Характеристики

Лицензия: Своя лицензия
Архитектура: segformer · dense
Базовая модель: segformer-b5-finetuned-cityscapes-1024-1024
Опубликована: 4 мар 2024

Ссылки

segformer-b5-cityscapes-finetuned-coastTrain

Сегментация изображенийИзображение → mask

Дообученная SegFormer B5 для попиксельной сегментации прибрежных изображений на классы природных поверхностей и застройки.

8скачиваний за 30 дней0лайковвеса обновлены6 мар 2024

Бенчмарки

Как модель справилась с задачами пользователей ailist и что о ней публикуют открытые лидерборды

Публичные оценки

Из открытых лидербордов. Сравнивайте результаты одного теста при одинаковых условиях — отсутствие оценки не значит, что модель хуже.

В подключённых источниках результатов для этой модели пока нет.

Что умеет модель

Описание составлено по карточке модели и метаданным Hugging Face

peldrak/segformer-b5-cityscapes-finetuned-coastTrain — модель семантической сегментации изображений на базе SegFormer B5, дообученная автором на наборе данных peldrak/coastTrain. Она присваивает пиксели изображениям классы прибрежной среды: вода, прибойная пена, осадочные породы, другие природные поверхности, растительность, застройка и неизвестная область. Модель предназначена для анализа прибрежных сцен и картирования их элементов. В оценке на тестовом наборе указаны mean IoU 0,8269 и общая точность 0,9424; IoU для класса Whitewater составляет 0,6328, ниже, чем для большинства перечисленных классов.

Модель использует архитектуру SegFormer B5 и хранится в формате safetensors. Базовая модель была дообучена на Cityscapes с разрешением 1024 × 1024, а эта версия адаптирована для набора coastTrain. Лицензия в метаданных обозначена как other; конкретные условия использования следует сверить с карточкой репозитория.

Сценарии применения

сегментация воды и прибойной пены на снимках побережья
картирование осадочных пород и природных поверхностей
выделение растительности на прибрежных изображениях
разделение природного ландшафта и застройки
анализ прибрежных сцен для мониторинга территории

Ограничения

На тестовом наборе IoU для класса Whitewater ниже, чем для большинства остальных классов; качество на других типах побережья и изображений в карточке не описано. Лицензия указана как other, поэтому перед использованием нужно проверить её условия в репозитории.

Карточка модели

README от автора на Hugging Face, как он его опубликовал

Открыть на Hugging Face

segformer-b5-cityscapes-finetuned-coastTrain

This model is a fine-tuned version of nvidia/segformer-b5-finetuned-cityscapes-1024-1024 on the peldrak/coastTrain dataset. It achieves the following results on the evaluation set:

Loss: 0.1881
Mean Iou: 0.8269
Mean Accuracy: 0.9060
Overall Accuracy: 0.9424
Accuracy Water: 0.9707
Accuracy Whitewater: 0.7995
Accuracy Sediment: 0.8682
Accuracy Other Natural Terrain: 0.8548
Accuracy Vegetation: 0.9256
Accuracy Development: 0.9477
Accuracy Unknown: 0.9753
Iou Water: 0.9366
Iou Whitewater: 0.6328
Iou Sediment: 0.8126
Iou Other Natural Terrain: 0.7202
Iou Vegetation: 0.8474
Iou Development: 0.8731
Iou Unknown: 0.9654
F1 Score: 0.9423

