Enable AutoML for PyTorch

../../../_images/colab_logo_32px.pngRun in Google Colab  ../../../_images/GitHub-Mark-32px.pngView source on GitHub

In this guide we will describe how to enable automated hyper-parameter search for PyTorch using Orca AutoEstimator.

Step 0: Prepare Environment

Conda is needed to prepare the Python environment for running this example. Please refer to the install guide for more details.

conda create -n zoo python=3.7 # zoo is conda environment name, you can use any name you like.
conda activate zoo
pip install analytics-zoo[ray]
pip install torch==1.7.1 torchvision==0.8.2

Step 1: Init Orca Context

from zoo.orca import init_orca_context, stop_orca_context

if cluster_mode == "local":
    init_orca_context(cores=4, memory="2g", init_ray_on_spark=True) # run in local mode
elif cluster_mode == "k8s":
    init_orca_context(cluster_mode="k8s", num_nodes=2, cores=4, init_ray_on_spark=True) # run on K8s cluster
elif cluster_mode == "yarn":
      cluster_mode="yarn-client", cores=4, num_nodes=2, memory="2g", init_ray_on_spark=True, 
      driver_memory="10g", driver_cores=1) # run on Hadoop YARN cluster

This is the only place where you need to specify local or distributed mode. View Orca Context for more details.

Note: You should export HADOOP_CONF_DIR=/path/to/hadoop/conf/dir when running on Hadoop YARN cluster. View Hadoop User Guide for more details.

Step 2: Define the Model

You may define your model, loss and optimizer in the same way as in any standard PyTorch program.

import torch
import torch.nn as nn
import torch.nn.functional as F

class LeNet(nn.Module):
    def __init__(self, fc1_hidden_size=500):
        super(LeNet, self).__init__()
        self.conv1 = nn.Conv2d(1, 20, 5, 1)
        self.conv2 = nn.Conv2d(20, 50, 5, 1)
        self.fc1 = nn.Linear(4*4*50, fc1_hidden_size)
        self.fc2 = nn.Linear(fc1_hidden_size, 10)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = F.max_pool2d(x, 2, 2)
        x = F.relu(self.conv2(x))
        x = F.max_pool2d(x, 2, 2)
        x = x.view(-1, 4*4*50)
        x = F.relu(self.fc1(x))
        x = self.fc2(x)
        return F.log_softmax(x, dim=1)

criterion = nn.NLLLoss()

After defining your model, you need to define a Model Creator Function that returns an instance of your model, and a Optimizer Creator Function that returns a PyTorch optimizer. Note that both the Model Creator Function and the Optimizer Creator Function should take config as input and get the hyper-parameter values from config.

def model_creator(config):
    model = LeNet(fc1_hidden_size=config["fc1_hidden_size"])
    return model

def optim_creator(model, config):
    return torch.optim.Adam(model.parameters(), lr=config["lr"])

Step 3: Define Dataset

You can define the train and validation datasets using Data Creator Function that takes config as input and returns a PyTorch DataLoader.

import torch
from torchvision import datasets, transforms

dir = './dataset'
test_batch_size = 640

def train_loader_creator(config):
    train_loader = torch.utils.data.DataLoader(
        datasets.MNIST(dir, train=True, download=True,
                           transforms.Normalize((0.1307,), (0.3081,))
        batch_size=config["batch_size"], shuffle=True)
    return train_loader

def test_loader_creator(config):
    test_loader = torch.utils.data.DataLoader(
        datasets.MNIST(dir, train=False, download=True,
                           transforms.Normalize((0.1307,), (0.3081,))
        batch_size=test_batch_size, shuffle=False)
    return test_loader

Step 4: Define Search Space

You should define a dictionary as your hyper-parameter search space.

The keys are hyper-parameter names which should be the same with those in your creators, and you can specify how you want to sample each hyper-parameter in the values of the search space. See automl.hp for more details.

from zoo.orca.automl import hp

search_space = {
    "fc1_hidden_size": hp.choice([500, 600]),
    "lr": hp.choice([0.001, 0.003]),
    "batch_size": hp.choice([160, 320, 640]),

Step 5: Automatically Fit and Search with Orca AutoEstimator

First, create an AutoEstimator. You can refer to AutoEstimator API doc for more details.

from zoo.orca.automl.auto_estimator import AutoEstimator

auto_est = AutoEstimator.from_torch(model_creator=model_creator,
                                    resources_per_trial={"cpu": 2},

Next, use the AutoEstimator to fit and search for the best hyper-parameter set.


Finally, you can get the best learned model and the best hyper-parameters.

best_model = auto_est.get_best_model()
best_config = auto_est.get_best_config()

Note: You should call stop_orca_context() when your application finishes.