Skip to content

Latest commit

 

History

History
81 lines (43 loc) · 4.34 KB

coffee_classifier.md

File metadata and controls

81 lines (43 loc) · 4.34 KB
Raw

Coffee Classifier Documentation

1. Installation

The following instructions apply to Ubuntu Linux 17.10.

1.1. Install required software

We install requirements with pip, the Python package management system. (So unfortunately there will be redundancy if you already have OpenCV etc. installed with Ubuntu packages on your system.)

The file requirements.txt informs pip about the software versions required for this installation. Updated versions should work, too, but may not.

The installation also contains jupyter, a Python IDE to work with annotated source code ("notebooks"). This is used in Coffee Classifier Trainer.ipynb. However, this is optional, as you can always extract the actual Python source code itself from that file and run it.

  1. Install Python 3.x and pip for Python 3.x:

     sudo apt install python3
 sudo apt install python3-pip

  2. Create ~/.bash_aliases and enter:

     alias python=python3
 alias pip=pip3

  3. Re-login to make this effective.

  4. Confirm by executing python --version and pip --version that you have Python 3.x and pip for Python 3.x.

  5. Install the requirements with pip. Tensorflow is an unmentioned requirement of Keras (seems to be a bug?). h5py is a requirement of Keras when importing a pre-trained model.

     pip install -r classifier/requirements.txt
 pip install tensorflow==1.5
 pip install h5py

  6. Try to run the classifier with one image (see below). If this results in an error, it is because the newest version of Tensorflow requires CPU features that your (older) CPU does not support – see issue #17411. In that case, try downgrading to Tensorflow 1.6, and if that still does to Tensorflow 1.5. Downgrading gradually will result in the highest version of Tensorflow that works with your CPU, resulting in the highest speed you can achieve with it because new versions tend to use more CPU capabilities.

1.2. Install the coffee beans dataset

  1. Download the full dataset of singled-out coffee bean images. (Currently here on Google Drive, but will move soon.)

  2. Fill the images into the directory directory structure that you find inside the classifier/ directory:

  • data/
    • train/
      • good/
      • bad/
    • validation/
      • good/
      • bad/

2. Using the Classifier

  1. Edit classify.py to add the path to the image file you want to classify.

  2. Run the classifier:

     python classify.py --image data/validation/good/Set04-good.55.enhanced.25.png

This will use the model definition in model.py and the weights as saved in bean_classifier.h5 to classify the input image.

3. Trainig the Classifier

To train the classifier, run any of the classifier/train*.py scripts. Each script will create a model, train it with the data in the data/ directory, and save the resulting weights to a .h5 file. You can then save that file to a different name for later to prevent it from being overwritten.

Training a classifier is not easy (it's a "fuzzy" task – results will often be quite bad without a clear reason). So we collect here our experience of what helped to improve results.

Training runs and results:

  • train_inceptionv3.weights.2018-03-29.h5: 78% final accuracy. Ca. 1650 training images, 550 validation images. 10 epochs for the upper layers, 20 epochs for the lower layers.

  • train_inceptionv3.weights.2018-04-01.h5: 83% final accuracy. Parameters: ca. 2000 training images, 845 validation images. 150x150 px image size. 20 epochs for the upper layers, 30 epochs for the lower layers. Change from last run: ca. 450 images added to the training set and 300 images added to the validation set, all taken of bad beans, with a smartphone camera in good outdoor lighting and using automatic image splitting.

  • train_inceptionv3.weights.2018-04-02.h5: 78% final accuracy. Parameters: 2507 training images, 845 validation images. 150x150 px image size. 20 epochs for the upper layers, 30 epochs for the lower layers. Change from last run: ca. 550 images added to the training set, all taken of bad beans, with a smartphone camera in good outdoor lighting and using automatic image splitting.