Distributed Grid Descent: an algorithm for hyperparameter tuning guided by Bayesian inference, designed to run on multiple processes and potentially many machines with no central point of control

Last update: Jan 01, 2022

Related tags

Algorithms DistributedGridDescent

Overview

Distributed Grid Descent

An implementation of Distributed Grid Descent: an algorithm for hyperparameter tuning guided by Bayesian inference, designed to run on multiple processes and potentially many machines with no central point of control as described in Appendix B of Working Memory Graphs [Loynd et al., 2019].

Note: This project is a work in progress. Please contact me if you like to contribute and help to develop a fully fledged python library out of it.

Usage

import numpy as np
from dgd import DistributedGridDescent

model = ... # model wrapper
data = {
    "train_data": ...
}

param_grid = {
    "learning_rate":[3e-3, 1e-3, 3e-4, 1e-4, 3e-5, 1e-5],
    "optimizer":["adam", "rmsprop"],
    "lr_annealing":[False, 0.95, 0.99],
    "batch_size":[32, 64, 128, 256, 1024],
    "num_linear_layers":[1, 2, 4, 8, 16],
    "num_neurons":[512, 256, 128, 64, 32, 16],
    "dropout":[0.0, 0.1, 0.3, 0.5],
    "l2":[0.0, 0.01, 0.1]
}

dgd = DistributedGridDescent(model, param_grid, metric=np.mean, n_jobs=-1)
dgd.run(data)

print(dgd.best_params_)
df = pd.DataFrame(dgd.results_).set_index("ID").sort_values(by=["metric"],ascending=False)

Examples and Tutorials

See sklearn_example.py, pytorch_example.py, rosenbrock_example.py and tensorflow_example.py in the examples folder for examples of basic usage of dgd.
See rosenbrock_server_example.py for an example of distributed usage.

Strong and weak scaling analysis

Algorithm

Input: Set of hyperparameters H, each having a discrete, ordered set of possible values.  
Input: Maximum number of training steps N per run.  
repeat  
    Download any new results.  
    if no results so far then
        Choose a random configuration C from the grid defined by H.
    else
        Identify the run set S with the highest metric.
        Initialize neighborhood B to contain only S.
        Expand B by adding all possible sets whose configurations differ from that of S by one step in exactly one hyperparameter setting.
        Calculate a ceiling M = Count(B) + 1.
        Weight each run set x in B M - Count(x).
        Sample a random run set S' from B according to run set weights.
        Choose configuration C from S'.
    end if
    Perform one training run of N steps using C.
    Calculate the runs Metric.
    Log the result on shared storage.
until terminated by user.

See Appendix B of Loynd et al., 2019 for details.

Distributed Grid Descent: an algorithm for hyperparameter tuning guided by Bayesian inference, designed to run on multiple processes and potentially many machines with no central point of control

Related tags

Overview

Distributed Grid Descent

Usage

Examples and Tutorials

Strong and weak scaling analysis

Algorithm

Owner

Martin

A lightweight, pure-Python mobile robot simulator designed for experiments in Artificial Intelligence (AI) and Machine Learning, especially for Jupyter Notebooks

All algorithms implemented in Python for education

Solving a card game with three search algorithms: BFS, IDS, and A*

Fedlearn algorithm toolkit for researchers

BCI datasets and algorithms

A* (with 2 heuristic functions), BFS , DFS and DFS iterativeA* (with 2 heuristic functions), BFS , DFS and DFS iterative

Implementation of core NuPIC algorithms in C++

PickMush - A mini study/project on boosting algorithm

A Python program to easily solve the n-queens problem using min-conflicts algorithm

Rover. Finding the shortest pass by Dijkstra’s shortest path algorithm

Genetic algorithms are heuristic search algorithms inspired by the process that supports the evolution of life.

causal-learn: Causal Discovery for Python

Infomap is a network clustering algorithm based on the Map equation.

Pathfinding algorithm based on A*

This is an implementation of the QuickHull algorithm in Python. I

Visualisation for sorting algorithms. Version 2.0

Implementation for Evolution of Strategies for Cooperation

The test data, code and detailed description of the AW t-SNE algorithm

Wordle-solver - A program that solves a Wordle using a simple algorithm

Ralebel is an interpreted, Haitian Creole programming language that aims to help Haitians by starting with the fundamental algorithm