PyTorch Porting Tutorial¶

Determined provides a high-level framework APIs for PyTorch, Keras, and Estimators that let users describe their model without boilerplate code. Determined reduces boilerplate by providing a state-of-the-art training loop that provides distributed training, hyperparameter search, automatic mixed precision, reproducibility, and many more features.

In this guide, we’ll walk through an example and provide helpful hints to successfully organize PyTorch code into Determined’s PyTorchTrial API. Once your code is in the PyTorchTrial format, you can easily take advantage of Determined Ai’s open-source platform.

We suggest you follow along using the source code found on Determined’s GitHub repo.

While all codebases are different, code to perform deep learning training tends to follow a typical pattern. Usually, there is a model, optimizer, data, and a learning rate scheduler. determined.pytorch.PyTorchTrial follows this pattern to reduce porting friction. To port, we will copy the core machine learning code based on the traditional training pieces, while deleting the unnecessary boilerplate code. We suggest isolating these components in the following order:

Model
Optimizer
Data
Train/validate batch
Learning Rate Scheduler
Other features such as automatic mixed precision (AMP), gradient clipping, and others

We will port each section following the same steps: First, we copy all the relevant code over. Then, we will remove boilerplate code, and update all relevant objects to use the context object. Finally, we will replace all configurations or hyperparameters.

Preparation¶

Before we begin, we need to create our core files. Determined requires two files to be defined: the model definition and the experiment configuration.

Model Definition¶

The model definition contains the trial class, which contains the model definition and training loop. Your deep learning framework defines which Determined.Trial() class to inherit. In our case, we will be working with PyTorch, which means we will be working with determined.pytorch.PyTorchTrial. Once we inherit this class, we will be required to override specific functions that represent a training script’s core components. When starting a new port, it is helpful to work off of a skeleton to keep track of what is still required. A good starting template can be found below:

from typing import Any, Dict, Union, Sequence
from determined.pytorch import DataLoader, PyTorchTrial, PyTorchTrialContext

TorchData = Union[Dict[str, torch.Tensor], Sequence[torch.Tensor], torch.Tensor]

class MyTrial(PyTorchTrial):
    def __init__(self, context: PyTorchTrialContext) -> None:
        self.context = context

    def build_training_data_loader(self) -> DataLoader:
        return DataLoader()

    def build_validation_data_loader(self) -> DataLoader:
        return DataLoader()

    def train_batch(self, batch: TorchData, epoch_idx: int, batch_idx: int)  -> Dict[str, Any]:
        return {}

    def evaluate_batch(self, batch: TorchData) -> Dict[str, Any]:
        return {}

Experiment Configuration¶

We also need to create an experiment configuration file. This file defines specific experiment information such as: number of samples, training length, and hyperparameters.

We recommend that all hyperparameters be added to this file. A good starting point is taking any command line arguments - you may see this as parser args - from the original script and immediately adding them to your file.

In our case, since we’re using hyperparameters that don’t change, we’ve given this specific config the name const.yaml:

description: Imagenet_Pytorch_const
hyperparameters:
    global_batch_size: 256
    dense1: 128
    data: /mnt/data
    arch: resnet18
    workers: 4
    start-epoch: 0
    lr: 0.1
    momentum: 0.9
    weight_decay: 1e-4
    pretrained: True
records_per_epoch: 60000
searcher:
    name: single
    metric: val_loss
    smaller_is_better: false
    max_length:
        epochs: 10
entrypoint: model_def:ImagenetTrial
max_restarts: 0

For now, we don’t have to worry much about the other fields; however, we suggest setting max_restarts to zero so Determined will not retry running the experiment. For more information on experiment configuration, see the experiment configuration reference.

Model¶

Now that we’ve finished the prep work, we can begin porting by creating the model. Model code will be placed in the Trial’s __init__() function.

To refresh, as we work on the model, we want to follow this checklist:

Remove boilerplate code.
Copy all relevant code over.
Update all relevant objects to use the ‘context’ object.
Replace all configurations or hyperparameters.