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 6e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss	Mean Iou	Mean Accuracy	Overall Accuracy	Accuracy Water	Accuracy Whitewater	Accuracy Sediment	Accuracy Other Natural Terrain	Accuracy Vegetation	Accuracy Development	Accuracy Unknown	Iou Water	Iou Whitewater	Iou Sediment	Iou Other Natural Terrain	Iou Vegetation	Iou Development	Iou Unknown	F1 Score
1.5137	0.16	20	1.4093	0.2798	0.3909	0.5891	0.7019	0.0298	0.0296	0.0247	0.9567	0.4761	0.5177	0.5641	0.0196	0.0292	0.0213	0.3809	0.4281	0.5158	0.5595
1.055	0.31	40	1.0631	0.3577	0.4550	0.7052	0.8814	0.0052	0.2973	0.0004	0.9572	0.4259	0.6178	0.7052	0.0050	0.2796	0.0004	0.4917	0.4043	0.6175	0.6828
1.0713	0.47	60	0.8083	0.4894	0.5712	0.8248	0.9457	0.0002	0.6146	0.0	0.8897	0.6666	0.8818	0.8231	0.0002	0.4732	0.0	0.6786	0.5733	0.8773	0.8106
1.1381	0.62	80	0.7066	0.4937	0.5901	0.8202	0.8788	0.0	0.5956	0.0051	0.9164	0.8435	0.8913	0.8319	0.0	0.4887	0.0051	0.6175	0.6277	0.8850	0.8107
1.7365	0.78	100	0.6244	0.5151	0.6193	0.8271	0.8129	0.0	0.8459	0.0000	0.8727	0.9082	0.8954	0.7790	0.0	0.5186	0.0000	0.7262	0.6899	0.8922	0.8218
0.7307	0.93	120	0.5070	0.5352	0.6115	0.8598	0.9605	0.0	0.6408	0.0	0.9080	0.8753	0.8955	0.8718	0.0	0.5549	0.0	0.6907	0.7387	0.8904	0.8458
1.1675	1.09	140	0.5019	0.5353	0.6205	0.8593	0.9221	0.0	0.6875	0.0000	0.9250	0.8983	0.9108	0.8673	0.0	0.5808	0.0000	0.7074	0.7402	0.8514	0.8465
0.701	1.24	160	0.4414	0.5432	0.6326	0.8671	0.9344	0.0007	0.7689	0.0004	0.8593	0.9636	0.9009	0.8799	0.0007	0.6322	0.0004	0.7217	0.6796	0.8876	0.8559
0.7284	1.4	180	0.4495	0.5437	0.6268	0.8576	0.8939	0.0	0.7874	0.0233	0.9336	0.8510	0.8985	0.8509	0.0	0.5802	0.0233	0.7158	0.7453	0.8902	0.8485
0.4898	1.55	200	0.3884	0.5559	0.6380	0.8772	0.9541	0.0	0.7934	0.0144	0.8666	0.9343	0.9031	0.8912	0.0	0.6746	0.0144	0.7378	0.6831	0.8901	0.8656
0.3669	1.71	220	0.3874	0.5574	0.6333	0.8752	0.9796	0.0001	0.7557	0.0380	0.8465	0.9239	0.8892	0.8747	0.0001	0.6674	0.0380	0.7369	0.7045	0.8802	0.8625
0.6277	1.86	240	0.3964	0.5600	0.6406	0.8770	0.9221	0.0029	0.8164	0.0	0.9202	0.9157	0.9069	0.8787	0.0029	0.6458	0.0	0.7487	0.7474	0.8965	0.8657
0.3567	2.02	260	0.3927	0.5648	0.6499	0.8825	0.9215	0.0000	0.8792	0.0016	0.8867	0.9531	0.9073	0.8862	0.0000	0.6898	0.0016	0.7659	0.7127	0.8974	0.8717
0.4065	2.17	280	0.3756	0.5633	0.6377	0.8810	0.9551	0.0145	0.7178	0.0058	0.9349	0.9232	0.9125	0.9051	0.0145	0.6369	0.0058	0.7220	0.7586	0.9000	0.8686
0.7596	2.33	300	0.3224	0.5907	0.6599	0.9000	0.9569	0.0001	0.8956	0.0230	0.9042	0.9352	0.9044	0.9019	0.0001	0.7377	0.0230	0.7935	0.7806	0.8977	0.8879
0.3008	2.48	320	0.3164	0.6114	0.6773	0.9011	0.9518	0.0022	0.8608	0.1780	0.9504	0.8886	0.9089	0.8993	0.0022	0.7541	0.1777	0.7957	0.7517	0.8992	0.8919
