The Hitchiker's Guide to PyTorch

Grokking PyTorch

PyTorch is a flexible deep learning framework that allows automatic differentiation through dynamic neural networks (i.e., networks that utilise dynamic control flow like if statements and while loops). It supports GPU acceleration, distributed training, various optimisations, and plenty more neat features. These are some notes on how I think about using PyTorch, and don't encompass all parts of the library or every best practice, but may be helpful to others.

Neural networks are a subclass of computation graphs. Computation graphs receive input data, and data is routed to and possibly transformed by nodes which perform processing on the data. In deep learning, the neurons (nodes) in neural networks typically transform data with parameters and differentiable functions, such that the parameters can be optimised to minimise a loss via gradient descent. More broadly, the functions can be stochastic, and the structure of the graph can be dynamic. So while neural networks may be a good fit for dataflow programming, PyTorch's API has instead centred around imperative programming, which is a more common way for thinking about programs. This makes it easier to read code and reason about complex programs, without necessarily sacrificing much performance; PyTorch is actually pretty fast, with plenty of optimisations that you can safely forget about as an end user (but you can dig in if you really want to).

The rest of this document, based on the official MNIST example, is about grokking PyTorch, and should only be looked at after the official beginner tutorials. For readability, the code is presented in chunks interspersed with comments, and hence not separated into different functions/files as it would normally be for clean, modular code.

Imports

import argparse
import os
import torch
from torch import nn, optim
from torch.nn import functional as F
from torch.utils.data import DataLoader
from torchvision import datasets, transforms

These are pretty standard imports, with the exception of the torchvision modules that are used for computer vision problems in particular.

Setup

parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=64, metavar='N',
                    help='input batch size for training (default: 64)')
parser.add_argument('--epochs', type=int, default=10, metavar='N',
                    help='number of epochs to train (default: 10)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
                    help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
                    help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
                    help='random seed (default: 1)')
parser.add_argument('--save-interval', type=int, default=10, metavar='N',
                    help='how many batches to wait before checkpointing')
parser.add_argument('--resume', action='store_true', default=False,
                    help='resume training from checkpoint')
args = parser.parse_args()

use_cuda = torch.cuda.is_available() and not args.no_cuda
device = torch.device('cuda' if use_cuda else 'cpu')
torch.manual_seed(args.seed)
if use_cuda:
  torch.cuda.manual_seed(args.seed)

argparse is a standard way of dealing with command-line arguments in Python.

A good way to write device-agnostic code (benefitting from GPU acceleration when available but falling back to CPU when not) is to pick and save the appropriate torch.device, which can be used to determine where tensors should be stored. See the official docs for more tips on device-agnostic code. The PyTorch way is to put device placement under the control of the user, which may seem a nuisance for simple examples, but makes it much easier to work out where tensors are - which is useful for a) debugging and b) making efficient use of devices manually.

For repeatable experiments, it is necessary to set random seeds for anything that uses random number generation (including random or numpy if those are used too). Note that cuDNN uses nondeterministic algorithms, and it can be disabled using torch.backends.cudnn.enabled = False.

Data

data_path = os.path.join(os.path.expanduser('~'), '.torch', 'datasets', 'mnist')
train_data = datasets.MNIST(data_path, train=True, download=True,
                            transform=transforms.Compose([
                              transforms.ToTensor(),
                              transforms.Normalize((0.1307,), (0.3081,))]))
test_data = datasets.MNIST(data_path, train=False, transform=transforms.Compose([
                             transforms.ToTensor(),
                             transforms.Normalize((0.1307,), (0.3081,))]))

train_loader = DataLoader(train_data, batch_size=args.batch_size,
                          shuffle=True, num_workers=4, pin_memory=True)
test_loader = DataLoader(test_data, batch_size=args.batch_size,
                         num_workers=4, pin_memory=True)

Since torchvision models get stored under ~/.torch/models/, I like to store torchvision datasets under ~/.torch/datasets. This is my own convention, but makes it easier if you have lots of projects that depend on MNIST, CIFAR-10 etc. In general it's worth keeping datasets separately to code if you end up reusing several datasets.

torchvision.transforms contains lots of handy transformations for single images, such as cropping and normalisation.

