How to use Tune with PyTorch#

In this walkthrough, we will show you how to integrate Tune into your PyTorch training workflow. We will follow this tutorial from the PyTorch documentation for training a CIFAR10 image classifier.

../../_images/pytorch_logo.png

Hyperparameter tuning can make the difference between an average model and a highly accurate one. Often simple things like choosing a different learning rate or changing a network layer size can have a dramatic impact on your model performance. Fortunately, Tune makes exploring these optimal parameter combinations easy - and works nicely together with PyTorch.

As you will see, we only need to add some slight modifications. In particular, we need to

  1. wrap data loading and training in functions,

  2. make some network parameters configurable,

  3. add checkpointing (optional),

  4. and define the search space for the model tuning

Note

To run this example, you will need to install the following:

$ pip install ray torch torchvision

Setup / Imports#

Let’s start with the imports:

import numpy as np
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from filelock import FileLock
from torch.utils.data import random_split
import torchvision
import torchvision.transforms as transforms
import ray
from ray import train, tune
from ray.train import Checkpoint
from ray.tune.schedulers import ASHAScheduler

Most of the imports are needed for building the PyTorch model. Only the last three imports are for Ray Tune.

Data loaders#

We wrap the data loaders in their own function and pass a global data directory. This way we can share a data directory between different trials.

def load_data(data_dir="./data"):
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
    ])

    # We add FileLock here because multiple workers will want to
    # download data, and this may cause overwrites since
    # DataLoader is not threadsafe.
    with FileLock(os.path.expanduser("~/.data.lock")):
        trainset = torchvision.datasets.CIFAR10(
            root=data_dir, train=True, download=True, transform=transform)

        testset = torchvision.datasets.CIFAR10(
            root=data_dir, train=False, download=True, transform=transform)

    return trainset, testset

Configurable neural network#

We can only tune those parameters that are configurable. In this example, we can specify the layer sizes of the fully connected layers:

class Net(nn.Module):
    def __init__(self, l1=120, l2=84):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, l1)
        self.fc2 = nn.Linear(l1, l2)
        self.fc3 = nn.Linear(l2, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

The train function#

Now it gets interesting, because we introduce some changes to the example from the PyTorch documentation.

The full code example looks like this:

def train_cifar(config):
    net = Net(config["l1"], config["l2"])

    device = "cpu"
    if torch.cuda.is_available():
        device = "cuda:0"
        if torch.cuda.device_count() > 1:
            net = nn.DataParallel(net)
    net.to(device)

    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(net.parameters(), lr=config["lr"], momentum=0.9)

    # To restore a checkpoint, use `train.get_checkpoint()`.
    loaded_checkpoint = train.get_checkpoint()
    if loaded_checkpoint:
        with loaded_checkpoint.as_directory() as loaded_checkpoint_dir:
           model_state, optimizer_state = torch.load(os.path.join(loaded_checkpoint_dir, "checkpoint.pt"))
        net.load_state_dict(model_state)
        optimizer.load_state_dict(optimizer_state)

    data_dir = os.path.abspath("./data")
    trainset, testset = load_data(data_dir)

    test_abs = int(len(trainset) * 0.8)
    train_subset, val_subset = random_split(
        trainset, [test_abs, len(trainset) - test_abs])

    trainloader = torch.utils.data.DataLoader(
        train_subset,
        batch_size=int(config["batch_size"]),
        shuffle=True,
        num_workers=8)
    valloader = torch.utils.data.DataLoader(
        val_subset,
        batch_size=int(config["batch_size"]),
        shuffle=True,
        num_workers=8)

    for epoch in range(10):  # loop over the dataset multiple times
        running_loss = 0.0
        epoch_steps = 0
        for i, data in enumerate(trainloader, 0):
            # get the inputs; data is a list of [inputs, labels]
            inputs, labels = data
            inputs, labels = inputs.to(device), labels.to(device)

            # zero the parameter gradients
            optimizer.zero_grad()

            # forward + backward + optimize
            outputs = net(inputs)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()

            # print statistics
            running_loss += loss.item()
            epoch_steps += 1
            if i % 2000 == 1999:  # print every 2000 mini-batches
                print("[%d, %5d] loss: %.3f" % (epoch + 1, i + 1,
                                                running_loss / epoch_steps))
                running_loss = 0.0

        # Validation loss
        val_loss = 0.0
        val_steps = 0
        total = 0
        correct = 0
        for i, data in enumerate(valloader, 0):
            with torch.no_grad():
                inputs, labels = data
                inputs, labels = inputs.to(device), labels.to(device)

                outputs = net(inputs)
                _, predicted = torch.max(outputs.data, 1)
                total += labels.size(0)
                correct += (predicted == labels).sum().item()

                loss = criterion(outputs, labels)
                val_loss += loss.cpu().numpy()
                val_steps += 1

        # Here we save a checkpoint. It is automatically registered with
        # Ray Tune and can be accessed through `train.get_checkpoint()`
        # API in future iterations.
        os.makedirs("my_model", exist_ok=True)
        torch.save(
            (net.state_dict(), optimizer.state_dict()), "my_model/checkpoint.pt")
        checkpoint = Checkpoint.from_directory("my_model")
        train.report({"loss": (val_loss / val_steps), "accuracy": correct / total}, checkpoint=checkpoint)
    print("Finished Training")

As you can see, most of the code is adapted directly from the example.

