dlhpcstarter 0.1.3

Deep Learning and HPC Starter Pack

Aims of this library:

• To be simple and easy to understand so that the focus is on the data science.
• To reduce the time taken from implementation to results.
• To promote rapid development of models via configuration files, class composition and/or class inheritance.
• Reduce boilerplate code (sections of code that are repeated in multiple places with little to no variation).
• To simplify cluster management and distributed computing with High Performance Computing (HPC).
• Be able to easily accommodate multiple research avenues simultaneously.
• To cooperatively improve the functionality and documentation of this repository to make it better!

Features:

• The Lightning LightningModule and Trainer are used to implement, train, and test models. It allows for many of the above aims to be accomplished, such as simplified distributed computing and a reduction of boilerplate code. It also allows us to simply use class inheritance and composition, allowing for rapid development.
• The Compose API of Hydra is used to create a hierarchical configuration, allowing for rapid development.
• Neptune.ai is used to track experiments; metric scores are automatically uploaded to Neptune.ai, allowing you to easily track your experiments from your browser.
• Scripts for submission to a cluster manager, such as SLURM are written for you. Also, cluster manager jobs are automatically resubmitted and resumed if they haven't finished before the time-limit.

Installation

The Deep Learning and HPC starter pack is available on PyPI:

pip install dlhpcstarter

Table of Contents

• Package map
• Tasks
• Models
• Development via Model Composition and Inheritance
• Configuration YAML files and argparse
• Development via Configuration Files
• Next level: Configuration composition via Hydra
• Stages and Trainer
• Tying it all together: main.py
• Cluster manager and distributed computing
• Monitoring using Neptune.ai
• Where all the outputs go: exp_dir
• Repository Wish List

Package map

---

The package is structured as follows:

├──  dlhpcstarter
│    │
│    ├── tools                     - for all other modules; tools that are repeadetly used.
│    ├── __main__.py               - __main__.py does the following:
│    │                                    1. Reads command line arguments using argparse.
│    │                                    2. Imports the 'stages' function for the task from task/
│    │                                    3. Loads the specified configuration .yaml for the job from 
│    │                                    4. Submits the job (the configuration + 'stages') to the 
│    │                                       cluster manager (or runs it locally if 'submit' is false).
│    └── cluster.py                - contains the cluster management object.
│    └── command_line_arguments.py - argparse for reading command line arguments.
│    └── trainer.py                - contains an optional wrapper for the Lightning Trainer.
│    └── utils.py                  - small utility definitions.

Tasks

---

Tasks can have any name. The name could be based on the data or the type of inference being made. For example:

• Two tasks have the same data but require different names due to differing predictions, e.g., MS-COCO Detection and MS-COCO Caption.
• Two tasks may have similar predictions but require different names due to differing data, e.g., MNIST and Chinese MNIST.

Some publicly available tasks include:

• Image classification tasks, e.g., MNIST, CIFAR10, CIFAR100, ImageNet.
• Object detection tasks, e.g., MS-COCO Detection.
• Image captioning detection tasks, e.g., MS-COCO Caption.
• Speech recognition tasks, e.g., LibriSpeech.
• Chest X-Ray report generation, e.g., MIMIC-CXR.

What does the name do?

It is used to separate the outputs of the experiment from other tasks.

Models

---

Please familiarise yourself with the Lightning LightningModule in order to correctly implement a model: https://pytorch-lightning.readthedocs.io/en/latest/common/lightning_module.html

A model is created using a Lightning LightningModule. Everything we need for the model can be placed in the LightningModule, including commonly used libraries and objects, for example:

• torch.nn.Module: base class for all neural networks in PyTorch.
• transformers: a library containing pre-trained Transformer models.
• torchvision: a library for image pre-processing and pre-trained computer vision models.
• torch.utils.data.Dataset: an object that processes each instance of a dataset.
• torch.utils.data.DataLoader: an object that samples mini-batches from a torch.utils.data.Dataset.

Note:

• The data pipeline could be implemented within the LightningModule or seperately from a model using a Lightning LightningDataModule. The LightningModule instance would then have to be given separately to the Lightning Trainer instance.

Example:

• An example model for cifar10 is in task/cifar10/model/baseline.py.

