Before you begin please make sure that all required libraries listed in requirements.txt are available.
Copy or rename modules/config.template.py to modules/config.py and set up everything you need.
To predict pictures of possible micrometeorites you can use ModelPredict.ipynb. Pretrained Models for this task can be found HERE (Link to be added). Please keep in mind that these models have been trained with images that have been created using focus stacking. In addition these photos have been preprocessed.
The script lets you choose whether you want your pictures to be preprocessed or not by setting the variable preprocessing to True or False. Preprocessing the pictures means cutting out the biggest contour on the image, replacing the background with white color and creating a 1:1-format picture of the object without stretching or compressing it. To use this functionality, you’ll have to have a contour model. You can download one here.
Needed config.py:
- TRAINED_MODEL
- PREDICT_DATA
- PREDICT_RESULTS (optional)
- CONTOUR_MODEL (optional)
You can train your own models using ModelTraining.ipynb or Hyperparameter_BayesianOptimization.ipynb if you also want to do some Hyperparameteroptimization during the training process.
This Repository doesn't have any training data to share. If you only want to play around with ML you can use a data set provided by Microsoft which features pictures of cats and dogs. You can download it here.
If you have enough pictures of micrometeorites you can precprocess them with prepare_images.py if needed. Preprocessing the pictures means cutting out the biggest contour on the image, replacing the background with white color and creating a 1:1-format picture of the object without stretching or compressing it.
Needed config.py:
- TRAINING_DATA
- MODEL_PATH
- IMG_SIZE
- TRAINING_EPOCHS
- PERCENT_TEST
- CONTOUR_MODEL (optional)