0.346	2.64	340	0.2927	0.6143	0.6807	0.9064	0.9643	0.0198	0.8652	0.1390	0.9219	0.9389	0.9156	0.9072	0.0197	0.7542	0.1388	0.8234	0.7570	0.8999	0.8967
0.272	2.79	360	0.2769	0.6170	0.6786	0.9098	0.9711	0.0228	0.8956	0.1033	0.9136	0.9310	0.9130	0.9071	0.0226	0.7745	0.1033	0.8178	0.7908	0.9030	0.8992
1.1467	2.95	380	0.2737	0.6153	0.6746	0.9107	0.9669	0.0005	0.9234	0.0834	0.9132	0.9179	0.9170	0.9041	0.0005	0.7624	0.0834	0.8207	0.8251	0.9110	0.8995
0.2723	3.1	400	0.3110	0.6254	0.6931	0.8984	0.9796	0.0115	0.8782	0.2913	0.8216	0.9572	0.9124	0.8902	0.0115	0.7610	0.2882	0.7610	0.7626	0.9032	0.8900
0.5568	3.26	420	0.3337	0.6544	0.7155	0.9012	0.9500	0.0359	0.7990	0.4369	0.9408	0.9111	0.9345	0.8904	0.0356	0.7152	0.4322	0.7726	0.8280	0.9068	0.8948
1.1779	3.41	440	0.2947	0.6237	0.6857	0.9057	0.9679	0.0133	0.7896	0.2042	0.9229	0.9476	0.9541	0.9116	0.0133	0.7095	0.2041	0.7783	0.8135	0.9357	0.8961
0.4611	3.57	460	0.2659	0.6563	0.7149	0.9153	0.9562	0.0761	0.9265	0.2765	0.9315	0.9160	0.9215	0.9051	0.0754	0.7745	0.2759	0.8331	0.8170	0.9133	0.9084
0.3981	3.72	480	0.2964	0.6657	0.7269	0.9046	0.9789	0.0658	0.8862	0.4978	0.8601	0.8802	0.9193	0.8972	0.0648	0.7316	0.4900	0.7955	0.7687	0.9118	0.8991
0.2398	3.88	500	0.2707	0.7112	0.7714	0.9181	0.9553	0.2649	0.8826	0.5232	0.9516	0.8983	0.9236	0.9130	0.2463	0.7796	0.5159	0.8301	0.7813	0.9125	0.9154
0.251	4.03	520	0.2693	0.6980	0.7686	0.9135	0.9680	0.2384	0.8886	0.5380	0.8746	0.9509	0.9215	0.9247	0.2301	0.7644	0.5082	0.8135	0.7432	0.9019	0.9110
0.4024	4.19	540	0.2900	0.6988	0.7641	0.9099	0.9662	0.2524	0.8416	0.5462	0.8835	0.8924	0.9662	0.9022	0.2351	0.7297	0.4979	0.7912	0.8078	0.9278	0.9065
0.3683	4.34	560	0.2852	0.7226	0.7853	0.9153	0.9567	0.3140	0.8381	0.5886	0.9349	0.9398	0.9248	0.9165	0.2981	0.7355	0.5773	0.8091	0.8062	0.9154	0.9132
0.2335	4.5	580	0.3128	0.7163	0.7833	0.9052	0.9370	0.3077	0.8492	0.6411	0.9398	0.8903	0.9182	0.8904	0.2859	0.6985	0.6154	0.8052	0.8088	0.9103	0.9038
1.5312	4.65	600	0.2551	0.7199	0.7784	0.9199	0.9599	0.3232	0.8606	0.4929	0.9314	0.9399	0.9410	0.9213	0.3082	0.7482	0.4884	0.8132	0.8258	0.9342	0.9175
0.3685	4.81	620	0.2286	0.7447	0.8053	0.9245	0.9757	0.3914	0.8468	0.6323	0.9213	0.9265	0.9431	0.9183	0.3353	0.7791	0.5920	0.8197	0.8356	0.9329	0.9226
0.2664	4.96	640	0.3448	0.7422	0.8128	0.9120	0.9329	0.4951	0.8208	0.6164	0.9396	0.9166	0.9679	0.9033	0.4493	0.7043	0.5822	0.7999	0.8185	0.9377	0.9112
0.2662	5.12	660	0.2932	0.7448	0.8158	0.9168	0.9516	0.5045	0.8308	0.6171	0.9215	0.9234	0.9616	0.9128	0.4277	0.7379	0.5809	0.8022	0.8260	0.9261	0.9157
1.1342	5.27	680	0.2631	0.7772	0.8420	0.9267	0.9816	0.5460	0.7974	0.7644	0.9196	0.9208	0.9640	0.9220	0.4753	0.7476	0.6784	0.8126	0.8527	0.9521	0.9253
0.2968	5.43	700	0.2302	0.7884	0.8577	0.9318	0.9537	0.5790	0.8663	0.7551	0.9368	0.9450	0.9680	0.9260	0.5104	0.7765	0.6694	0.8353	0.8467	0.9547	0.9315