Remove Boilerplate Code and Copy All Relevant Code¶

Based on the checklist, we want to first remove all boilerplate code, such as code related to distributed training or device management. As we remove boilerplate code, we can immediately copy relevant code, such as the model creation, to our PyTorchTrial. In the original script most of the model code is found in lines 119-168, where it defines the model and sets up the GPU and script for distributed training. Since Determined handles much of this logic, we can remove a lot of this as boilerplate.

Let’s work through these lines:

if args.gpu is not None:
    print("Use GPU: {} for training".format(args.gpu))

In the experiment configuration file, we define the number of resources, usually GPUs; therefore, we can omit this from the model definition.

if args.distributed:
    if args.dist_url == "env://" and args.rank == -1:
        args.rank = int(os.environ["RANK"])
    if args.multiprocessing_distributed:
        # For multiprocessing distributed training, rank needs to be the
        # global rank among all the processes
        args.rank = args.rank * ngpus_per_node + gpu
    dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
                            world_size=args.world_size, rank=args.rank)

Determined will automatically set up horovod for the user. If you would like to access the rank (typically used to view per GPU training), you can get it by calling self.context.distributed.rank.

if args.pretrained:
    print("=> using pre-trained model '{}'".format(args.arch))
    model = models.__dict__[args.arch](pretrained=True)
else:
    print("=> creating model '{}'".format(args.arch))
    model = models.__dict__[args.arch]()

Here is where we actually define the model. We will copy and paste this code directly into our __init__ function. For now, we can leave it as it was copied, but we will return to update this code a bit later on.

if not torch.cuda.is_available():
    print('using CPU, this will be slow')
elif args.distributed:
    # For multiprocessing distributed, DistributedDataParallel constructor
    # should always set the single device scope, otherwise,
    # DistributedDataParallel will use all available devices.
    if args.gpu is not None:
        torch.cuda.set_device(args.gpu)
        model.cuda(args.gpu)
        # When using a single GPU per process and per
        # DistributedDataParallel, we need to divide the batch size
        # ourselves based on the total number of GPUs we have
        args.batch_size = int(args.batch_size / ngpus_per_node)
        args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node)
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
    else:
        model.cuda()
        # DistributedDataParallel will divide and allocate batch_size to all
        # available GPUs if device_ids are not set
        model = torch.nn.parallel.DistributedDataParallel(model)
elif args.gpu is not None:
    torch.cuda.set_device(args.gpu)
    model = model.cuda(args.gpu)
else:
    # DataParallel will divide and allocate batch_size to all available GPUs
    if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
        model.features = torch.nn.DataParallel(model.features)
        model.cuda()
    else:
        model = torch.nn.DataParallel(model).cuda()

This snippet sets up CUDA and converts it to a distributed model. Because Determined handles distributed training automatically, this code is unnecessary. When we create the model, we will wrap it with self.context.wrap_model(model), which will convert the model to distributed if needed.

Update Objects and Replace HP Configurations¶

We have copied all relevant code over and now need to clean it up. Our current __init__ function looks something like this:

def __init__(self, context: PyTorchTrialContext):
    self.context = context
    if args.pretrained:
        print("=> using pre-trained model '{}'".format(args.arch))
        model = models.__dict__[args.arch](pretrained=True)
    else:
        print("=> creating model '{}'".format(args.arch))
        model = models.__dict__[args.arch]()

First, we update all references to the parser arguments. Everywhere args are used will be changed to self.context.get_hparams(). This function will give you access to all the hyperparameters in the experiment configuration file. By converting the hyperparameters to be accessed in the experiment configuration, it allows for better experiment tracking, and makes it easier to quickly run a searcher experiment.

Finally, we need to wrap our model, so Determined can handle all the distributed training code we previously removed:

self.model = self.context.wrap_model(model)

After all of these changes, we are left with the code below:

def __init__(self, context: PyTorchTrialContext):
    self.context = context

    arch = self.context.get_hparam("arch")
    if self.context.get_hparam("pretrained"):
        print("=> using pre-trained model '{}'".format(arch))
        model = models.__dict__[arch](pretrained=True)
    else:
        print("=> creating model '{}'".format(arch))
        model = models.__dict__[arch]()

    self.model = self.context.wrap_model(model)

Optimizer/Loss¶

Next, we will port the optimizer and loss functions. The optimizer and loss will be placed in the __init__() function.

