TensorFlow 1.8.0.
In order to run the project from the command line, first make sure the PYTHONPATH system variable contains the path to the root directory. Ensure that the utils/constants.py contains the proper paths.
The project was developed using TensorFlow 1.8.0. If you use a newer version, you may receive deprecation warnings and some scripts may not work properly. In particular, the utils/freeze_graph.py script may produce errors since the format of the checkpoint files may differ in newer TensorFlow versions. This script is available in any TensorFlow version and so using the script provided in your TensorFlow distribution is recommended.
The latest version can be found here: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/freeze_graph.py
An implementation of the same network that is adapted to the latest TensorFlow version can be found on Kaggle in a Python Notebook: https://www.kaggle.com/mitch9090/fruit-dataset-image-classification-network
Use the utils/build_image_data.py script to generate the tfrecords files which are needed for the train and test scripts.
In the utils/build_image_data.py file you can modify the path to the dataset, the location where the tfrecords are saved and the path to the labels file. The labels file contains the labels(names of folders) from the dataset that you want to include in the tfrecord files, each label must be on a newline.
Alternatively, the utils/build_image_data.py can be run from the command line like this:
python build_image_data.py [flags]
The available options are:
--train_directory=path to the folder containing the train images
--validation_directory=path to the folder containing the validation images
--output_directory=path to where to output the tfrecord files
--labels_file=path to the label file
Run the network/fruit_train_net.py to train the network - currently the network runs for 75000 iterations and saves the network state every 1000 steps. By default, the network/fruit_train_net.py and the network/fruit_test_net.py scripts expect the tfrecords files(obtained using the utils/build_image_data.py script) to be in a folder called "data", at the same level as the network folder. This can be changed in the utils/constants.py by changing the "data_dir" variable. This script will create the fruit_models/ folder where the network parameters and the graph meta data are saved.
After training the network run the network/fruit_test_net.py to test the obtained model.
The fruit_detection/detect_fruits.py script contains a basic example on how to load the trained model, how to read an image and how to feed it to the network and how to get the output and interpret it.
The script has only a flag:
--image_path=path to the image to be read and classified
The network_structure/fruit_network.py file contains the network definition as well as parameters used for building the network:
batch_size #amount of images used in each iteration
input_size #amount of pixels in the first layer of the network
num_classes #number of classes on the output layer -> this must match the number of classes that is used to generate the tf record from the utils/build_image_data.py script
dropout #probability to keep the weights resulted after a train iteration
initial_learning_rate #start value for the learning rate
final_learning_rate #lower bound for the learning rate
The network_structure/utils.py file contains a few helper methods. The most important ones are:
| Method name | Effect |
|---|---|
| augment_image | performs data augmentation on images by altering the hue and saturation and by randomly flipping the image vertically and orizontally; afterwards this method calls the build_hsv_grayscale_image method |
| build_hsv_grayscale_image | converts the image to HSV and concatenates a grayscale representation ofthe image |
The utils/freeze_graph.py file creates a single file with the operations graph definition and weights values. This file is provided with every version of TensorFlow.
For running the script, these flags are mandatory:
--input_graph=path to the pbtxt file
--input_checkpoint=path to the ckpt file
--output_graph=name of the output file
--output_node_names=name of the last layer of the network (found in the network_structure/fruit_network.py file, in the convnet method, in this case the name of the last layer is ”out/out”)