CNN-Inference-Didactic

In this repositary you will find a C library written with didactic intentions using straight-forward algorithms, simple structures and lots of comments.

Channels First

This library uses a channels first convention. Anyone can read the code with this in mind, but in order to run experiments by replicating your model, its weights must be configured accordingly.

Documentation

In order to generate the documentation for this framework, open terminal/command prompt in the directory and run doxygen Doxyfile. Two versions of the documentation will be generated(one in html, one in LaTeX). In order to view the documentation website, open /html/index.html in your browser.

Preparing the Input

The input to your Convolutional Neural Network must be an RGB image. In order to prepare your input images for inference, you must run rgb_prep.py with the arguments <path/to/image> <CNN input height> <CNN input width>. This will automatically generate the input files for the C model to read.

Replicating Your Model

In order to experiment with this library, it is necessary to build a replica of your model in C. If your model is a Keras Sequential Model or was built in Keras in a sequential way, then you may choose to automate this process by interfacing with Keras directly(see below). Otherwise you will have to manually build your model using the library functions. See sample_model.c for an example model.

Extracting Trained Weights

Once you've built your model, you will need to load your trained weights into it. This is handled by the functions in h5_format.c. As the name suggests, the functions in this file were designed to work with .h5 style weights from Keras. Even if your model isn't sequential, all you need to do is to run generator.extract_weights() as shown below with your model's h5py file. This will generate the necessary files for the C model. All you need to do is to keep track of the names of your layers as shown in sample_model.c. This framework assumes all your layers have unique names.

import keras
import generator

generator.extract_weights(keras.models.load_model("path/to/your/model"))

Interfacing With Keras

If you have a Keras Sequential Model, then you may choose to directly interface with Keras and jump right into running your experiments after preparing your input(see above). In order to generate your model in C automatically, you can either run generator.py with the argument <path/to/your/model> or you may call the generator functions from another scripts like this:

import keras
import generator

model = keras.models.load_model("path/to/your/model")
generator.generate(model)
generator.extract_weights(model)

Running Your Experiment

After successfully preparing your input, replicating your model and extracting your trained weights it is time for running your inference.

Run

gcc model.c cnn_inference.c h5_format.c rgb_format.c -o infer.exe

to compile.