Teacher-student project with SOTA teachers to improve student models results

• Teacher-student project with SOTA teachers to improve student models results
  • Generate weak labels
    • With CenterNet
    • With mmdetection (ATSS, GCNET)

Generate weak labels

With CenterNet

First of all set the proper params on the config.

Choose over which data make the inferences:

Go to:

models/CenterNet/config.py And set (at least) params:

# Directories
self._configs["data_dir"] = "/home/$USER/datasets/openimages"
self._configs["filenames_dir"] = "/tmp/oi_names.txt"
self._configs["cache_dir"] = "/home/$USER/Desktop/teacher-student/models/CenterNet/cache"
self._configs["snapshot_name"] = "CenterNet-104_480000"
self._configs["result_dir"] = "/opt/results"
self._configs["model_config"] = "/home/$USER/Desktop/teacher-student/models/CenterNet/config/CenterNet104_teacher_student.json"

• data_dir: Folder with images
• filenames_dir: .txt with image names
• cached_dir + snapshot_name: pretrained model should be in cache/nnet/$snapshot_name
• result_dir: folder to store the results
• model_config: hyper params config file

NOTE Only JSON files will be saved as results. If you want to see also labeled images change debug=False -> True in models/CenterNet/test/openimages.py

def testing(db, nnet, result_dir, debug=False):
    # This way images will be saved too
    return globals()[system_configs.sampling_function](db, nnet, result_dir, debug=True)

When all config is ready just execute get_teacher_inferences.py centernet {gpu number}

With mmdetection (ATSS, GCNET)

Execute get_teacher_inferences.py mmdetection {gpu number} and follow the steps

Recommended config:

OS: Ubuntu 18.04
CUDA: 9.0
NVCC: 9.0
GCC(G++): 5.5