mlmono

Mono repo for all machine learning work. For now this package rely on the tensorflow estimator feature.

Adding a model

Models should all go under the directory models/.

1. Write a class that inherits from MLGraph

   class CNN(MLGraph):
   def __init__(self):
       # define the variables that you want to interface with the outside
       # mostly for debugging
       # input and output are the default names for the forward model as defined in `MLGraph`
       # self.input = None
       # self.output = None

2. Defined a forward-pass function for the model

   class CNN(MLGraph):
   ...
   def add_forward_pass(self, features):
       # define the forward model
       # returns the prediction output
       # self.input and self.output must now refer to a variable node

3. Write a config file

Config files are YAML files that describes a test case. They technically can be anywhere but we suggest that you put them under configs/ and keep track of them along with the version controll system. A config file should have at least io > dataset, model and trainer fields. The corresponding modules of the model will be initialized according to the variables under those fields. Refer to the example config for more details.

How to run

Before you can run the script, you usually need to run

python setup.py develop

in order to set up the entry point. This is necessary to make the import paths correct.

The entry script is cli.py so basically you can run the mlmono.cli module in most of the cases.

To train a model

mlmono train --config configs/mnist.yml

To do prediction

mlmono predict --config configs/mnist.yml

How it works