0.3812	5.58	720	0.2051	0.7779	0.8372	0.9336	0.9772	0.4866	0.8627	0.7113	0.9184	0.9430	0.9612	0.9189	0.3999	0.8046	0.6717	0.8473	0.8508	0.9517	0.9325
0.1902	5.74	740	0.2044	0.7815	0.8584	0.9330	0.9651	0.5500	0.8494	0.8160	0.9469	0.9219	0.9597	0.9253	0.4528	0.7914	0.6608	0.8459	0.8452	0.9490	0.9325
0.2853	5.89	760	0.2718	0.7844	0.8523	0.9239	0.9479	0.5953	0.8483	0.7417	0.9288	0.9486	0.9554	0.9179	0.5321	0.7413	0.6981	0.8124	0.8399	0.9495	0.9238
0.1833	6.05	780	0.2191	0.7903	0.8613	0.9288	0.9598	0.6288	0.8627	0.7658	0.9246	0.9327	0.9546	0.9209	0.5326	0.7646	0.6852	0.8253	0.8539	0.9494	0.9286
0.2128	6.2	800	0.1897	0.8066	0.8748	0.9392	0.9703	0.6067	0.9032	0.8183	0.9098	0.9555	0.9599	0.9274	0.5356	0.8336	0.7080	0.8593	0.8315	0.9506	0.9388
0.2327	6.36	820	0.2148	0.7999	0.8734	0.9341	0.9531	0.6428	0.9396	0.7743	0.8991	0.9443	0.9604	0.9131	0.5264	0.8037	0.6901	0.8537	0.8609	0.9518	0.9342
1.0808	6.51	840	0.2443	0.7872	0.8597	0.9273	0.9631	0.6632	0.8257	0.7386	0.9302	0.9392	0.9580	0.9267	0.5440	0.7644	0.6807	0.8081	0.8341	0.9524	0.9270
0.2563	6.67	860	0.2376	0.7825	0.8473	0.9288	0.9637	0.5780	0.8420	0.7267	0.9379	0.9213	0.9619	0.9198	0.4945	0.7607	0.6624	0.8265	0.8596	0.9542	0.9282
0.209	6.82	880	0.2008	0.7996	0.8652	0.9353	0.9713	0.5989	0.8857	0.7859	0.9103	0.9475	0.9567	0.9236	0.5153	0.8064	0.7099	0.8452	0.8452	0.9516	0.9349
0.1402	6.98	900	0.2454	0.7832	0.8475	0.9295	0.9673	0.6082	0.8135	0.7021	0.9518	0.9292	0.9603	0.9271	0.5157	0.7680	0.6419	0.8117	0.8636	0.9544	0.9289
0.383	7.13	920	0.2064	0.7999	0.8706	0.9334	0.9720	0.6260	0.8447	0.8266	0.9219	0.9434	0.9599	0.9269	0.5389	0.7808	0.7077	0.8303	0.8620	0.9528	0.9329
0.5665	7.29	940	0.2529	0.7892	0.8836	0.9279	0.9513	0.7850	0.8478	0.7735	0.9296	0.9380	0.9601	0.9205	0.5403	0.7625	0.6755	0.8270	0.8449	0.9539	0.9284
0.2031	7.44	960	0.2445	0.7782	0.8562	0.9276	0.9491	0.6769	0.8635	0.6824	0.9435	0.9101	0.9680	0.9197	0.5111	0.7486	0.6226	0.8341	0.8524	0.9592	0.9277
0.1641	7.6	980	0.1669	0.8201	0.8875	0.9475	0.9762	0.6290	0.9007	0.8526	0.9303	0.9519	0.9719	0.9374	0.5367	0.8539	0.7152	0.8790	0.8560	0.9624	0.9472
0.1214	7.75	1000	0.1696	0.8279	0.8914	0.9479	0.9685	0.7061	0.9217	0.7959	0.9472	0.9376	0.9630	0.9371	0.5860	0.8561	0.7139	0.8779	0.8673	0.9570	0.9478
0.0926	7.91	1020	0.2069	0.8005	0.8677	0.9339	0.9733	0.6746	0.8183	0.7692	0.9485	0.9307	0.9595	0.9405	0.5605	0.7717	0.7126	0.8248	0.8397	0.9540	0.9334
0.4292	8.06	1040	0.1985	0.8140	0.8845	0.9408	0.9620	0.6695	0.9283	0.8152	0.9259	0.9310	0.9596	0.9263	0.5722	0.8215	0.6976	0.8728	0.8538	0.9535	0.9409
0.1741	8.22	1060	0.2038	0.8082	0.8877	0.9352	0.9619	0.7341	0.8483	0.8190	0.9322	0.9496	0.9685	0.9296	0.5834	0.7820	0.7134	0.8411	0.8503	0.9579	0.9351
