We have seen how to setup the ingredients for an experiment.
We have also seen how to create a recipe for a task that will act as boilerplate code for any number of experiments
Let’s go through some more of AutoTrain’s ingredients…
We have seen how to setup the ingredients for an experiment. The next step is to create the necessary files to use these ingredients in a way that training and validation can be done in one line.
All the recipes can be summarized into the following files
config.ini --> Ingredients
custom_functions.py --> User inputs that suppliment config.ini
model.py --> Base model (and optional dataloaders + learner during training)
train.py --> Training logic
validate.py --> Validation logic
main.py --> Command line and API logic
Let’s create the recepie for MNIST ingredients that we have created in the previous article.
... continue reading ...Config First ApproachAs we have discussed in the previous blog, our goal is to decouple moving parts from fixed parts so that a user can focus more time on tweaking parameters and hyperparameters of an experiment and less time setting up the code to train. The first step is going to create a framework that lets user condense, as much as possible, all important information for an experiment into a readable format.
To achieve this we have indentified and tweaked spacy-project’s awesome parser. Let’s dive straight into a sample configuration file
Below is sample file for storing information about downloading data for mnist -
[project]
version = 0.0.1
name = mnist
root = /home/me/projects/${project.name}
[project.data]
source = https://files.fast.ai/data/examples/mnist_tiny.tgz
root = ${root}/data/
What if there is a framework that uses SOTA techniques to train Deep Learning models that needs (almost) no coding?
... continue reading ...