Development via Model Composition and Inheritance

---

To promote rapid development of models, one solution is to use class composition and/or inheritance. For example, we may have a baseline that not only includes a basic model, but also the data pipeline:

from lightning.pytorch import LightningModule
from torch.utils.data import DataLoader, random_split
import torchvision
import torch

class Baseline(LightningModule):
    def __init__(self, lr, ..., **kwargs):
        super(Baseline, self).__init__()
        self.save_hyperparameters()
        self.lr = lr
        self.model = torchvision.models.resnet18(...)

    def setup(self, stage=None):
        if stage == 'fit' or stage is None:
            train_set = torchvision.datasets.CIFAR10(...)
            self.train_set, self.val_set = random_split(train_set, [45000, 5000])

        if stage == 'test' or stage is None:
            self.test_set = torchvision.datasets.CIFAR10(...)

    def train_dataloader(self, shuffle=True):
        return DataLoader(self.train_set, ...)

    def val_dataloader(self):
        return DataLoader(self.val_set, ...)

    def test_dataloader(self):
        return DataLoader(self.test_set, ...)

    def configure_optimizers(self):
        optimiser = {'optimizer': torch.optim.SGD(self.parameters(), lr=self.lr, momentum=0.9)}
        return optimiser

    def forward(self, images):
        return self.model(images)

    def training_step(self, batch, batch_idx):
        images, labels = batch
        y_hat = self(images)
        loss = self.loss(y_hat, labels)
        self.log_dict({'train_loss': loss}, ...)
        return loss

    def validation_step(self, batch, batch_idx):
        images, labels = batch
        y_hat = self(images)
        loss = self.loss(y_hat, labels)
        self.val_accuracy(torch.argmax(y_hat['logits'], dim=1), labels)
        self.log_dict({'val_acc': self.val_accuracy, 'val_loss': loss}, ...)

    def test_step(self, batch, batch_idx):
        images, labels = batch
        y_hat = self(images)
        self.test_accuracy(torch.argmax(y_hat['logits'], dim=1), labels)
        self.log_dict({'test_acc': self.test_accuracy}, ...)

After training and testing the baseline, we may want to improve upon its performance. For example, if we wanted to make the following modifications:

• Use a DenseNet instead of a ResNet.
• Use the AdamW optimiser.
• Use a warmup learning rate scheduler.

All we would need to do is inherit the baseline and make our modifications:

from transformers import get_constant_schedule_with_warmup

class Inheritance(Baseline):

    def __init__(self, num_warmup_steps, **kwargs):
        super(Inheritance, self).__init__(**kwargs)
        self.save_hyperparameters()
        self.num_warmup_steps = num_warmup_steps
        self.model = torchvision.models.densenet121(...)

    def configure_optimizers(self):
        optimiser = {'optimizer': torch.optim.AdamW(self.parameters(), lr=self.lr)}
        optimiser['scheduler'] = {
            'scheduler': get_constant_schedule_with_warmup(optimiser['optimizer'], self.num_warmup_steps),
            'interval': 'step',
            'frequency': 1,
        }
        return optimiser

We could also construct a model that is the combination of the two via composition. For example, we may want to use everything from Baseline, but the optimiser from Inheritance:

from lightning.pytorch import LightningModule

class Composite(LightningModule):
    def __init__(self, **kwargs):
        self.baseline = Baseline(self, **kwargs)

    def setup(self, stage=None):
        self.baseline.setup(stage)

    def train_dataloader(self, shuffle=True):
        return self.baseline.train_dataloader(shuffle)

    def val_dataloader(self):
        return self.baseline.val_dataloader()

    def test_dataloader(self):
        return self.baseline.test_dataloader()

    def configure_optimizers(self):
        return Inheritance.configure_optimizers(self)  # Use configure_optimizers() from Inheritance.

    def forward(self, images):
        return self.baseline.forward(images)

    def training_step(self, batch, batch_idx):
        return self.baseline.training_step(batch, batch_idx)

    def validation_step(self, batch, batch_idx):
        return self.baseline.validation_step(batch, batch_idx)

    def test_step(self, batch, batch_idx):
        return self.baseline.test_step(batch, batch_idx)

Configuration YAML files and argparse

---

Currently, there are two methods for giving arguments:

1. Via command line arguments using the argparse module. argparse mainly handles paths, development stage flags (e.g., training and testing flags), and cluster manager arguments.
2. Via a configuration file stored in YAML format. Can handle all the arguments defined by the argparse plus more, including hyperparameters for the model.

The mandatory arguments include:

1. task, the name of the task.
2. config, relative or absolute path to the configuration file (with or without the extension).
3. module, the module that the model definition is housed.
4. definition, the class representing the model.
5. exp_dir, the experiment directory, i.e., where all outputs, including model checkpoints will be saved.
6. monitor, metric to monitor for ModelCheckpoint and EarlyStopping (optional), as well as test checkpoint loading (e.g., 'val_loss').
7. monitor_mode, whether the monitored metric is to be maximised or minimised ('max' or 'min').

task and config must be given as command line arguments for argparse:

dlhpcstarter --config task/cifar10/config/baseline --task cifar10

module, definition, and exp_dir can be given either as command line arguments, or be placed in the configuration file.