0.218	8.37	1080	0.1888	0.8054	0.8766	0.9409	0.9770	0.7073	0.8644	0.7571	0.9442	0.9211	0.9653	0.9381	0.5579	0.8125	0.6568	0.8547	0.8599	0.9581	0.9406
0.1687	8.53	1100	0.1887	0.8208	0.8996	0.9424	0.9617	0.7261	0.9048	0.8567	0.9210	0.9514	0.9754	0.9320	0.6109	0.8214	0.6963	0.8650	0.8558	0.9643	0.9425
0.1115	8.68	1120	0.1924	0.8137	0.8967	0.9390	0.9737	0.7728	0.8477	0.8500	0.9296	0.9305	0.9726	0.9374	0.6039	0.7985	0.6894	0.8367	0.8657	0.9640	0.9389
0.2189	8.84	1140	0.1950	0.8102	0.8791	0.9397	0.9718	0.6680	0.8758	0.8029	0.9293	0.9407	0.9655	0.9321	0.5657	0.8183	0.6952	0.8497	0.8525	0.9581	0.9395
0.2046	8.99	1160	0.1955	0.8183	0.8889	0.9409	0.9707	0.7188	0.8764	0.8236	0.9345	0.9287	0.9693	0.9349	0.5922	0.8067	0.7146	0.8500	0.8690	0.9610	0.9407
0.0574	9.15	1180	0.1943	0.8258	0.9017	0.9423	0.9642	0.7756	0.8757	0.8409	0.9354	0.9389	0.9813	0.9362	0.6173	0.8057	0.7304	0.8544	0.8714	0.9653	0.9423
0.1683	9.3	1200	0.1685	0.8342	0.8967	0.9487	0.9695	0.7096	0.9281	0.8261	0.9382	0.9376	0.9676	0.9368	0.6100	0.8515	0.7291	0.8807	0.8726	0.9590	0.9486
0.1265	9.46	1220	0.1684	0.8278	0.8829	0.9482	0.9764	0.6761	0.9042	0.7785	0.9515	0.9278	0.9657	0.9404	0.5943	0.8480	0.7065	0.8732	0.8729	0.9595	0.9479
0.1052	9.61	1240	0.1743	0.8176	0.8969	0.9441	0.9750	0.7880	0.8907	0.7930	0.9261	0.9333	0.9721	0.9343	0.5809	0.8331	0.6799	0.8665	0.8654	0.9634	0.9441
0.1972	9.77	1260	0.2006	0.8135	0.8781	0.9385	0.9731	0.6571	0.8506	0.8227	0.9316	0.9379	0.9735	0.9315	0.5703	0.7969	0.7257	0.8386	0.8669	0.9643	0.9380
0.2727	9.92	1280	0.1952	0.8091	0.8666	0.9412	0.9714	0.6697	0.8829	0.7132	0.9561	0.9058	0.9672	0.9344	0.5762	0.8125	0.6553	0.8544	0.8714	0.9594	0.9407
0.1478	10.08	1300	0.1935	0.8153	0.8976	0.9395	0.9675	0.7368	0.8569	0.8718	0.9352	0.9506	0.9644	0.9317	0.6118	0.7976	0.6876	0.8580	0.8621	0.9584	0.9395
0.2574	10.23	1320	0.2107	0.8154	0.8888	0.9384	0.9586	0.7452	0.8745	0.8039	0.9545	0.9184	0.9668	0.9285	0.5996	0.7961	0.7035	0.8515	0.8681	0.9603	0.9385
0.9414	10.39	1340	0.1919	0.8221	0.8939	0.9428	0.9653	0.7195	0.8869	0.8314	0.9350	0.9456	0.9736	0.9319	0.5980	0.8148	0.7137	0.8644	0.8668	0.9652	0.9427
0.1415	10.54	1360	0.2000	0.8247	0.8946	0.9417	0.9687	0.7613	0.8609	0.8152	0.9407	0.9392	0.9761	0.9386	0.6259	0.8026	0.7224	0.8437	0.8722	0.9673	0.9416
0.1511	10.7	1380	0.1881	0.8269	0.9060	0.9424	0.9707	0.7995	0.8682	0.8548	0.9256	0.9477	0.9753	0.9366	0.6328	0.8126	0.7202	0.8474	0.8731	0.9654	0.9423

Framework versions

Transformers 4.37.0
Pytorch 2.1.2
Datasets 2.18.0
Tokenizers 0.15.1

Как запустить

На своей GPU в ailist. Цена — по текущим предложениям провайдеров

GPU в ailist

Универсальный рантайм · image-segmentation

Можно запустить

Подбираем GPU…

Класс архитектуры: SegformerForSemanticSegmentation
Ревизия весов: 6ecbb8d