DataLoader contains many options, but beyond batch_size and shuffle, num_workers and pin_memory are worth knowing for efficiency. num_workers > 0 uses subprocesses to asynchronously load data, rather than making the main process block on this. The typical use-case is when loading data (e.g. images) from disk and maybe transforming them too - this can be done in parallel with the network processing the data. You will want to tune the amount to a) minimise the number of workers and hence CPU and RAM usage (each worker loads a separate batch, not individual samples within a batch) b) minimise the time the network is waiting for data. pin_memory uses pinned memory (as opposed to paged memory) to speed up any RAM to GPU transfers (and does nothing for CPU-only code).

Model

class Net(nn.Module):
  def __init__(self):
    super(Net, self).__init__()
    self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
    self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
    self.conv2_drop = nn.Dropout2d()
    self.fc1 = nn.Linear(320, 50)
    self.fc2 = nn.Linear(50, 10)

  def forward(self, x):
    x = F.relu(F.max_pool2d(self.conv1(x), 2))
    x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
    x = x.view(-1, 320)
    x = F.relu(self.fc1(x))
    x = self.fc2(x)
    return F.log_softmax(x, dim=1)

model = Net().to(device)
optimiser = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)

if args.resume:
  model.load_state_dict(torch.load('model.pth'))
  optimiser.load_state_dict(torch.load('optimiser.pth'))

Network initialisation typically includes member variables, layers which contain trainable parameters, and maybe separate trainable parameters and non-trainable buffers. The forward pass then uses these in conjunction with functions from F that are purely functional (don't contain parameters). Some people prefer to have completely functional networks (e.g., keeping parameters separately and using F.conv2d instead of nn.Conv2d) or networks completely made of layers (e.g., nn.ReLU instead of F.relu).

.to(device) is a convenient way of sending the device parameters (and buffers) to GPU if device is set to GPU, doing nothing otherwise (when device is set to CPU). It's important to transfer the network parameters to the appropriate device before passing them to the optimiser, otherwise the optimiser will not be keeping track of the parameters properly!

Both neural networks (nn.Module) and optimisers (optim.Optimizer) have the ability to save and load their internal state, and .load_state_dict(state_dict) is the recommended way to do so - you'll want to reload the state of both to resume training from previously saved state dictionaries. Saving the entire object can be error prone. If you have saved tensors on GPU and want to load them on CPU or another GPU, the easiest way is to directly load them onto CPU using the map_location option, e.g., torch.load('model.pth', map_location='cpu').

Some points of note not shown here are that the forward pass can make use of control flow (e.g., a member variable or even the data itself can determine the execution of an if statement. It is also perfectly valid to print tensors in the middle, making debugging much easier. Finally, the forward pass can make use of multiple arguments. A short snippet (not tied to any sensible idea) to illustrate this is below:

def forward(self, x, hx, drop=False):
  hx2 = self.rnn(x, hx)
  print(hx.mean().item(), hx.var().item())
  if hx.max.item() > 10 or self.can_drop and drop:
    return hx
  else:
    return hx2

Training

model.train()
train_losses = []

for i, (data, target) in enumerate(train_loader):
  data = data.to(device=device, non_blocking=True)
  target = target.to(device=device, non_blocking=True)
  optimiser.zero_grad()
  output = model(data)
  loss = F.nll_loss(output, target)
  loss.backward()
  train_losses.append(loss.item())
  optimiser.step()

  if i % 10 == 0:
    print(i, loss.item())
    torch.save(model.state_dict(), 'model.pth')
    torch.save(optimiser.state_dict(), 'optimiser.pth')
    torch.save(train_losses, 'train_losses.pth')

Network modules are by default set to training mode - which impacts the way some modules work, most noticeably dropout and batch normalisation. It's best to set this manually anyway with .train(), which propagates the training flag down all children modules.

Here the .to() method not only takes the device, but also sets non_blocking=True, which enables asynchronous data copies to GPU from pinned memory, hence allowing the CPU to keep operating during the transfer; non_blocking=True is simply a no-op otherwise.

Before collecting a new set of gradients with loss.backward() and doing backpropagation with optimiser.step(), it's necessary to manually zero the gradients of the parameters being optimised with optimiser.zero_grad(). By default, PyTorch accumulates gradients, which is very handy when you don't have enough resources to calculate all the gradients you need in one go.

PyTorch uses a tape-based automatic gradient (autograd) system - it collects which operations were done on tensors in order, and then replays them backwards to do reverse-mode differentiation. This is why it is super flexible and allows arbitrary computation graphs. If none of the tensors require gradients (you'd have to set requires_grad=True when constructing a tensor for this - as by default new tensors don't require gradients) then no graph is stored! However, networks tend to have parameters that require gradients, so any computation done from the output of a network will be stored in the graph. So if you want to store data resulting from this, you'll need to manually disable gradients or, more commonly, store it as a Python number (via .item() on a PyTorch scalar) or numpy array. See the official docs for more on autograd.