For each model of a task, we define a configuration. Hyperparameters, paths, as well as the device configuration can be stored in a configuration file. Configurations are in YAML format, e.g., task/cifar10/config/baseline.yaml.

Development via Configuration Files

---

If we have the following configuration file for the aforementioned CIFAR10 Baseline model, task/cifar10/config/baseline.yaml:

train: True
test: True
module: task.cifar10.model.baseline
definition: Baseline
monitor: 'val_acc'
monitor_mode: 'max'
lr: 1e-3
max_epochs: 32
mbatch_size: 32
num_workers: 5
exp_dir: /my/experiment/directory
dataset_dir: /my/datasets/directory

Another way we can improve upon the baseline model, i.e., the baseline configuration, is by modifying its hyperparameters. For example, we can still use Baseline, but alter the learning rate in task/cifar10/config/baseline_rev_a.yaml:

train: True
test: True
module: task.cifar10.model.baseline
definition: Baseline
monitor: 'val_acc'
monitor_mode: 'max'
lr: 1e-4  # modify this.
max_epochs: 32
mbatch_size: 32
num_workers: 5
exp_dir: /my/experiment/directory
dataset_dir: /my/datasets/directory

dlhpcstarter --config task/cifar10/config/baseline_rev_a --task cifar10

Next level: Configuration composition via Hydra

---

If your new configuration only modifies a few arguments of another configuration file, you can take advantage of the composition feature of Hydra. This makes creating task/cifar10/config/baseline_rev_a.yaml from the previous section easy. We simply add the arguments from task/cifar10/config/baseline.yaml by adding its name to the defaults list:

defaults:
  - baseline
  - _self_

lr: 1e-4

Note that other configuration files are imported with reference to the current configuration path (not the working directory).

Please note that groups are not being used, and packages should be placed using @_global_ if the configurations being used for composition are not in the same directory. For example, the following would not work with this repository as the arguments in hpc_paths will be grouped under paths:

defaults:
  - paths/hpc_paths
  - _self_

train: True
test: True
resumable: True
module: task.cifar10.model.baseline
definition: Baseline
monitor: 'val_acc'
monitor_mode: 'max'
lr: 1e-3
max_epochs: 3
mbatch_size: 32
num_workers: 5

To get around this, simply place @_global_ to remove the grouping:

defaults:
  - paths/hpc_paths@_global_  # changed here to remove "paths" grouping.
  - _self_

train: True
test: True
resumable: True
module: task.cifar10.model.baseline
definition: Baseline
monitor: 'val_acc'
monitor_mode: 'max'
lr: 1e-3
max_epochs: 3
mbatch_size: 32
num_workers: 5

This also allows us to organise configurations easily. For example, if we have the following directory structure:

├── task
│   └──  cifar10          
│        └── config  
│            ├── cluster
│            │    ├── 2hr.yaml
│            │    └── 24hr.yaml
│            │
│            ├── distributed
│            │    ├── 1gpu.yaml
│            │    ├── 4gpu.yaml
│            │    └── 4gpu4node.yaml
│            │
│            ├── paths
│            │    ├── local.yaml
│            │    └── hpc.yaml
│            │
│            └── baseline.yaml

With task/cifar10/config/baseline.yaml as:

defaults:
  - cluster/2hr@_global_
  - distributed/4gpu@_global_
  - paths/hpc_paths@_global_
  - _self_

train: True
test: True
resumable: True
module: task.cifar10.model.baseline
definition: Baseline
monitor: 'val_acc'
monitor_mode: 'max'
lr: 1e-3
max_epochs: 3
mbatch_size: 32
num_workers: 5

Where task/cifar10/config/baseline.yaml will now include arguments from the following example sub-configurations:

• task/cifar10/config/cluster/2hr.yaml:

  memory: 32GB
  time_limit: '02:00:00'
  venv_path: /path/to/my/venv/bin/activate
  

• task/cifar10/config/distributed/4gpu.yaml:

  num_gpus: 2
  strategy: ddp
  

• task/cifar10/config/paths/hpc.yaml:

  exp_dir: /path/to/my/experiments
  dataset_dir: /path/to/my/dataset
  

See the following documentation for more information:

• https://hydra.cc/docs/1.2/tutorials/basic/your_first_app/defaults/
• https://hydra.cc/docs/1.2/advanced/defaults_list/#composition-order
• https://hydra.cc/docs/1.2/advanced/overriding_packages/

Stages and Trainer

---

In each task directory is a Python module called stages.py, which contains the stages definition. This definition takes an object as input that houses the configuration for a job.