Test set accuracy#

Commonly the performance of a machine learning model is tested on a hold-out test set with data that has not been used for training the model. We also wrap this in a function:

def test_best_model(best_result):
    best_trained_model = Net(best_result.config["l1"], best_result.config["l2"])
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    best_trained_model.to(device)

    checkpoint_path = os.path.join(best_result.checkpoint.to_directory(), "checkpoint.pt")

    model_state, optimizer_state = torch.load(checkpoint_path)
    best_trained_model.load_state_dict(model_state)

    trainset, testset = load_data()

    testloader = torch.utils.data.DataLoader(
        testset, batch_size=4, shuffle=False, num_workers=2)

    correct = 0
    total = 0
    with torch.no_grad():
        for data in testloader:
            images, labels = data
            images, labels = images.to(device), labels.to(device)
            outputs = best_trained_model(images)
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()


    print("Best trial test set accuracy: {}".format(correct / total))

As you can see, the function also expects a device parameter, so we can do the test set validation on a GPU.

Configuring the search space#

Lastly, we need to define Tune’s search space. Here is an example:

config = {
    "l1": tune.sample_from(lambda _: 2**np.random.randint(2, 9)),
    "l2": tune.sample_from(lambda _: 2**np.random.randint(2, 9)),
    "lr": tune.loguniform(1e-4, 1e-1),
    "batch_size": tune.choice([2, 4, 8, 16]),
}

The tune.sample_from() function makes it possible to define your own sample methods to obtain hyperparameters. In this example, the l1 and l2 parameters should be powers of 2 between 4 and 256, so either 4, 8, 16, 32, 64, 128, or 256. The lr (learning rate) should be uniformly sampled between 0.0001 and 0.1. Lastly, the batch size is a choice between 2, 4, 8, and 16.

At each trial, Tune will now randomly sample a combination of parameters from these search spaces. It will then train a number of models in parallel and find the best performing one among these. We also use the ASHAScheduler which will terminate bad performing trials early.

You can specify the number of CPUs, which are then available e.g. to increase the num_workers of the PyTorch DataLoader instances. The selected number of GPUs are made visible to PyTorch in each trial. Trials do not have access to GPUs that haven’t been requested for them - so you don’t have to care about two trials using the same set of resources.

Here we can also specify fractional GPUs, so something like gpus_per_trial=0.5 is completely valid. The trials will then share GPUs among each other. You just have to make sure that the models still fit in the GPU memory.

After training the models, we will find the best performing one and load the trained network from the checkpoint file. We then obtain the test set accuracy and report everything by printing.

The full main function looks like this:

def main(num_samples=10, max_num_epochs=10, gpus_per_trial=2):
    config = {
        "l1": tune.sample_from(lambda _: 2 ** np.random.randint(2, 9)),
        "l2": tune.sample_from(lambda _: 2 ** np.random.randint(2, 9)),
        "lr": tune.loguniform(1e-4, 1e-1),
        "batch_size": tune.choice([2, 4, 8, 16])
    }
    scheduler = ASHAScheduler(
        max_t=max_num_epochs,
        grace_period=1,
        reduction_factor=2)
    
    tuner = tune.Tuner(
        tune.with_resources(
            tune.with_parameters(train_cifar),
            resources={"cpu": 2, "gpu": gpus_per_trial}
        ),
        tune_config=tune.TuneConfig(
            metric="loss",
            mode="min",
            scheduler=scheduler,
            num_samples=num_samples,
        ),
        param_space=config,
    )
    results = tuner.fit()
    
    best_result = results.get_best_result("loss", "min")

    print("Best trial config: {}".format(best_result.config))
    print("Best trial final validation loss: {}".format(
        best_result.metrics["loss"]))
    print("Best trial final validation accuracy: {}".format(
        best_result.metrics["accuracy"]))

    test_best_model(best_result)

main(num_samples=2, max_num_epochs=2, gpus_per_trial=0)

2022-07-22 16:38:53,384	INFO services.py:1483 -- View the Ray dashboard at http://127.0.0.1:8273
2022-07-22 16:38:56,785	WARNING function_trainable.py:619 --
== Status ==
Current time: 2022-07-22 16:40:13 (running for 00:01:16.43)
Memory usage on this node: 10.7/16.0 GiB
Using AsyncHyperBand: num_stopped=2 Bracket: Iter 2.000: -1.421571186053753 | Iter 1.000: -1.7652838359832763
Resources requested: 0/16 CPUs, 0/0 GPUs, 0.0/5.63 GiB heap, 0.0/2.0 GiB objects
Current best trial: 66098_00000 with loss=1.421571186053753 and parameters={'l1': 128, 'l2': 128, 'lr': 0.00046907397024184945, 'batch_size': 4}
Result logdir: /Users/kai/ray_results/train_cifar_2022-07-22_16-38-50
Number of trials: 2/2 (2 TERMINATED)
Trial name status loc batch_size l1 l2 lr iter total time (s) loss accuracy
train_cifar_66098_00000TERMINATED127.0.0.1:53065 4 128 1280.000469074 2 72.61761.42157 0.4877
train_cifar_66098_00001TERMINATED127.0.0.1:53078 4 128 640.00993903 1 64.97211.90462 0.2915
 


2022-07-22 16:38:57,794	INFO plugin_schema_manager.py:52 -- Loading the default runtime env schemas: ['/Users/kai/coding/ray/python/ray/_private/runtime_env/../../runtime_env/schemas/working_dir_schema.json', '/Users/kai/coding/ray/python/ray/_private/runtime_env/../../runtime_env/schemas/pip_schema.json'].

