Model Manager

Utility app for managing Keras models records and associated files

API

functional API

here are some methods used to interact with saved data

• show_all_models

• show_models_with_tag (tag : str)

• dump_models_to_df

• get_history (kmodel : KModel instance)

• plot_history_many (in_models : [] , lateral_compare : bool, ignore_none : bool) in_models : list of any combinations of str or db kmodel instance
  lateral_compare : bool, whether the plots will be grouped by models or metrics
  ignore_none : bool, whether to ignore models that don't have a history artifact
  return : (a matplotlib figure obj, data dict)

KModel

inherits from django.db.Model

additional methods :

• add_note (note : string) this method appends notes to kmodel_instance.notes along with a time stamp

• add_tags (tags : list) this method accepts a list of tags, str or Tag model instance. then adds a many-to-many relationship from self to the Tag obj

• remove_tags (tags : list) this method accepts a list of tags, str or Tag model instance. then removes the many-to-many relationship from self to the Tag obj but not the Tag itself

• add_artifact (artifact : arbitrary python obj, descriptor : str, plot : bool) descriptor is a string describing the type of data this artifact contains, does not have to be unqiue
  dumps artifact obj into pickle, create db record, returns django model instance if plot is true and a pyplot.plot object is passed as artifact, a png of the plot would be saved instead of pickle

Artifact

inherits from django.db.Model
This class is meant to be a record of pickled python objects such as model outputs, custom loss functions or Keras model factories.
Note : artifact couldn't easily modified once created, this will be fixed in the future.

additional methods : - payload @property this property returns pickled object, multiple calls of this method on the same object will not result in multiple loads.

RecorderCallback

inherents from keras.callbacks.Callback

How Does this work?

Here is a Jupyter Notebook Demo hosted on Google Colab

Saving model

Save the model and training history by passing a instance of RecorderCallback to fit or fit_generator

Saving data associated with models

call call KModel.add_artifact with arbitrary python object and it will be dumped in to a pickle file
and an Artifact object instance will be save to db and returned the path to the pickle file is saved as an attribute to the Artifact object

Interacting with database record

query the database by using Django Queryset API

###Work In Progress Stay tuned!