diyago / smokefire-classification Goto Github PK

1. Current results

TTA results increase metrics on test (flip_lr and flip_ud)

OOF metrics:

• MEAN METRIC: 0.953 std 0.051
• ALL METRICS: [0.999, 0.976, 0.868, 0.97]

Test metrics:

• ROC AUC 0.947
• Precision 0.816
• Recall 0.74

TRAIN:

• ROC AUC 0.996

• Precision 0.937

• Recall 0.975

• F1_score 0.956

• MAP 0.997

• TEST:

• ROC AUC 0.829

• Precision 0.471

• Recall 1.0

• F1_score 0.64

• MAP 0.833

3. Training data

Fire: 31 unique location, only 7 with target 1. 2288 - total annotated images Smoke train: 4 location 118 correct 54 incorrect Smoke val: 2 location 33 correct, 18 incorect

3. To train model

a. Prepare data by prepare_data.py

b. Adjust config in config/config_classification.yml

c. train models run train.py

d. Watch TensorBoard logs tensorboard --logdir ./lightning_logs/

e. All inference and deploy param specified in deploy_model folder

d. Collect up-to-date requirements.txt call pipreqs --force