(train_cifar pid=53065) Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to /Users/kai/ray_results/train_cifar_2022-07-22_16-38-50/train_cifar_66098_00000_0_batch_size=4,l1=128,l2=128,lr=0.0005_2022-07-22_16-38-57/data/cifar-10-python.tar.gz

  0%|          | 0/170498071 [00:00<?, ?it/s]
  0%|          | 33792/170498071 [00:00<14:09, 200766.21it/s]
  0%|          | 197632/170498071 [00:00<04:21, 650251.40it/s]
  0%|          | 492544/170498071 [00:00<02:02, 1393215.11it/s]
  1%|          | 1082368/170498071 [00:00<01:00, 2821583.21it/s]
  1%|          | 1950720/170498071 [00:00<00:36, 4640856.44it/s]
  2%|▏         | 2835456/170498071 [00:00<00:28, 5924997.25it/s]
  2%|▏         | 3965952/170498071 [00:00<00:22, 7537390.37it/s]
  3%|▎         | 5063680/170498071 [00:00<00:19, 8499565.01it/s]
  4%|▎         | 6128640/170498071 [00:01<00:17, 9134765.49it/s]
  4%|▍         | 7406592/170498071 [00:01<00:15, 10222961.60it/s]
  5%|▌         | 8553472/170498071 [00:01<00:15, 10510446.68it/s]
  6%|▌         | 9700352/170498071 [00:01<00:14, 10735559.18it/s]
  6%|▋         | 10863616/170498071 [00:01<00:14, 11000864.50it/s]
  7%|▋         | 11971584/170498071 [00:01<00:14, 10807721.84it/s]
  8%|▊         | 13059072/170498071 [00:01<00:14, 10666283.43it/s]
  8%|▊         | 14130176/170498071 [00:01<00:14, 10574609.87it/s]
  9%|▉         | 15303680/170498071 [00:01<00:14, 10892197.45it/s]
 10%|▉         | 16396288/170498071 [00:01<00:14, 10675729.31it/s]
 10%|█         | 17548288/170498071 [00:02<00:14, 10904960.72it/s]
 11%|█         | 18641920/170498071 [00:02<00:15, 9863782.34it/s] 
 12%|█▏        | 20268032/170498071 [00:02<00:12, 11570721.42it/s]
 13%|█▎        | 21452800/170498071 [00:02<00:12, 11541443.91it/s]
 13%|█▎        | 22742016/170498071 [00:02<00:12, 11907361.68it/s]
 14%|█▍        | 23948288/170498071 [00:02<00:12, 11767051.50it/s]
 15%|█▍        | 25136128/170498071 [00:02<00:12, 11573913.54it/s]
 15%|█▌        | 26362880/170498071 [00:02<00:12, 11761937.08it/s]
 16%|█▌        | 27545600/170498071 [00:02<00:12, 11044330.05it/s]
 17%|█▋        | 28662784/170498071 [00:03<00:12, 11026608.99it/s]
 17%|█▋        | 29819904/170498071 [00:03<00:12, 11138417.94it/s]
 18%|█▊        | 30940160/170498071 [00:03<00:12, 11010962.50it/s]
 19%|█▉        | 32046080/170498071 [00:03<00:13, 10129993.91it/s]
 19%|█▉        | 33074176/170498071 [00:03<00:13, 9943367.80it/s] 
 20%|██        | 34259968/170498071 [00:03<00:13, 10448733.26it/s]
 21%|██        | 35521536/170498071 [00:03<00:12, 11062770.09it/s]
 22%|██▏       | 36799488/170498071 [00:03<00:11, 11524350.69it/s]
 22%|██▏       | 37961728/170498071 [00:03<00:11, 11438058.60it/s]
 23%|██▎       | 39112704/170498071 [00:04<00:11, 11458573.52it/s]
 24%|██▎       | 40263680/170498071 [00:04<00:11, 11424542.73it/s]
 24%|██▍       | 41409536/170498071 [00:04<00:11, 11346691.19it/s]
 25%|██▌       | 42697728/170498071 [00:04<00:10, 11786332.70it/s]
 26%|██▌       | 43879424/170498071 [00:04<00:11, 11446463.24it/s]
 26%|██▋       | 45028352/170498071 [00:04<00:11, 10783870.46it/s]
 27%|██▋       | 46115840/170498071 [00:04<00:11, 10678464.51it/s]
 28%|██▊       | 47190016/170498071 [00:04<00:11, 10442301.65it/s]
 28%|██▊       | 48350208/170498071 [00:04<00:11, 10759276.29it/s]
 29%|██▉       | 49447936/170498071 [00:04<00:11, 10792782.74it/s]
 30%|██▉       | 50731008/170498071 [00:05<00:10, 11385649.79it/s]
 31%|███       | 52003840/170498071 [00:05<00:10, 11752264.60it/s]
 31%|███       | 53183488/170498071 [00:05<00:10, 11665484.87it/s]
 32%|███▏      | 54360064/170498071 [00:05<00:09, 11694832.15it/s]
 33%|███▎      | 55531520/170498071 [00:05<00:10, 11476978.86it/s]
 33%|███▎      | 56681472/170498071 [00:05<00:10, 11202902.06it/s]
 35%|███▍      | 59196416/170498071 [00:05<00:09, 11613442.08it/s]
 35%|███▌      | 60392448/170498071 [00:05<00:09, 11697365.01it/s]
 36%|███▌      | 61568000/170498071 [00:05<00:09, 11682261.49it/s]
 37%|███▋      | 62740480/170498071 [00:06<00:09, 11452394.76it/s]
 37%|███▋      | 63889408/170498071 [00:06<00:09, 11380145.84it/s]
 38%|███▊      | 65030144/170498071 [00:06<00:09, 10851178.13it/s]
 39%|███▉      | 66126848/170498071 [00:06<00:09, 10801561.44it/s]
 39%|███▉      | 67224576/170498071 [00:06<00:09, 10829254.00it/s]
 40%|████      | 68322304/170498071 [00:06<00:09, 10860535.53it/s]
 41%|████      | 69410816/170498071 [00:06<00:09, 10588112.87it/s]
 41%|████▏     | 70472704/170498071 [00:06<00:09, 10246538.71it/s]
 42%|████▏     | 71500800/170498071 [00:06<00:09, 9998803.68it/s] 
 43%|████▎     | 72631296/170498071 [00:07<00:09, 10246250.37it/s]
 43%|████▎     | 73746432/170498071 [00:07<00:09, 10505166.41it/s]
 44%|████▍     | 74843136/170498071 [00:07<00:09, 10607834.22it/s]
 45%|████▍     | 75907072/170498071 [00:07<00:09, 10449028.08it/s]
 45%|████▌     | 76954624/170498071 [00:07<00:08, 10432352.97it/s]
 46%|████▌     | 78021632/170498071 [00:07<00:08, 10390722.73it/s]
 46%|████▋     | 79119360/170498071 [00:07<00:08, 10527472.73it/s]
 47%|████▋     | 80173056/170498071 [00:07<00:08, 10185476.35it/s]
 48%|████▊     | 81265664/170498071 [00:07<00:08, 10271324.45it/s]
 48%|████▊     | 82396160/170498071 [00:07<00:08, 10568711.71it/s]
 49%|████▉     | 83456000/170498071 [00:08<00:08, 10544461.91it/s]
 50%|████▉     | 84624384/170498071 [00:08<00:07, 10847079.54it/s]
 50%|█████     | 85754880/170498071 [00:08<00:07, 10930136.27it/s]
 51%|█████     | 86934528/170498071 [00:08<00:07, 11111543.18it/s]
 52%|█████▏    | 88179712/170498071 [00:08<00:07, 11438251.85it/s]
 52%|█████▏    | 89375744/170498071 [00:08<00:07, 11528186.37it/s]
 53%|█████▎    | 90620928/170498071 [00:08<00:06, 11741163.72it/s]
 54%|█████▍    | 91833344/170498071 [00:08<00:06, 11844882.32it/s]
 55%|█████▍    | 93019136/170498071 [00:08<00:07, 10859729.47it/s]
 55%|█████▌    | 94120960/170498071 [00:09<00:07, 10842087.69it/s]
 56%|█████▌    | 95216640/170498071 [00:09<00:07, 10396612.64it/s]
 56%|█████▋    | 96266240/170498071 [00:09<00:07, 10348003.47it/s]
 57%|█████▋    | 97354752/170498071 [00:09<00:06, 10497379.20it/s]
 58%|█████▊    | 98550784/170498071 [00:09<00:06, 10859176.06it/s]
 59%|█████▊    | 99746816/170498071 [00:09<00:06, 11154133.97it/s]
 59%|█████▉    | 101041152/170498071 [00:09<00:05, 11656717.30it/s]
 61%|██████    | 103482368/170498071 [00:09<00:05, 11894491.00it/s]
 61%|██████▏   | 104776704/170498071 [00:09<00:05, 12198050.59it/s]
 62%|██████▏   | 106021888/170498071 [00:10<00:05, 12267825.95it/s]
 63%|██████▎   | 107250688/170498071 [00:10<00:05, 12168871.59it/s]
 64%|██████▎   | 108469248/170498071 [00:10<00:05, 11595891.87it/s]
 64%|██████▍   | 109635584/170498071 [00:10<00:05, 11559630.86it/s]
 65%|██████▍   | 110795776/170498071 [00:10<00:05, 11349017.87it/s]
 66%|██████▌   | 111943680/170498071 [00:10<00:05, 11386490.70it/s]
 66%|██████▋   | 113085440/170498071 [00:10<00:05, 11297820.57it/s]
 67%|██████▋   | 114216960/170498071 [00:10<00:05, 11067298.70it/s]
 68%|██████▊   | 115325952/170498071 [00:10<00:05, 10487747.55it/s]
 68%|██████▊   | 116442112/170498071 [00:11<00:05, 10659499.65it/s]
 69%|██████▉   | 117720064/170498071 [00:11<00:04, 11250043.47it/s]
 70%|██████▉   | 118981632/170498071 [00:11<00:04, 11635244.90it/s]
 70%|███████   | 120151040/170498071 [00:11<00:04, 11648387.01it/s]
 71%|███████   | 121357312/170498071 [00:11<00:04, 11768670.42it/s]
 72%|███████▏  | 122586112/170498071 [00:11<00:04, 11905710.02it/s]
 73%|███████▎  | 123779072/170498071 [00:11<00:03, 11731120.04it/s]
 73%|███████▎  | 124954624/170498071 [00:11<00:03, 11738276.15it/s]
 74%|███████▍  | 126130176/170498071 [00:11<00:03, 11630197.21it/s]
 75%|███████▍  | 127370240/170498071 [00:11<00:03, 11827478.91it/s]
 75%|███████▌  | 128631808/170498071 [00:12<00:03, 12036486.08it/s]
 76%|███████▌  | 129860608/170498071 [00:12<00:03, 12070353.14it/s]
 77%|███████▋  | 131068928/170498071 [00:12<00:03, 11930703.88it/s]
 78%|███████▊  | 132262912/170498071 [00:12<00:03, 11640168.06it/s]
 78%|███████▊  | 133429248/170498071 [00:12<00:03, 11474541.42it/s]
 79%|███████▉  | 134611968/170498071 [00:12<00:03, 11412959.43it/s]
 80%|████████  | 137004032/170498071 [00:12<00:02, 11687905.40it/s]
 81%|████████  | 138174464/170498071 [00:13<00:05, 5764496.82it/s] 
 83%|████████▎ | 141067264/170498071 [00:13<00:03, 9739825.72it/s]
 84%|████████▎ | 142558208/170498071 [00:13<00:02, 9526524.50it/s]
 84%|████████▍ | 143872000/170498071 [00:13<00:03, 6741235.46it/s]
 86%|████████▌ | 146703360/170498071 [00:13<00:02, 10136683.18it/s]

