Kuzushiji-recognition

The Kaggle page of the competition can be found here: https://www.kaggle.com/c/kuzushiji-recognition.

Working environment

This is the recommended environment, but it may work also with other versions.

• Python 3.6.8 (≥ 3.6 required)
• Tensorflow 1.14.0

Project setup

1. Clone/download the repository
2. Install the required packages: pip install -r requirements.txt
3. Modify configuration parameters as described in next section.

Test configuration

To evaluate the models follow this steps:

1. Download the Kaggle dataset from this page (about 7 GB), create a folder with the following structure:

       repo_folder/
       ├── datasets/
       │   └── kaggle/
       │       ├── testing/
       │       │   └── images
       │       └── training/
       │           └── images
   

   Decompress test_images.zip and place all the images in testing/images/. Place train images in training/images.

Insert the following files into the datasets/kaggle folder:

• image_labels_map.csv (i.e. train.csv)
• sample_submission.csv
• unicode_translation.csv

2. Adjust configuration in networks/configuration/params_model_CenterNet.json. Use the evaluate parameters to print metrics. Detector and classifier have separate configurations. To print metrics set evaluate to true, leaving the other parameters to false. You can find a list of prepared JSON in here. You may adjust the batch size via the batch_size_predict parameter.

3. Create a folder in networks/experiments with the following structure:

        experiments/
        ├── run_name
        ├── run_name_2/
        │   └── weights
        ├── run_name_3/
        │   └── weights
   

4. Download the desired weights from GDrive. Place the detector weights inside run_name_2/weigths/ and the classifier weights in run_name_3/weights/

5. Weights file are named as weights.11-0.13.hdf5. Here 11 is the epoch number.

   • Set this number in initial_epoch parameter inside the JSON file for both models
   • Set the parameter run_id to the chosen 'run_name'
   • Set restore_weights to true for both models.

6. Run program with python3 __centernet__.py

Train configuration

Use the same configuration, setting train parameter to true. restore_weights can be set to true, to continue training from previous weights. In this case training will restart from specified initial epoch. Otherwise set it to false and set initial_epoch to 0, to restart. To make new experiments without overwriting existing weights, just change the run_id parameter: a new experiment folder will be created.