Remove Boilerplate Code and Copy All Relevant Code¶

Once again, we copy the relevant optimizer and loss definitions. In the original model, the optimizer is defined with one line, which we copy over directly:

optimizer = torch.optim.SGD(model.parameters(), args.lr,
                            momentum=args.momentum,
                            weight_decay=args.weight_decay)

For the loss, this example uses CrossEntropyLoss(). This can be added to PyTorchTrial with one line.

self.criterion = nn.CrossEntropyLoss()

Update Objects and Replace HP Configurations¶

Now we update the arguments to reference the experiment configuration.

optimizer = torch.optim.SGD(self.model.parameters(), self.context.get_hparam("lr"), momentum=self.context.get_hparam("momentum"), weight_decay=self.context.get_hparam("weight_decay"))
self.optimizer = self.context.wrap_optimizer(optimizer)

You may notice self.context.get_hparams() can become long. A simple trick is to set self.context.get_hparams to self.hparams. Then you can use self.hparams[“variable”].

The init function should now look something like this.

def __init__(self, context: PyTorchTrialContext):
    self.context = context

    arch = self.context.get_hparam("arch")
    if self.context.get_hparam("pretrained"):
        print("=> using pre-trained model '{}'".format(arch))
        model = models.__dict__[arch](pretrained=True)
    else:
        print("=> creating model '{}'".format(arch))
        model = models.__dict__[arch]()

    self.model = self.context.wrap_model(model)

    optimizer = torch.optim.SGD(self.model.parameters(), self.context.get_hparam("lr"), momentum=self.context.get_hparam("momentum"), weight_decay=self.context.get_hparam("weight_decay"))
    self.optimizer = self.context.wrap_optimizer(optimizer)

    self.criterion = nn.CrossEntropyLoss()

We have been able to remove over 80 lines of code by porting to Determined!

Data¶

Now, we can fill out build_train_data_loader() and build_validation_data_loader(). Both of these data loading functions return a determined.DataLoader. A determined.DataLoader expects the same parameters as a torch.DataLoader and will handle distributed training setup.

The original script handles the data in lines 202 - 233. For the data loaders, we follow the same procedure for porting.

Remove Boilerplate Code and Copy All Relevant Code¶

In the original code, the data is loaded based on the path, and prepared for distributed training as seen below:

traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                std=[0.229, 0.224, 0.225])

train_dataset = datasets.ImageFolder(
    traindir,
    transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        normalize,
    ]))
if args.distributed:
    train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
else:
    train_sampler = None

We will bring all the code over except the if args.distribued clause since Determined will automatically do the right thing when running a distributed training job.

Update Objects and Replace HP Configurations¶

There are a few pieces that need to be changed. First, the data location should be set to a class variable: self.download_directory. During distributed training, the data should be downloaded to unique directories based on rank to prevent multiple download processes (one process per GPU) from conflicting with one another. This root directory will be defined based on self.hparams and will point to where the data is stored within the docker container. If you want to learn more about how to access data with Determined, check out our documentation.

We also update the torch.Dataloader to be a determined.pytorch.DataLoader. The batch_size will be set to self.context.get_per_slot_batch_size(). We set batch_size to self.context.get_per_slot_batch_size() which automatically calculates the per-gpu batch size based on global_batch_size and slots_per_trial as defined in the experiment configuration. By using self.context.get_per_slot_batch_size(), Determined will assign the appropriate per GPU batch size.

The train function will look something like this:

def build_training_data_loader(self):
    traindir = os.path.join(self.download_directory, 'train')
    self.normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                std=[0.229, 0.224, 0.225])

    train_dataset = datasets.ImageFolder(
        traindir,
        transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            self.normalize,
        ]))

    train_loader = determined.pytorch.DataLoader(
        train_dataset, batch_size=self.context.get_per_slot_batch_size(), shuffle=True,num_workers=self.context.get_hparam("workers", pin_memory=True))

In this example, we are using ImageNet as the dataset. If you do not have access to the dataset, the CIFAR-10 dataset can be accessed with the code below:

def build_training_data_loader(self):
    transform = transforms.Compose(
        [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
    )
    trainset = torchvision.datasets.CIFAR10(
        root=self.download_directory, train=True, download=True, transform=transform
    )
    return DataLoader(trainset, batch_size=self.context.get_per_slot_batch_size())

Train / Validation Batch¶

It’s time to set up the train_batch function. Typically in PyTorch, you loop through the DataLoader to access and train your model one batch at a time. You can usually identify this code by finding the common code snippet: for batch in dataloader. In Determined, train_batch() also provides one batch at a time, so we can copy the code directly into our function.