Typically, the following things happen in stages():

• The LightningModule model is imported via the model argument, e.g.,

  from src import importer
  
  Model = importer(definition=args.definition, module=args.module)
  model = Model(**vars(args))
  

  See src.utils.importer for a handy function that imports based on strings.
• A Lightning Trainer instance is created, e.g., trainer = lightning.pytorch.Trainer(...).
• The model is trained using trainer: trainer.fit(model).
• The model is tested using trainer: trainer.test(model).

It handles the training and testing of a model for a task by using a Lightning Trainer.

A helpful wrapper at src/trainer.py exists that passes frequently used and useful callbacks, loggers, and plugins to a Lightning Trainer instance:

from src.dlhpcstarter.trainer import trainer_instance

trainer = trainer_instance(**vars(args))

Place any of the parameters for the trainer detailed at https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-class-api in your configuration file, and they will be passed to the Lightning Trainer instance.

Tying it all together: dlhpcstarter

---

This is an overview of what occurs when the entrypoint dlhpcstarter is executed, this is not necessary to understand to use the package.

dlhpcstarter does the following:

• Gets the command line arguments using argparse, e.g., arguments like this:

  dlhpcstarter --config task.cifar10.config.baseline --task cifar10
  

• Imports the stages definition for the task using src.utils.importer.
• Reads the configuration .yaml and combines it with the command line arguments.
• Submits stages to the cluster manager if args.submit = True or runs stages locally. The command line arguments and the configuration arguments are passed to stages in both cases.

Cluster manager and distributed computing

---

The following arguments are used for distributed computing:

Argument	Description	Default
num_workers	No. of workers per DataLoader & GPU.	1
num_gpus	Number of GPUs per node.	None
num_nodes	Number of nodes (should only be used with submit = True).	1

The following arguments are used to configure a job for a cluster manager (the default cluster manager is SLURM):

Argument	Description	Default
memory	Amount of memory per node.	'16GB'
time_limit	Job time limit.	'02:00:00'
submit	Submit job to the cluster manager.	None
resumable	Resumable training; Automatic resubmission to cluster manager.	None
qos	Quality of service.	None
begin	When to begin the Slurm job, e.g. now+1hour.	None
email	Email for cluster manager notifications.	None
venv_path	Path to ''bin/activate'' of a venv.	None

These can be given as command line arguments:

dlhpcstarter --config task/cifar10/config/baseline --task cifar10 --submit 1 --num-gpus 4 --num-workers 5 --memory 32GB

Or they can be placed in the configuration .yaml file:

num_gpus: 4  # Added.
num_workers: 5  # Added.
memory: '32GB'  # Added.

train: True
test: True
module: task.cifar10.model.baseline
definition: Baseline
monitor: 'val_acc'
monitor_mode: 'max'
lr: 1e-3
max_epochs: 32
mbatch_size: 32
num_workers: 5
exp_dir: /my/experiment/directory
dataset_dir: /my/datasets/directory

And executed with:

dlhpcstarter --config task/cifar10/config/baseline --task cifar10 --submit True

If using a cluster manager, add the path to the bin/activate of your virtual environment:

...
venv_path: /my/env/name/bin/activate
...

Monitoring using Neptune.ai

Simply sign up at https://neptune.ai/ and add your username and API token to your configuration file:

...
neptune_username: my_username
neptune_api_key: df987y94y2q9hoiusadhc9wy9tr82uq408rjw98ch987qwhtr093q4jfi9uwehc987wqhc9qw4uf9w3q4h897324th
...

The PyTorch Lightning Trainer will then automatically upload metrics using the Neptune Logger to Neptune.ai. Once logged in to https://neptune.ai/, you will be able to monitor your task. See here for information about using the online UI: https://docs.neptune.ai/you-should-know/displaying-metadata.

Where all the outputs go: exp_dir

---

The experiments directory is where all your outputs will be saved, including model checkpoints, metric scores. This is also where the cluster manager script, as well as where stderr and stdout are saved.

Note: the trial number also sets the seed number for your experiment.

***Description to be finished.

Repository Wish List

---

• Transfer cluster management over to submitit: https://ai.facebook.com/blog/open-sourcing-submitit-a-lightweight-tool-for-slurm-cluster-computation/
• Add description about how to use https://neptune.ai/.
• Use https://hydra.cc/ instead of argparse (or have the option to use either).
• https://docs.ray.io/en/latest/tune/index.html for hyperparameter optimisation.
• Notebook examples.