levanpon98 / zaloaihackathon2019

Nine Dash Detection

This is an implementation of EfficientDet for object detection on Keras and Tensorflow. The project is based on fizyr/keras-retinanet and the qubvel/efficientnet. The pretrained EfficientNet weights files are downloaded from Callidior/keras-applications/releases

Thanks for their hard work. This project is released under the Apache License. Please take their licenses into consideration too when use this project.

Train

• STEP1: python3 train.py --snapshot imagenet --phi {0, 1, 2, 3, 4, 5, 6} --gpu 0 --random-transform --freeze-backbone --batch-size 32 --steps 1000 coco path/to/dataset to start training. The init lr is 1e-3.
• STEP2: python3 train.py --snapshot xxx.h5 --phi {0, 1, 2, 3, 4, 5, 6} --gpu 0 --random-transform --compute-val-loss --freeze-bn --batch-size 4 --steps 10000 coco path/to/dataset to start training when val mAP can not increase during STEP1. The init lr is 1e-4 and decays to 1e-5 when val mAP keeps dropping down.
• Optional arguments:
  • --snapshot SNAPSHOT Resume training from a snapshot.
  • --freeze-backbone Freeze training of backbone layers.
  • --freeze-bn Freeze training of BatchNormalization layers.
  • --weighted-bifpn Use weighted BiFPN
  • --batch-size BATCH_SIZE Size of the batches.
  • --phi {0,1,2,3,4,5,6} Hyper parameter phi
  • --gpu GPU Id of the GPU to use (as reported by nvidia-smi).
  • --num_gpus NUM_GPUS Number of GPUs to use for parallel processing.
  • --multi-gpu-force Extra flag needed to enable (experimental) multi-gpu support.
  • --epochs EPOCHS Number of epochs to train.
  • --steps STEPS Number of steps per epoch.
  • --snapshot-path SNAPSHOT_PATH Path to store snapshots of models during training
  • --tensorboard-dir TENSORBOARD_DIR Log directory for Tensorboard output
  • --no-snapshots Disable saving snapshots.
  • --no-evaluation Disable per epoch evaluation.
  • --random-transform Randomly transform image and annotations.
  • --compute-val-loss Compute validation loss during training
  • --multiprocessing Use multiprocessing in fit_generator.
  • --workers WORKERS Number of generator workers.
  • --max-queue-size MAX_QUEUE_SIZE Queue length for multiprocessing workers in fit_generator.

Test

• predict.py --data-path /pat/to/dataset --score-threshold 0.5 --model-path path/to/our/model to start testing, the result will store in result.json
• Optional arguments:
  • --data-path DATA_PATH Data for prediction
  • --target-path TARGET_PATH Target path
  • --split SPLIT Target path
  • --max-detections MAX_DETECTIONS Max detection
  • --ninedash-category-id NINEDASH_CATEGORY_ID Ninedash category ID
  • --model-path MODEL_PATH Model path of the network
  • --score-threshold SCORE_THRESHOLD Minimum score threshold
  • --phi {0,1,2,3,4,5,6} Hyper parameter phi
  • --weighted-bifpn Use weighted BiFPN
  • --batch-size BATCH_SIZE Size of the batches.
  • --num-classes NUM_CLASSES Number of classes
  • --gpu GPU Id of the GPU to use (as reported by nvidia-smi).