Ax integrates easily with different scheduling frameworks and distributed training frameworks. In this example, Ax-driven optimization is executed in a distributed fashion using RayTune.
RayTune is a scalable framework for hyperparameter tuning that provides many state-of-the-art hyperparameter tuning algorithms and seamlessly scales from laptop to distributed cluster with fault tolerance. RayTune leverages Ray's Actor API to provide asynchronous parallel and distributed execution.
Ray 'Actors' are a simple and clean abstraction for replicating your Python classes across multiple workers and nodes. Each hyperparameter evaluation is asynchronously executed on a separate Ray actor and reports intermediate training progress back to RayTune. Upon reporting, RayTune then uses this information to performs actions such as early termination, re-prioritization, or checkpointing.
import logging
from ray import tune
from ray.tune import report
from ray.tune.suggest.ax import AxSearch
logger = logging.getLogger(tune.__name__)
logger.setLevel(
level=logging.CRITICAL
) # Reduce the number of Ray warnings that are not relevant here.
import numpy as np
import torch
from ax.plot.contour import plot_contour
from ax.plot.trace import optimization_trace_single_method
from ax.service.ax_client import AxClient
from ax.utils.notebook.plotting import init_notebook_plotting, render
from ax.utils.tutorials.cnn_utils import CNN, evaluate, load_mnist, train
init_notebook_plotting()
[INFO 02-24 16:35:08] ax.utils.notebook.plotting: Injecting Plotly library into cell. Do not overwrite or delete cell.
We specify enforce_sequential_optimization as False, because Ray runs many trials in parallel. With the sequential optimization enforcement, AxClient would expect the first few trials to be completed with data before generating more trials.
When high parallelism is not required, it is best to enforce sequential optimization, as it allows for achieving optimal results in fewer (but sequential) trials. In cases where parallelism is important, such as with distributed training using Ray, we choose to forego minimizing resource utilization and run more trials in parallel.
ax = AxClient(enforce_sequential_optimization=False)
[INFO 02-24 16:35:08] ax.service.ax_client: Starting optimization with verbose logging. To disable logging, set the `verbose_logging` argument to `False`. Note that float values in the logs are rounded to 2 decimal points.
Here we set up the search space and specify the objective; refer to the Ax API tutorials for more detail.
MINIMIZE = False # Whether we should be minimizing or maximizing the objective
ax.create_experiment(
name="mnist_experiment",
parameters=[
{"name": "lr", "type": "range", "bounds": [1e-6, 0.4], "log_scale": True},
{"name": "momentum", "type": "range", "bounds": [0.0, 1.0]},
],
objective_name="mean_accuracy",
minimize=MINIMIZE,
)
[INFO 02-24 16:35:08] ax.service.utils.instantiation: Inferred value type of ParameterType.FLOAT for parameter lr. If that is not the expected value type, you can explicity specify 'value_type' ('int', 'float', 'bool' or 'str') in parameter dict.
[INFO 02-24 16:35:08] ax.service.utils.instantiation: Inferred value type of ParameterType.FLOAT for parameter momentum. If that is not the expected value type, you can explicity specify 'value_type' ('int', 'float', 'bool' or 'str') in parameter dict.
[INFO 02-24 16:35:08] ax.modelbridge.dispatch_utils: Using Bayesian Optimization generation strategy: GenerationStrategy(name='Sobol+GPEI', steps=[Sobol for 5 trials, GPEI for subsequent trials]). Iterations after 5 will take longer to generate due to model-fitting.
ax.experiment.optimization_config.objective.minimize
False
load_mnist(data_path="~/.data") # Pre-load the dataset before the initial evaluations are executed.
1.9%
Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz to /home/runner/.data/MNIST/raw/train-images-idx3-ubyte.gz
100.1%
Extracting /home/runner/.data/MNIST/raw/train-images-idx3-ubyte.gz to /home/runner/.data/MNIST/raw
8.9%
Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz to /home/runner/.data/MNIST/raw/train-labels-idx1-ubyte.gz Extracting /home/runner/.data/MNIST/raw/train-labels-idx1-ubyte.gz to /home/runner/.data/MNIST/raw Downloading http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz to /home/runner/.data/MNIST/raw/t10k-images-idx3-ubyte.gz
180.4%/opt/hostedtoolcache/Python/3.7.10/x64/lib/python3.7/site-packages/torchvision/datasets/mnist.py:480: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:141.)
Extracting /home/runner/.data/MNIST/raw/t10k-images-idx3-ubyte.gz to /home/runner/.data/MNIST/raw Downloading http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz to /home/runner/.data/MNIST/raw/t10k-labels-idx1-ubyte.gz Extracting /home/runner/.data/MNIST/raw/t10k-labels-idx1-ubyte.gz to /home/runner/.data/MNIST/raw Processing... Done!