 87%|████████▋ | 148267008/170498071 [00:14<00:02, 10106749.67it/s]
 88%|████████▊ | 149662720/170498071 [00:14<00:02, 10207494.38it/s]
 89%|████████▊ | 150955008/170498071 [00:14<00:01, 10359984.46it/s]
 89%|████████▉ | 152185856/170498071 [00:14<00:01, 10695074.47it/s]
 90%|████████▉ | 153402368/170498071 [00:14<00:01, 10797069.40it/s]
 91%|█████████ | 154587136/170498071 [00:14<00:01, 10637510.80it/s]
 91%|█████████▏| 155722752/170498071 [00:14<00:01, 10801079.07it/s]
 92%|█████████▏| 156895232/170498071 [00:14<00:01, 11046731.64it/s]
 93%|█████████▎| 158041088/170498071 [00:14<00:01, 11153902.38it/s]
 93%|█████████▎| 159286272/170498071 [00:15<00:00, 11495953.86it/s]
 94%|█████████▍| 160531456/170498071 [00:15<00:00, 11759881.01it/s]
 95%|█████████▍| 161809408/170498071 [00:15<00:00, 12040688.94it/s]
 96%|█████████▌| 163026944/170498071 [00:15<00:00, 12066384.42it/s]
 96%|█████████▋| 164242432/170498071 [00:15<00:00, 12030570.11it/s]
 97%|█████████▋| 165451776/170498071 [00:15<00:00, 11715975.80it/s]
 98%|█████████▊| 166629376/170498071 [00:15<00:00, 11429772.28it/s]
 98%|█████████▊| 167777280/170498071 [00:15<00:00, 11396536.97it/s]
 99%|█████████▉| 168921088/170498071 [00:15<00:00, 11335778.16it/s]
170499072it [00:16, 10634117.63it/s]                               