One way to cut the computation graph is to use .detach(), which you may use when passing on a hidden state when training RNNs with truncated backpropagation-through-time. It's also handy when differentiating a loss where one component is the output of another network, but this other network shouldn't be optimised with respect to the loss - examples include training a discriminator from a generator's outputs in GAN training, or training the policy of an actor-critic algorithm using the value function as a baseline (e.g. A2C). Another technique for preventing gradient calculations that is efficient in GAN training (training the generator from the discriminator) and typical in fine-tuning is to loop through a networks parameters and set param.requires_grad = False.

Apart from logging results in the console/in a log file, it's important to checkpoint model parameters (and optimiser state) just in case. You can also use torch.save() to save normal Python objects, but other standard choices include the built-in pickle. Note that here that if the parameters are on GPU they will be saved on GPU, but can be recovered on CPU using map_location, as discussed previously; casting the model parameters to CPU before using .state_dict() would therefore require casting back to GPU before resuming the rest of your code.

Testing

model.eval()
test_loss, correct = 0, 0

with torch.inference_mode():
  for data, target in test_loader:
    data = data.to(device=device, non_blocking=True)
    target = target.to(device=device, non_blocking=True)
    output = model(data)
    test_loss += F.nll_loss(output, target, reduction='sum').item()
    pred = output.argmax(1, keepdim=True)
    correct += pred.eq(target.view_as(pred)).sum().item()

test_loss /= len(test_data)
acc = correct / len(test_data)
print(acc, test_loss)

In response to .train() earlier, networks should explicitly be set to evaluation mode using .eval().

As mentioned previously, the computation graph would normally be made when using a network. By using the inference_mode context manager via with torch.inference_mode() this is prevented from happening. Previous versions of PyTorch only had the no_grad context manager, which can be used to disable gradient computations within a context, but allow the variables from the context to be used as part of a computation graph. For full details of what this means and what it entails, see the official docs on locally disabling gradients.

Extra

This is an extra section just to add a few useful asides.

Memory problems? Check the official docs for tips.

CUDA errors? They are a pain to debug, and are usually a logic problem that would come up with a more intelligible error message on CPU. It's best to be able to easily switch between CPU and GPU if you are planning on using the GPU. A more general development tip is to be able to set up your code so that it's possible to run through all of the logic quickly to check it before launching a proper job - examples would be preparing a small/synthetic dataset, running one train + test epoch, etc. If it is a CUDA error, or you really can't switch to CPU, setting CUDA_LAUNCH_BLOCKING=1 will make CUDA kernel launches synchronous and as a result provide better error messages.

A note for torch.multiprocessing, or even just running multiple PyTorch scripts at once. Because PyTorch uses multithreaded BLAS libraries to speed up linear algebra computations on CPU, it'll typically use several cores. If you want to run several things at once, with multiprocessing or several scripts, it may be useful to manually reduce these by setting the environment variable OMP_NUM_THREADS to 1 or another small number - this reduces the chance of CPU thrashing. The official docs have some other notes for multiprocessing in particular.

Owner
Kai Arulkumaran
Researcher, programmer, DJ, transhumanist.
Kai Arulkumaran
Similar Resources

PyTorch tutorials and best practices.

Effective PyTorch Table of Contents Part I: PyTorch Fundamentals PyTorch basics Encapsulate your model with Modules Broadcasting the good and the ugly

Nov 28, 2022

A scalable template for PyTorch projects, with examples in Image Segmentation, Object classification, GANs and Reinforcement Learning.

A scalable template for PyTorch projects, with examples in Image Segmentation, Object classification, GANs and Reinforcement Learning.

