4-add-new-models

Setup

Make sure have the environemnt setup (../README.md).

Tutorial

1. Wrapping the Iris classifier into Kipoi

Go through the adding_new_models.ipynb notebook.

This notebook provides an end-to-end example of adding a new model to the Kipoi format and explains how to share it to https://github.com/kipoi/models/.

2. Adding a custom model with a pre-defined dataloader from Kipoiseq

Next, try to write a model in models/gc_bias which will compute the fraction of GC nucleotides in the sequence. The model should expect a simple string as input and should return a float.

You can use kipoiseq.dataloaders.StringSeqIntervalDl for the dataloader.

Once you are done implementing the model in models/gc_bias, see if you can run the tests:

source activate kipoi-shared__envs__kipoi-py3-keras2
kipoi test models/gc_bias

3. Adding multiple Keras models in a single batch

I used the following colab notebook to train three models.

I created a Kipoi model out of a single model (model1) and saved it to: models/keras.

Read models/keras/model.yaml and run kipoi test models/keras to see if it works.

Your task is now to use model templating to simulteneusly add all there models to models/multiple/ by writing the following two files:

model-template.yaml
model.tsv

To get more information about these files, read the 'Contributing multiple very similar models with a template' paragraph in Kipoi 0.6 release notes.

Note: use models/keras/model.yaml as a starting point for models/multiple/model-template.yaml.

You can find all the links to the model under models/multiple/README.md.

Test if you can successfully run the models using:

kipoi test models/multiple/model3

4. Testing if predictions match the expected ones

Generate the expected predictions

kipoi test models/keras -o models/keras/expect.h5

Test if they match:

kipoi test models/keras -e models/keras/expect.h5

Note: you can add the following entry to model.yaml in order to automatically check predictions for every kipoi test <model>:

test:
  expect:
    url: <url to expect.h5>
    md5: <md5 hash of expect.h5>
  precision_decimal: 6

Links

• https://github.com/kipoi/models/ - explore the how existing models are implemented
• https://github.com/kipoi/kipoi/tree/master/example/models - example models used for testing the Kipoi API
• http://kipoi.org/docs/contributing/01_Getting_started/ - getting started to contributing models
• Kipoi 0.6 release notes - blog post explaining model templaes, Kipoiseq, hosting remote files on Zenodo and prediction testing

Next step: 5-appendix