(train_cifar pid=53065) Extracting /Users/kai/ray_results/train_cifar_2022-07-22_16-38-50/train_cifar_66098_00000_0_batch_size=4,l1=128,l2=128,lr=0.0005_2022-07-22_16-38-57/data/cifar-10-python.tar.gz to /Users/kai/ray_results/train_cifar_2022-07-22_16-38-50/train_cifar_66098_00000_0_batch_size=4,l1=128,l2=128,lr=0.0005_2022-07-22_16-38-57/data
(train_cifar pid=53065) Files already downloaded and verified

(train_cifar pid=53065) /Users/kai/.pyenv/versions/3.7.7/lib/python3.7/site-packages/torch/nn/functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at  ../c10/core/TensorImpl.h:1156.)
(train_cifar pid=53065)   return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

(train_cifar pid=53078) Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to /Users/kai/ray_results/train_cifar_2022-07-22_16-38-50/train_cifar_66098_00001_1_batch_size=4,l1=128,l2=64,lr=0.0099_2022-07-22_16-39-01/data/cifar-10-python.tar.gz

  0%|          | 0/170498071 [00:00<?, ?it/s]
  0%|          | 33792/170498071 [00:00<13:49, 205600.99it/s]
  0%|          | 197632/170498071 [00:00<03:15, 869741.20it/s]
  0%|          | 443392/170498071 [00:00<02:17, 1233031.00it/s]
  1%|          | 1868800/170498071 [00:00<00:37, 4499043.51it/s]
  2%|▏         | 2589696/170498071 [00:00<00:31, 5313052.46it/s]
  2%|▏         | 3671040/170498071 [00:00<00:23, 6978488.44it/s]
  3%|▎         | 4703232/170498071 [00:00<00:20, 7988630.70it/s]
  3%|▎         | 5817344/170498071 [00:01<00:18, 8937646.63it/s]
  4%|▍         | 6996992/170498071 [00:01<00:16, 9785571.02it/s]
  5%|▍         | 8225792/170498071 [00:01<00:15, 10363873.64it/s]
  5%|▌         | 9307136/170498071 [00:01<00:15, 10452563.74it/s]
  6%|▌         | 10361856/170498071 [00:01<00:15, 10414666.33it/s]
  7%|▋         | 11469824/170498071 [00:01<00:14, 10611373.09it/s]
  7%|▋         | 12600320/170498071 [00:01<00:14, 10817759.86it/s]
  8%|▊         | 13685760/170498071 [00:01<00:14, 10674349.48it/s]
  9%|▊         | 14779392/170498071 [00:01<00:14, 10735551.44it/s]
  9%|▉         | 15893504/170498071 [00:01<00:14, 10841588.52it/s]
 10%|▉         | 17024000/170498071 [00:02<00:14, 10935080.88it/s]
 11%|█         | 18118656/170498071 [00:02<00:14, 10734313.48it/s]
 11%|█▏        | 19193856/170498071 [00:02<00:14, 10736033.30it/s]
 12%|█▏        | 20284416/170498071 [00:02<00:13, 10784705.38it/s]
 13%|█▎        | 22594560/170498071 [00:02<00:13, 11197004.04it/s]
 14%|█▍        | 23774208/170498071 [00:02<00:12, 11327905.78it/s]
 15%|█▍        | 24953856/170498071 [00:02<00:12, 11441768.69it/s]
 15%|█▌        | 26098688/170498071 [00:02<00:12, 11317906.35it/s]
 16%|█▌        | 27231232/170498071 [00:02<00:12, 11304969.94it/s]

