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.
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)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.
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.