DeepFramework is an abstract layer designed to be on top of any of your deep learning projects. This framework intends to provide a layer to separate your models logic (what is actually training, predicting...) from all the scripting needed to execute that, so that you can encapsulate and separate the core of your program from everything else. With that you can actually focus on designing your architecture and forget about all the surrounding boilerplate, and what it's more important, it will speed up your prototyping and the scaffolding of your project.
What it does:
- Ease the process of starting a new DL project
- Provides classes to avoid boilerplate code
- Helps you building a processing pipeline
- Encapsulates the computing logic (training a model, validating it...) from all the other stuff so that you can easily change the underlying layer without worrying about refactor all your program
What it does not:
- Perform any kind of DL computation
- dframe.dataset: Classes to help you load and persist your dataset. The Dataset class provides common functions such as a batch generator, get all your dataset inputs as an array...
- dframe.model: Interface to encapsulate your models logic and definition. Right now it's only a skeleton, I expect to add more functionality on that
- dframe.pipeline: Once you already have a trained model, you can use the classes provided here to create a pipeline processing system to actually perform your application's logic. It already gives you all the process management, and a really simple and useful structure
The project has been designed with two big scenarios in mind: the research and application. In both of them I have found a lack in terms of libraries supporting it and that's why I developed this framework.
For people looking to train and design a model. In order to do that, there's usually a need to construct a model, load a dataset, format it so that the model can process its data... The packages dframe.dataset and dframe.model provide classes to easily code the scripts needed to do that in a clean and structured way. I've found out that when doing research the code tends to get messy and difficult to adapt and scale (because of the different amount of experiments needed to be conducted) so I've created those packages to, not only speed up this process, but to encourage best program practices.
Training a model sometime it's just a matter of wanting the results (e.g. for a publication) but other times you want to use it in your own application (an object detection model for your killer-app for example). In this case, dframe.pipeline package it's gonna suit your needs as it provides an easy way to implement a pipeline system formed by cores. Each core is a processing unit that could have your trained model inside that takes an input, process it and outputs the result. The pipeline is in charge of connecting all the cores and of the communication with the outside world.
This framework has been uploaded to PyPI. For an easy install:
pip install DeepFramework
In your own project you will probably use two main classes: Dataset
and Model
. Their names clearly specify what are they for, so let's see how to use them in a toy example.
First you will need a dataset holding your train set, validation set or test set. The Dataset
class provides you with an easy to use interface and some helper methods so that you
don't have to implement everything by yourself. A dataset is basically a collection of Sample
. Each sample holds the input features and optionally the associated label/output.
This is how you creat such samples and dataset:
from dframe.dataset.sample import Sample
from dframe.dataset.dataset import Dataset
# Create a Dataset with a single sample that in turn has 3 input features and a single output
dataset = Dataset([Sample([1, 2, 3], 4)])
# You can add samples a posteriori
dataset.add(Sample([5, 6, 7], 8))
# Now you can get the input matrix for example
X = dataset.get_input()
y = dataset.get_output()
# A batch generator is already implemented and ready to use so you can easily feed your model with batches
# You only need to pass this function execution
dataset.batch_generator(batch_size=256) # Pass this to the model
Now it is time to create your model. You just need to create a class that extends from Model
and implement its abstract methods:
from dframe.model.model import Model
class MyModel(Model):
def __init__(self):
# Your initialization goes here. Typically you can add the model definition here. In the new version that wil come soon, a build method is added to the interface
# so that you can add there your model architecture definition and compilation
print 'Hello world!'
def train(self, dataset):
# Logic to train the model. You will typically use a DL library for that. Here you can use dataset.batch_generator to yield a batch of data
print 'Training...'
def validate(self, dataset):
print 'Validating...'
def test(self, dataset):
print 'Testing...'
def predict(self, sample):
print 'Predict output for sample'
Finally you just need to pass the dataset to the model to train it:
model = MyModel()
model.train(dataset)
- More documentation and how to use tutorial (yes, I know but... it's so boring to do). Working on some Jupyter notebooks
- Improve model package to provide a model library to manage all your models
This project is in a beta state. After developing what I thought it was the main core of the framework I am releasing it so that everyone can give it a try and, most important, provide some feedback! All of this is based on my experiences in programming deep learning projects, but as you can imagine, this is limitted, so one of the main reasons of releasing the beta is to gather opinions and suggestions from people that have worked on that longer than I am and have more experience than me. I also want to mention that I am not a guru of Python so you might see some piece of code that it is not as pythonic as it could be, or maybe there is an efficiency improvement that it can be done. I would be very grateful if you could make me notice of those so I can improve this framework to something greater! (or do it yourself and then send the merge request)