(train_cifar pid=53065) [1,  2000] loss: 2.289

 17%|█▋        | 28558336/170498071 [00:03<00:11, 11857850.10it/s]
 17%|█▋        | 29745152/170498071 [00:03<00:11, 11779914.21it/s]
 18%|█▊        | 30923776/170498071 [00:03<00:12, 11470109.27it/s]
 19%|█▉        | 32072704/170498071 [00:03<00:12, 10962954.99it/s]
 19%|█▉        | 33244160/170498071 [00:03<00:12, 11138489.87it/s]
 20%|██        | 34362368/170498071 [00:03<00:12, 11060657.86it/s]
 21%|██        | 35488768/170498071 [00:03<00:12, 11075482.88it/s]
 22%|██▏       | 36733952/170498071 [00:03<00:11, 11463576.76it/s]
 22%|██▏       | 37946368/170498071 [00:03<00:11, 11644667.40it/s]
 23%|██▎       | 39207936/170498071 [00:03<00:11, 11771113.78it/s]
 24%|██▍       | 40551424/170498071 [00:04<00:10, 12239670.49it/s]
 25%|██▍       | 41777152/170498071 [00:04<00:10, 11724623.32it/s]
 25%|██▌       | 42959872/170498071 [00:04<00:10, 11728894.89it/s]
 26%|██▌       | 44237824/170498071 [00:04<00:10, 11980491.97it/s]
 27%|██▋       | 45474816/170498071 [00:04<00:10, 12094090.27it/s]
 27%|██▋       | 46687232/170498071 [00:04<00:10, 11965519.76it/s]
 28%|██▊       | 47886336/170498071 [00:04<00:12, 10214803.94it/s]
 29%|██▉       | 49693696/170498071 [00:04<00:09, 12272562.75it/s]
 30%|██▉       | 50979840/170498071 [00:04<00:09, 12216219.96it/s]
 31%|███       | 52242432/170498071 [00:05<00:10, 11819987.87it/s]
 31%|███▏      | 53508096/170498071 [00:05<00:09, 12050298.93it/s]
 32%|███▏      | 54735872/170498071 [00:05<00:09, 12065940.19it/s]
 33%|███▎      | 55958528/170498071 [00:05<00:09, 12089312.70it/s]
 34%|███▎      | 57274368/170498071 [00:05<00:09, 12399347.53it/s]
 34%|███▍      | 58523648/170498071 [00:05<00:09, 12262574.22it/s]
 35%|███▌      | 59756544/170498071 [00:05<00:09, 12035557.70it/s]
 36%|███▌      | 60998656/170498071 [00:05<00:09, 12133520.71it/s]
 37%|███▋      | 62276608/170498071 [00:05<00:08, 12211085.29it/s]
 37%|███▋      | 63501312/170498071 [00:06<00:08, 12042366.84it/s]
 38%|███▊      | 64708608/170498071 [00:06<00:08, 12023498.31it/s]
 39%|███▊      | 65912832/170498071 [00:06<00:08, 11757085.89it/s]
 39%|███▉      | 67090432/170498071 [00:06<00:09, 11320601.68it/s]
 40%|████      | 68227072/170498071 [00:06<00:09, 11258567.04it/s]
 41%|████      | 69355520/170498071 [00:06<00:09, 11082937.77it/s]
 41%|████▏     | 70465536/170498071 [00:06<00:09, 10911735.04it/s]
 42%|████▏     | 71558144/170498071 [00:06<00:09, 10780589.37it/s]
 43%|████▎     | 72680448/170498071 [00:06<00:08, 10893260.16it/s]
 43%|████▎     | 73771008/170498071 [00:06<00:08, 10841439.34it/s]
 44%|████▍     | 74856448/170498071 [00:07<00:08, 10769785.94it/s]
 45%|████▍     | 75940864/170498071 [00:07<00:08, 10763959.14it/s]
 45%|████▌     | 77054976/170498071 [00:07<00:08, 10849984.90it/s]
 46%|████▌     | 78140416/170498071 [00:07<00:08, 10726097.11it/s]
 47%|████▋     | 79315968/170498071 [00:07<00:08, 11026350.29it/s]
 47%|████▋     | 80462848/170498071 [00:07<00:08, 11146851.17it/s]
 48%|████▊     | 81593344/170498071 [00:07<00:08, 11105620.54it/s]

(train_cifar pid=53065) [1,  4000] loss: 1.058

 49%|████▊     | 82854912/170498071 [00:07<00:07, 11527296.83it/s]
 49%|████▉     | 84008960/170498071 [00:07<00:07, 11414359.53it/s]
 50%|████▉     | 85151744/170498071 [00:07<00:07, 10991470.33it/s]
 51%|█████     | 86254592/170498071 [00:08<00:07, 10779941.74it/s]
 51%|█████▏    | 87426048/170498071 [00:08<00:07, 10916081.56it/s]
 52%|█████▏    | 88548352/170498071 [00:08<00:07, 11004500.00it/s]
 53%|█████▎    | 89670656/170498071 [00:08<00:07, 11052931.99it/s]
 53%|█████▎    | 90850304/170498071 [00:08<00:07, 11002613.99it/s]
 54%|█████▍    | 92144640/170498071 [00:08<00:06, 11551701.57it/s]
 55%|█████▍    | 93357056/170498071 [00:08<00:06, 11700840.36it/s]
 55%|█████▌    | 94553088/170498071 [00:08<00:06, 11639934.12it/s]
 56%|█████▌    | 95724544/170498071 [00:08<00:06, 11661673.96it/s]
 57%|█████▋    | 96891904/170498071 [00:09<00:06, 11214261.52it/s]
 57%|█████▋    | 98017280/170498071 [00:09<00:06, 10988710.06it/s]
 58%|█████▊    | 99120128/170498071 [00:09<00:06, 10959602.56it/s]
 59%|█████▉    | 100218880/170498071 [00:09<00:06, 10938597.56it/s]
 59%|█████▉    | 101314560/170498071 [00:09<00:06, 10739220.45it/s]
 60%|██████    | 102417408/170498071 [00:09<00:06, 10815208.28it/s]
 61%|██████    | 103500800/170498071 [00:09<00:06, 10820566.49it/s]
 61%|██████▏   | 104678400/170498071 [00:09<00:05, 11089700.09it/s]
 62%|██████▏   | 105907200/170498071 [00:09<00:05, 11414720.98it/s]
 64%|██████▎   | 108348416/170498071 [00:10<00:05, 11853907.96it/s]
 64%|██████▍   | 109593600/170498071 [00:10<00:05, 12026577.90it/s]
 65%|██████▍   | 110797824/170498071 [00:10<00:05, 11649481.70it/s]
 66%|██████▌   | 111966208/170498071 [00:10<00:05, 11295371.19it/s]
 67%|██████▋   | 114279424/170498071 [00:10<00:05, 11174140.17it/s]
 68%|██████▊   | 115491840/170498071 [00:10<00:04, 11428872.46it/s]
 68%|██████▊   | 116638720/170498071 [00:10<00:04, 11376576.33it/s]
 69%|██████▉   | 117779456/170498071 [00:10<00:04, 11128529.20it/s]
 70%|██████▉   | 118894592/170498071 [00:10<00:04, 11105995.24it/s]
 70%|███████   | 120046592/170498071 [00:11<00:04, 11194307.10it/s]
 71%|███████   | 121291776/170498071 [00:11<00:04, 11529028.42it/s]
 72%|███████▏  | 122504192/170498071 [00:11<00:04, 11678012.53it/s]
 73%|███████▎  | 123673600/170498071 [00:11<00:04, 11616160.53it/s]
 73%|███████▎  | 124896256/170498071 [00:11<00:03, 11792603.53it/s]
 74%|███████▍  | 126092288/170498071 [00:11<00:03, 11722249.72it/s]
 75%|███████▍  | 127265792/170498071 [00:11<00:03, 11438563.16it/s]
 75%|███████▌  | 128411648/170498071 [00:11<00:03, 11383743.46it/s]
 76%|███████▌  | 129582080/170498071 [00:11<00:03, 11459099.89it/s]
 77%|███████▋  | 130728960/170498071 [00:12<00:03, 11417116.03it/s]
 77%|███████▋  | 131924992/170498071 [00:12<00:03, 11562162.11it/s]