Remove Boilerplate Code and Copy All Relevant Code¶

In the original implementation, we find the core training loop.

for i, (images, target) in enumerate(train_loader):
    # measure data loading time
    data_time.update(time.time() - end)

    if args.gpu is not None:
        images = images.cuda(args.gpu, non_blocking=True)
    if torch.cuda.is_available():
        target = target.cuda(args.gpu, non_blocking=True)

    # compute output
    output = model(images)
    loss = criterion(output, target)

    # measure accuracy and record loss
    acc1, acc5 = accuracy(output, target, topk=(1, 5))
    losses.update(loss.item(), images.size(0))
    top1.update(acc1[0], images.size(0))
    top5.update(acc5[0], images.size(0))

    # compute gradient and do SGD step
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    # measure elapsed time
    batch_time.update(time.time() - end)
    end = time.time()

    if i % args.print_freq == 0:
        progress.display(i)

As you noticed above, the loop manages the per-batch metrics. Determined automatically averages and displays the metrics returned in train_batch allowing us to remove print frequency code and the metric arrays.

Update Objects and Replace HP Configurations¶

Now, we will convert some PyTorch functions to now use Determined’s equivalent. We need to change loss.backward(), optim.zero_grad(), and optim.step(). The self.context object will be used to call loss.backwards and handle zeroing and stepping the optimizer. We update these functions respectively:

self.context.backward(loss)
self.context.step_optimizer(self.optimizer)

The final train_batch will look like:

def train_batch(self, batch: TorchData, epoch_idx: int, batch_idx: int):
    images, target = batch
    output = self.model(images)
    loss = self.criterion(output, target)
    acc1, acc5 = self.accuracy(output, target, topk=(1, 5))

    self.context.backward(loss)
    self.context.step_optimizer(self.optimizer)

    return {"loss": loss.item(), 'top1': acc1[0], 'top5': acc5[0]}

Code Check Point¶

At this point, you should be able to run your Determined model. Confirm that your model weights are loaded correctly, it can functionally run a batch, and all your hyperparameters are correctly accessing experiment configuration.

Learning Rate Scheduler¶

Determined has a few ways of managing the learning rate. Determined can automatically update every batch or epoch, or you can manage it yourself. In this case, we are doing the latter by using a custom function to handle the learning rate adjustment. We define it in the __init__() function and wrap it with self.context.wrap_lr_scheduler.

Next, we call the function in train_batch(). Since our model runs, we can also print the learning rate per batch or epoch to confirm the accuracy. In this case, we will update the learning rate to use torch.optim.StepLR() and wrap it with self.context.wrap_lr_scheduler.

def __init__(self, context):

    . . .

    lr_sch = torch.optim.lr_scheduler.StepLR(self.optimizer, gamma=.1, step_size=2)
    self.lr_sch = self.context.wrap_lr_scheduler(lr_sch, step_mode=LRScheduler.StepMode.STEP_EVERY_EPOCH)

Other Functionality¶

At this point, you can begin adding other features of your model. This may include using 16 FP (automatic mixed precision) or gradient clipping. It’s best to add one at a time to make it easier to check that each component is properly working. Determined has a wide range of examples to demonstrate several real-world use cases. Examples can be found on Determined’s github:

Helpful Hints¶

During porting, most of the time you can remove distributed training code.

If you are having trouble porting your model and would like to debug it prior to finishing the rest of the code, you can use fake data in the data loader. This lets you run and test other parts of the model_def.py.

Sometimes it’s useful just getting an “ugly” version of the code. This is where you first directly place the original code in the right function without updating any pieces.

Saving the extra model “features” until later helps you ensure the core functions are correct. This makes it easier to debug other portions of the script.

For more debugging tips, check out the how-to guide on model debugging.