit-b2 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 procedure

Training hyperparameters

The following hyperparameters were used during training:

• Number of Epochs: 32.0
• Learning Rate: 1e-05
• 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: No
• Data Augmentation: No

Training results

Epoch	Validation Loss	Learning Rate
1	0.7078850269317627	1e-05
2	0.6609442234039307	1e-05
3	0.6220470666885376	1e-05
4	0.5949200987815857	1e-05
5	0.5844070315361023	1e-05
6	0.5725482702255249	1e-05
7	0.5633664131164551	1e-05
8	0.5647215247154236	1e-05
9	0.5570588707923889	1e-05
10	0.5523818135261536	1e-05
11	0.5505710244178772	1e-05
12	0.5467283725738525	1e-05
13	0.5548911094665527	1e-05
14	0.545318067073822	1e-05
15	0.5459777116775513	1e-05
16	0.5529848337173462	1e-05
17	0.5465011596679688	1e-05
18	0.5460110902786255	1e-05
19	0.5581657290458679	1e-05
20	0.5426208972930908	1e-05
21	0.5669962167739868	1e-05
22	0.5413931608200073	1e-05
23	0.560343325138092	1e-05
24	0.5521411895751953	1e-05
25	0.5587380528450012	1e-05
26	0.5514724850654602	1e-05
27	0.552064061164856	1e-05
28	0.5605921745300293	1e-05
29	0.5458650588989258	1.0000000000000002e-06
30	0.5459087491035461	1.0000000000000002e-06
31	0.547154426574707	1.0000000000000002e-06
32	0.545670747756958	1.0000000000000002e-06

---

Test results

See https://github.com/SeatizenDOI/the-point-is-the-mask/blob/master/config_base.json to get all the data to perform the same results. Use python train.py -oe

📊 Evaluating zone: config/drone_test_polygon_troudeau.geojson

✅ Pixel Accuracy: 0.9297

✅ Mean Accuracy : 0.8092

✅ Mean IoU : 0.5609

Pixel Accuracy Per Class:

• Acropore_branched: 0.8005
• Acropore_tabular: 0.8930
• No_acropore_massive: 0.8727
• No_acropore_sub_massive: 0.5100
• Sand: 0.9695

IoU Per Class:

• Acropore_branched: 0.2471
• Acropore_tabular: 0.5531
• No_acropore_massive: 0.5913
• No_acropore_sub_massive: 0.4544
• Sand: 0.9586

📊 Evaluating zone: config/drone_test_polygon_stleu.geojson

✅ Pixel Accuracy: 0.8016

✅ Mean Accuracy : 0.7548

✅ Mean IoU : 0.5354

Pixel Accuracy Per Class:

• Acropore_branched: 0.7272
• Acropore_tabular: 0.0000
• No_acropore_massive: 0.9735
• No_acropore_sub_massive: 0.3808
• Sand: 0.9376

IoU Per Class:

• Acropore_branched: 0.4750
• Acropore_tabular: 0.0000
• No_acropore_massive: 0.4427
• No_acropore_sub_massive: 0.3150
• Sand: 0.9089

📦 Micro-Averaged Metrics Across Zones (all pixels):

Pixel Accuracy Per Class:

• Acropore_branched: 0.7312
• Acropore_tabular: 0.8930
• No_acropore_massive: 0.9434
• No_acropore_sub_massive: 0.4130
• Sand: 0.9549

IoU Per Class:

• Acropore_branched: 0.4505
• Acropore_tabular: 0.4017
• No_acropore_massive: 0.4757
• No_acropore_sub_massive: 0.3479
• Sand: 0.9356

✅ Pixel Accuracy: 0.8607

✅ Mean Accuracy : 0.7871

✅ Mean IoU : 0.5223

---

Framework Versions

• Transformers: 4.51.3
• Pytorch: 2.6.0+cu124
• Datasets: 3.6.0
• Tokenizers: 0.21.1