(train_cifar pid=53065) [1,  6000] loss: 0.633

 78%|███████▊  | 133082112/170498071 [00:12<00:04, 8762729.97it/s] 
 79%|███████▊  | 134251520/170498071 [00:12<00:03, 9459387.31it/s]
 79%|███████▉  | 135414784/170498071 [00:12<00:03, 10010405.01it/s]
 80%|████████  | 136610816/170498071 [00:12<00:03, 10523298.83it/s]
 82%|████████▏ | 138980352/170498071 [00:12<00:02, 11012610.93it/s]
 82%|████████▏ | 140198912/170498071 [00:12<00:02, 11337292.82it/s]
 83%|████████▎ | 141378560/170498071 [00:13<00:02, 11433312.45it/s]
 84%|████████▎ | 142537728/170498071 [00:13<00:02, 11225954.39it/s]
 84%|████████▍ | 143672320/170498071 [00:13<00:02, 11106773.35it/s]
 85%|████████▍ | 144884736/170498071 [00:13<00:02, 11392243.15it/s]
 86%|████████▌ | 146064384/170498071 [00:13<00:02, 11461605.17it/s]
 86%|████████▋ | 147276800/170498071 [00:13<00:01, 11644469.12it/s]
 87%|████████▋ | 148587520/170498071 [00:13<00:01, 12032938.55it/s]
 88%|████████▊ | 149793792/170498071 [00:13<00:01, 11625996.36it/s]
 89%|████████▊ | 150961152/170498071 [00:13<00:01, 11353618.86it/s]
 89%|████████▉ | 152100864/170498071 [00:13<00:01, 11210110.17it/s]
 90%|████████▉ | 153225216/170498071 [00:14<00:01, 10989692.27it/s]
 91%|█████████ | 154327040/170498071 [00:14<00:02, 6395856.39it/s] 
 92%|█████████▏| 157336576/170498071 [00:14<00:01, 10963246.92it/s]
 93%|█████████▎| 158843904/170498071 [00:14<00:01, 10672622.13it/s]
 94%|█████████▍| 160198656/170498071 [00:15<00:01, 7535629.34it/s] 
 96%|█████████▌| 163070976/170498071 [00:15<00:00, 11093551.80it/s]
 97%|█████████▋| 164648960/170498071 [00:15<00:00, 11143733.74it/s]
 97%|█████████▋| 166092800/170498071 [00:15<00:00, 11190526.07it/s]
 98%|█████████▊| 167443456/170498071 [00:15<00:00, 11480616.98it/s]
 99%|█████████▉| 168762368/170498071 [00:15<00:00, 11429373.68it/s]
170499072it [00:15, 10815137.24it/s]                               

(train_cifar pid=53078) Extracting /Users/kai/ray_results/train_cifar_2022-07-22_16-38-50/train_cifar_66098_00001_1_batch_size=4,l1=128,l2=64,lr=0.0099_2022-07-22_16-39-01/data/cifar-10-python.tar.gz to /Users/kai/ray_results/train_cifar_2022-07-22_16-38-50/train_cifar_66098_00001_1_batch_size=4,l1=128,l2=64,lr=0.0099_2022-07-22_16-39-01/data
(train_cifar pid=53065) [1,  8000] loss: 0.434
(train_cifar pid=53078) Files already downloaded and verified

(train_cifar pid=53078) /Users/kai/.pyenv/versions/3.7.7/lib/python3.7/site-packages/torch/nn/functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at  ../c10/core/TensorImpl.h:1156.)
(train_cifar pid=53078)   return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

(train_cifar pid=53065) [1, 10000] loss: 0.325
(train_cifar pid=53078) [1,  2000] loss: 2.117
Result for train_cifar_66098_00000:
  accuracy: 0.4004
  date: 2022-07-22_16-39-47
  done: false
  experiment_id: 6512b700fdb64a458c3496f36ea1776c
  hostname: Kais-MacBook-Pro.local
  iterations_since_restore: 1
  loss: 1.625945699906349
  node_ip: 127.0.0.1
  pid: 53065
  should_checkpoint: true
  time_since_restore: 45.849108934402466
  time_this_iter_s: 45.849108934402466
  time_total_s: 45.849108934402466
  timestamp: 1658504387
  timesteps_since_restore: 0
  training_iteration: 1
  trial_id: '66098_00000'
  warmup_time: 0.003801107406616211
  
