bikit-models

Please visit Building Inspection Toolkit for further details on how to use these models.

New Usage daclnets

from bikit_models.daclnet import build_dacl
from bikit_models.utils import process_img_daclnet, make_prediciton

img_path = "assets/image_0000468_crop_0000001.png"
cp_path = "models/checkpoints/codebrim-classif-balancedcodebrim-classif-balanced_ResNet50_hta.pth"
model, cat_to_name = build_dacl(cp_path=cp_path)
img = process_img_daclnet(img_path)
make_prediciton(model, img, cat_to_name)

New Usage vistranet_0

from bikit_models.vistranet_0 import build_vistra_0
from bikit_models.utils import process_img_vistranet, make_prediciton

img_path = "assets/image_0000468_crop_0000001.png"
cp_path = "models/checkpoints/codebrim-classif-balanced/codebrim-classif-balanced_ViT_s8_0.pth"
model, cat_to_name = build_vistra_0(cp_path=cp_path)
img = process_img_vistranet(img_path)
make_prediciton(model, img, cat_to_name)

New Usage vistranet_1

from bikit_models.vistranet_1 import build_vistra_1
from bikit_models.utils import process_img_vistranet, make_prediciton

img_path = "assets/image_0000468_crop_0000001.png"
cp_path = "models/checkpoints/codebrim-classif-balanced/codebrim-classif-balanced_ViT_s8_1.pth"
model, cat_to_name = build_vistra_1(cp_path=cp_path)
img = process_img_vistranet(img_path)
make_prediciton(model, img, cat_to_name, activation=False) # activation is already performed 

Old Usage

from bikit.utils import load_model, load_img_from_url
from bikit.models import make_prediction

img = load_img_from_url("https://github.com/phiyodr/building-inspection-toolkit/raw/master/bikit/data/11_001990.jpg")
model, metadata = load_model("codebrim-classif-balanced_MobileNetV3-Large_hta", add_metadata=True)
prob, pred = make_prediction(model, img, metadata, print_predictions=True, preprocess_image=True)

Performances of available models

Cp-name	TrainedOn	TrainCropSize	MetricsReportedOn	ExactMatchRatio	F1	Precision	Recall	Accuracy	AUROC	Recall-NoDamage	Recall-Crack	Recall-Efflorescence	Recall-Spalling	Recall-BarsExposed	Recall-Rust	Scaling	Other
codebrim-classif_MobileNetV3-Large_hta	codebrim-classif	224	itself	70.57	83.04	86.27	81.07	92.25	96.67	94	84	82.67	65.1	84.67	76
codebrim-classif-balanced_ViT_s8_1	codebrim-classif-balanced	224	itself	79.59	89.82	91.3	88.54	86.65		94	89.33	85.91	84	90	88
codebrim-classif-balanced_ViT_s8_0	codebrim-classif-balanced	224	itself	77.53	89.62	89.65	89.76	86.05		95.33	92	84.56	90.67	90.67	85.33
codebrim-classif-balanced_ResNet50_hta	codebrim-classif-balanced	224	itself	71.36	84.13	85.33	83.2	92.59	96.99	93.33	85.33	77.18	84.67	85.33	73.33
codebrim-classif-balanced_MobileNetV3-Large_hta	codebrim-classif-balanced	224	itself	68.99	82.77	84.36	81.75	91.98	96.45	94	80	72.48	84.67	86.67	72.67
codebrim-classif-balanced_EfficientNetV1-B0_hta	codebrim-classif-balanced	224	itself	65.66	81	80.33	82.52	90.88	96.06	90	77.33	67.79	88.67	92.67	78.67
dacl1k_MobileNetV3-Large_dhb	dacl1k	512	itself	31.51	66.94	76.24	60.49	78.54	83.41	73.91	42.50	52.27	68.89	56.60	68.75
dacl1k_MobileNetV3-Large_hta	dacl1k	224	itself	23.29	56.94	75.72	46.95	76.18	82.58	65.22	22.5	43.18	44.44	35.85	70.54
mcds_bikit_MobileNetV3-Large_hta	mcds_bikit	224	itself	54.44	65.52	79.48	59.44	90.65	93.67	70	76.67	90	58.89	21.67	68.33	43.33	46.67
mcds_bikit_EfficientNetV1-B0_dha	mcds_bikit	224	itself	51.85	64.55	77.72	58.06	90.23	91.91	46.67	73.33	80	61.11	38.33	75	43.33	46.67
mcds_bikit_ResNet50_dha	mcds_bikit	224	itself	48.15	62.33	80.88	54.93	89.81	93.07	66.67	73.33	86.67	44.44	23.33	65	36.67	43.33
meta2_MobileNetV3-Large_hta	meta2	224	itself	70.41	82.99	87.43	80.1	92.39	96.5	94.44	88.33	70.39	82.22	68.57	76.67
meta2+dacl1k_MobileNetV3-Large_dhb	meta2+dacl1k	512	dacl1k	31.51	67.08	82.09	57.88	80.44	85.77	65.22	31.25	54.55	65.56	56.60	74.11
meta2+dacl1k_MobileNetV3-Large_hta	meta2+dacl1k	224	itself	49.22	66.48	72.17	66.48	85.45	89.27	32.35	83.13	61.8	78.33	65.02	78.26
meta3_MobileNetV3-Large_hta	meta3	224	itself	81.52	85.28	90.84	81.2	95.31	98.03	97.29	93.75	65.92	82.59	71.43	76.19
meta3+dacl1k_MobileNetV3-Large_dhb	meta3+dacl1k	512	dacl1k	32.42	68.94	83.03	60.24	81.13	84.93	67.39	36.25	50.00	73.33	60.38	74.11
meta3+dacl1k_MobileNetV3-Large_hta	meta3+dacl1k	224	itself	75.22	82.14	90.41	76.18	93.4	96.85	93.54	85.54	60.67	78.33	65.4	73.6
meta4_MobileNetV3-Large_hta	meta4	224	itself	77.84	79.4	87.68	74.85	93.01	97.69	99.17	60.82	67.6	81.48	69.52	70.48
meta4+dacl1k_MobileNetV3-Large_hta	meta4+dacl1k	224	itself	76.81	76.44	86.19	71.11	92.76	97.4	98.42	61.85	59.18	73.61	59.7	73.91

** The performance of the models trained on codebrim-classif-balanced dataset in the bikit-models repo differ from the original bikit paper due to sanity changes in bikit. The original models from the paper can be found at dacl-demo repo.