(<torch.utils.data.dataloader.DataLoader at 0x7f9ef72c4c90>, <torch.utils.data.dataloader.DataLoader at 0x7f9f7c637d90>, <torch.utils.data.dataloader.DataLoader at 0x7f9ef5a88dd0>)
Since we use the Ax Service API here, we evaluate the parameterizations that Ax suggests, using RayTune. The evaluation function follows its usual pattern, taking in a parameterization and outputting an objective value. For detail on evaluation functions, see Trial Evaluation.
def train_evaluate(parameterization):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_loader, valid_loader, test_loader = load_mnist(data_path="~/.data")
net = train(
net=CNN(),
train_loader=train_loader,
parameters=parameterization,
dtype=torch.float,
device=device,
)
report(
mean_accuracy=evaluate(
net=net,
data_loader=valid_loader,
dtype=torch.float,
device=device,
)
)
Execute the Ax optimization and trial evaluation in RayTune using AxSearch algorithm:
tune.run(
train_evaluate,
num_samples=30,
search_alg=AxSearch(
ax_client=ax,
max_concurrent=3,
mode="min" if MINIMIZE else "max", # Ensure that `mode` aligns with the `minimize` setting in `AxClient`
),
verbose=0, # Set this level to 1 to see status updates and to 2 to also see trial results.
# To use GPU, specify: resources_per_trial={"gpu": 1}.
)
2021-02-24 16:35:12,375 INFO services.py:1174 -- View the Ray dashboard at http://127.0.0.1:8265
--------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) <ipython-input-9-967449f35314> in <module> 7 mode="min" if MINIMIZE else "max", # Ensure that `mode` aligns with the `minimize` setting in `AxClient` 8 ), ----> 9 verbose=0, # Set this level to 1 to see status updates and to 2 to also see trial results. 10 # To use GPU, specify: resources_per_trial={"gpu": 1}. 11 ) /opt/hostedtoolcache/Python/3.7.10/x64/lib/python3.7/site-packages/ray/tune/tune.py in run(run_or_experiment, name, metric, mode, stop, time_budget_s, config, resources_per_trial, num_samples, local_dir, search_alg, scheduler, keep_checkpoints_num, checkpoint_score_attr, checkpoint_freq, checkpoint_at_end, verbose, progress_reporter, log_to_file, trial_name_creator, trial_dirname_creator, sync_config, export_formats, max_failures, fail_fast, restore, server_port, resume, queue_trials, reuse_actors, trial_executor, raise_on_failed_trial, callbacks, loggers, ray_auto_init, run_errored_only, global_checkpoint_period, with_server, upload_dir, sync_to_cloud, sync_to_driver, sync_on_checkpoint) 362 # Create syncer callbacks 363 callbacks = create_default_callbacks( --> 364 callbacks, sync_config, metric=metric, loggers=loggers) 365 366 runner = TrialRunner( /opt/hostedtoolcache/Python/3.7.10/x64/lib/python3.7/site-packages/ray/tune/utils/callback.py in create_default_callbacks(callbacks, sync_config, loggers, metric) 110 last_logger_index = len(callbacks) - 1 111 if not has_tbx_logger: --> 112 callbacks.append(TBXLoggerCallback()) 113 last_logger_index = len(callbacks) - 1 114 /opt/hostedtoolcache/Python/3.7.10/x64/lib/python3.7/site-packages/ray/tune/logger.py in __init__(self) 589 def __init__(self): 590 try: --> 591 from tensorboardX import SummaryWriter 592 self._summary_writer_cls = SummaryWriter 593 except ImportError: ModuleNotFoundError: No module named 'tensorboardX'
best_parameters, values = ax.get_best_parameters()
best_parameters
{'lr': 0.0022512732432975525, 'momentum': 3.990447334350426e-15}
means, covariances = values
means
{'mean_accuracy': 0.9661663659714574}
render(
plot_contour(
model=ax.generation_strategy.model,
param_x="lr",
param_y="momentum",
metric_name="mean_accuracy",
)
)
# `plot_single_method` expects a 2-d array of means, because it expects to average means from multiple
# optimization runs, so we wrap out best objectives array in another array.
best_objectives = np.array(
[[trial.objective_mean * 100 for trial in ax.experiment.trials.values()]]
)
best_objective_plot = optimization_trace_single_method(
y=np.maximum.accumulate(best_objectives, axis=1),
title="Model performance vs. # of iterations",
ylabel="Accuracy",
)
render(best_objective_plot)