Here I will add notebooks about how to use keras for training private datasets. This means, wild data that you own or are not a built-in function of other python library.
For the moment I have the following tutorials:
Here you can see:
- How to train a keras model using Fully Connected (FC) Layers. FC are also called multilayer perpectrons.
- How to train images from disk (HDD) by batch. In order words, only load them into memory when training in a given batch.
- How to train using manual features, in this case a histogram of all the channels of an image.
Here you can see:
- how to train an image classification model using Convolutional Neural Networks (CNN).
- how to train using images inside a zip file, by batch. In order words, only load them into memory when training in a given batch, without unziping all the content of the zip file.
- how to save and restore a keras model.
Here you can see:
- How to annotate the position of an object in a group of images for a further object localization process.
- How to use the VGG annotator tool for this purpose.
- How to export the metadata that holds the position of the objects as a JSON file.
- How to extract the data from the JSON file into a usable format, a pandas DataFrame.
You can see the process here: https://youtu.be/MRkdgOoUqFk
This is an example of a html file alone that shows how to use a keras model, transformed using the tensorflow.js tool, for prediction of a Star craft 2 unit. Here you can see:
- How to load the tensorflow.js
- How to load your model (model.js)
- How to upload an image and change its format so you can feed it into the keras trained model.
Important: In order to make it work you need to place the model into a web server, just your browser is not enough.