mobilenet-v2-custom-dataset

Using Keras MobileNet-v2 model with your custom images dataset

The Keras implementation of MobileNet-v2 (from Keras-Application package) uses by default famous datasets such as imagenet, cifar in a encoded format. But what if we want to use our own custom dataset?

The problem is that if we load all images in a single numpy array, the memory will quickly overload, that's why in this repository we use keras ImageDataGenerator class to generate batches during the runtime. The advantage of using ImageDataGenerator to generate batches instream of making a hand-made loop over our dataset is that it is directly supported by keras models and we just have to call fit_generator method to train on the batches. Moreover, we can easily activate the data augmentation option.

Our custom dataset need to have the following structure: for every class create a folder containing .jpg sample images:

dataset_folder
    class1
        image1.jpg
        image2.jpg
    class2
        image1.jpg
        image2.jpg
        image3.jpg

How to use?

1. Configure the parameters in config.json

2. Train the model using python train.py

3. Evaluate the model on test dataset using: python test.py