Gridsearch

When trying to find the right hyperparameters for your neural network, sometimes you just have to do a lot of trial and error. Currently, our Gridsearch implementation is pretty basic, but it allows you to supply ranges of input values for various metaparameters and then executes training runs in either random or sequential fashion.

Warning: this class is a bit underdeveloped. Tread with care.

Gridsearch

class pywick.gridsearch.gridsearch.GridSearch(function, grid_params, search_behavior='exhaustive', args_as_dict=True)[source]

Simple GridSearch to apply to a generic function

Parameters:
  • function – (function): function to perform grid search on
  • grid_params

    (dict): dictionary mapping variable names to lists of possible inputs aka..

    {‘input_a’:[‘dog’, ‘cat’, ‘stuff’], ‘input_b’:[3, 10, 22]}

  • search_behavior

    (string): how to perform the search. Options are: ‘exhaustive’, ‘sampled_x.x’ (where x.x is sample threshold 0.0 < 1.0)

    exhaustive - try every parameter in order they are specified in the dictionary (last key gets all its values searched first)

    sampled - sample from the dictionary of params with specified threshold. The random tries below the threshold will be executed

  • args_as_dict

    (bool): There are two ways to pass parameters into a function:

    1. Simply use each key in grid_params as a variable to pass to the function (and change those variable values according to the mapping inside grid_params)

    2. Pass a single dictionary to the function where the keys of the dictionary themselves are changed according to the grid_params

    defaults to dict

run()[source]

Runs GridSearch by iterating over options as specified :return:

Pipeline

class pywick.gridsearch.pipeline.Pipeline(ordered_func_list, func_args=None)[source]

Defines a pipeline for operating on data. Output of first function will be passed to the second and so forth.

Parameters:
  • ordered_func_list – (list): list of functions to call
  • func_args – (dict): optional dictionary of params to pass to functions in addition to last output the dictionary should be in the form of: func_name: list(params)
add_after(func, args_dict=None)[source]

Add a function to be applied at the end of the pipeline

Parameters:func – The function to apply
add_before(func, args_dict=None)[source]

Add a function to be applied before the rest in the pipeline

Parameters:func – The function to apply
call(input_)[source]

Apply the functions in current Pipeline to an input.

Parameters:input – The input to process with the Pipeline.
static identity(x)[source]

Return a copy of the input.

This is here for serialization compatibility with pickle.