(train_cifar pid=53078) [1,  4000] loss: 0.983
(train_cifar pid=53065) [2,  2000] loss: 1.582
(train_cifar pid=53078) [1,  6000] loss: 0.647
(train_cifar pid=53065) [2,  4000] loss: 0.758
(train_cifar pid=53078) [1,  8000] loss: 0.489
(train_cifar pid=53065) [2,  6000] loss: 0.499
(train_cifar pid=53065) [2,  8000] loss: 0.365
(train_cifar pid=53078) [1, 10000] loss: 0.388
Result for train_cifar_66098_00001:
  accuracy: 0.2915
  date: 2022-07-22_16-40-09
  done: true
  experiment_id: 6410c16837024e5e903317c212a4af63
  hostname: Kais-MacBook-Pro.local
  iterations_since_restore: 1
  loss: 1.9046219720602036
  node_ip: 127.0.0.1
  pid: 53078
  should_checkpoint: true
  time_since_restore: 64.97207283973694
  time_this_iter_s: 64.97207283973694
  time_total_s: 64.97207283973694
  timestamp: 1658504409
  timesteps_since_restore: 0
  training_iteration: 1
  trial_id: '66098_00001'
  warmup_time: 0.0027120113372802734
  
(train_cifar pid=53065) [2, 10000] loss: 0.285
Result for train_cifar_66098_00000:
  accuracy: 0.4877
  date: 2022-07-22_16-40-13
  done: true
  experiment_id: 6512b700fdb64a458c3496f36ea1776c
  hostname: Kais-MacBook-Pro.local
  iterations_since_restore: 2
  loss: 1.421571186053753
  node_ip: 127.0.0.1
  pid: 53065
  should_checkpoint: true
  time_since_restore: 72.61763620376587
  time_this_iter_s: 26.768527269363403
  time_total_s: 72.61763620376587
  timestamp: 1658504413
  timesteps_since_restore: 0
  training_iteration: 2
  trial_id: '66098_00000'
  warmup_time: 0.003801107406616211
  

2022-07-22 16:40:14,050	INFO tune.py:738 -- Total run time: 77.27 seconds (76.42 seconds for the tuning loop).

Best trial config: {'l1': 128, 'l2': 128, 'lr': 0.00046907397024184945, 'batch_size': 4}
Best trial final validation loss: 1.421571186053753
Best trial final validation accuracy: 0.4877
Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to ./data/cifar-10-python.tar.gz

Extracting ./data/cifar-10-python.tar.gz to ./data
Files already downloaded and verified

/Users/kai/.pyenv/versions/3.7.7/lib/python3.7/site-packages/torch/nn/functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at  ../c10/core/TensorImpl.h:1156.)
  return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

Best trial test set accuracy: 0.4939

If you run the code, an example output could look like this:

  Number of trials: 10 (10 TERMINATED)
  +-------------------------+------------+-------+------+------+-------------+--------------+---------+------------+----------------------+
  | Trial name              | status     | loc   |   l1 |   l2 |          lr |   batch_size |    loss |   accuracy |   training_iteration |
  |-------------------------+------------+-------+------+------+-------------+--------------+---------+------------+----------------------|
  | train_cifar_87d1f_00000 | TERMINATED |       |   64 |    4 | 0.00011629  |            2 | 1.87273 |     0.244  |                    2 |
  | train_cifar_87d1f_00001 | TERMINATED |       |   32 |   64 | 0.000339763 |            8 | 1.23603 |     0.567  |                    8 |
  | train_cifar_87d1f_00002 | TERMINATED |       |    8 |   16 | 0.00276249  |           16 | 1.1815  |     0.5836 |                   10 |
  | train_cifar_87d1f_00003 | TERMINATED |       |    4 |   64 | 0.000648721 |            4 | 1.31131 |     0.5224 |                    8 |
  | train_cifar_87d1f_00004 | TERMINATED |       |   32 |   16 | 0.000340753 |            8 | 1.26454 |     0.5444 |                    8 |
  | train_cifar_87d1f_00005 | TERMINATED |       |    8 |    4 | 0.000699775 |            8 | 1.99594 |     0.1983 |                    2 |
  | train_cifar_87d1f_00006 | TERMINATED |       |  256 |    8 | 0.0839654   |           16 | 2.3119  |     0.0993 |                    1 |
  | train_cifar_87d1f_00007 | TERMINATED |       |   16 |  128 | 0.0758154   |           16 | 2.33575 |     0.1327 |                    1 |
  | train_cifar_87d1f_00008 | TERMINATED |       |   16 |    8 | 0.0763312   |           16 | 2.31129 |     0.1042 |                    4 |
  | train_cifar_87d1f_00009 | TERMINATED |       |  128 |   16 | 0.000124903 |            4 | 2.26917 |     0.1945 |                    1 |
  +-------------------------+------------+-------+------+------+-------------+--------------+---------+------------+----------------------+


  Best trial config: {'l1': 8, 'l2': 16, 'lr': 0.0027624906698231976, 'batch_size': 16, 'data_dir': '...'}
  Best trial final validation loss: 1.1815014744281769
  Best trial final validation accuracy: 0.5836
  Best trial test set accuracy: 0.5806

As you can see, most trials have been stopped early in order to avoid wasting resources. The best performing trial achieved a validation accuracy of about 58%, which could be confirmed on the test set.

So that’s it! You can now tune the parameters of your PyTorch models.

See More PyTorch Examples#

  • MNIST PyTorch Example: Converts the PyTorch MNIST example to use Tune with the function-based API. Also shows how to easily convert something relying on argparse to use Tune.

  • PBT ConvNet Example: Example training a ConvNet with checkpointing in function API.

  • MNIST PyTorch Trainable Example: Converts the PyTorch MNIST example to use Tune with Trainable API. Also uses the HyperBandScheduler and checkpoints the model at the end.