PyTorch Project Template is being sponsored by the following tool; please help to support us by taking a look and signing up to a free trial PyTorch P

Nov 27, 2022

Some example scripts on pytorch

pytorch-practice Some example scripts on pytorch CONLL 2000 Chunking task Uses BiLSTM CRF loss with char CNN embeddings. To run use: cd data/conll2000

Nov 29, 2022

Example of network fine-tuning in pytorch for the kaggle competition Dogs vs. Cats Redux: Kernels Edition

Example of network fine-tuning in pytorch for the kaggle competition Dogs vs. Cats Redux: Kernels Edition Currently

Sep 22, 2022

ConvNet training using pytorch

Convolutional networks using PyTorch This is a complete training example for Deep Convolutional Networks on various datasets (ImageNet, Cifar10, Cifar

Nov 4, 2022

simple generative adversarial network (GAN) using PyTorch

Generative Adversarial Networks (GANs) in PyTorch Running Run the sample code by typing: ./gan_pytorch.py ...and you'll train two nets to battle it o

Jun 14, 2020

Torch Containers simplified in PyTorch

Torch Containers simplified in PyTorch

pytorch-containers This repository aims to help former Torchies more seamlessly transition to the "Containerless" world of PyTorch by providing a list

Apr 25, 2022

The Hitchiker's Guide to PyTorch

The Hitchiker's Guide to PyTorch

Nov 2, 2022

Short, introductory guide for the Python programming language

Short, introductory guide for the Python programming language

100 Page Python Intro This book is a short, introductory guide for the Python programming language.

Nov 26, 2022

Guide: Finetune GPT2-XL (1.5 Billion Parameters) and GPT-NEO (2.7 B) on a single 16 GB VRAM V100 Google Cloud instance with Huggingface Transformers using DeepSpeed

Guide: Finetune GPT2-XL (1.5 Billion Parameters) and GPT-NEO (2.7 Billion Parameters) on a single 16 GB VRAM V100 Google Cloud instance with Huggingfa

Dec 1, 2022

The guide to tackle with the Text Summarization

The guide to tackle with the Text Summarization

The guide to tackle with the Text Summarization

Dec 1, 2022

Data Visualization Guide for Presentations, Reports, and Dashboards

Data Visualization Guide for Presentations, Reports, and Dashboards

This is a highly practical and example-based guide on visually representing data in reports and dashboards.

Nov 25, 2022

A complete guide to start and improve in machine learning (ML)

 A complete guide to start and improve in machine learning (ML)

A complete guide to start and improve in machine learning (ML), artificial intelligence (AI) in 2021 without ANY background in the field and stay up-to-date with the latest news and state-of-the-art techniques!

Dec 3, 2022

Paddle-RLBooks is a reinforcement learning code study guide based on pure PaddlePaddle.

Paddle-RLBooks is a reinforcement learning code study guide based on pure PaddlePaddle.

Paddle-RLBooks Welcome to Paddle-RLBooks which is a reinforcement learning code study guide based on pure PaddlePaddle. 欢迎来到Paddle-RLBooks,该仓库主要是针对强化学

Nov 22, 2022

An audio guide for destroying oracles in Destiny's Vault of Glass raid

An audio guide for destroying oracles in Destiny's Vault of Glass raid

prophet An audio guide for destroying oracles in Destiny's Vault of Glass raid. This project allows you to make any encounter with oracles without hav

Sep 15, 2022

Guide & Examples to create deeplearning gstreamer plugins and use them in your pipeline

Guide & Examples to create deeplearning gstreamer plugins and use them in your pipeline

upai-gst-dl-plugins Guide & Examples to create deeplearning gstreamer plugins and use them in your pipeline Introduction Thanks to the work done by @j

Apr 20, 2022

Step by Step Guide To Install Discord Py Master Branch on Replit

Step by Step Guide To Install Discord Py Master Branch on Replit

Guide to Install Discord Py Master Branch on Replit Step 1 Create an empty repl on replit Step 2 Add this Basic Code to the file main.py so as to chec

Nov 18, 2022

step by step guide for beginners for getting started with open source

step by step guide for beginners for getting started with open source

Step-by-Step Guide for beginners for getting started with Open-Source Here The Contribution Begins 💻 If you are a beginner then this repository is fo

Nov 21, 2022

A guide for using Bootstrap 5 classes in Dash Bootstrap Components V1

A guide for using Bootstrap 5 classes in Dash Bootstrap Components V1

dash-bootstrap-cheatsheet This handy interactive cheatsheet makes it easy to use the Bootstrap 5 classes with your Dash app made with the latest versi

Nov 2, 2022
Comments
  • Correction?

    Correction?

    Thanks for writing this, Kai!

    I believe:

    you'd have to set requires_grad=True when constructing a tensor for this

    should be:

    you'd have to set requires_grad=False when constructing a tensor for this

  • Python has stoped working

    Python has stoped working

    This line of my code correct += (predicted[0] == labels[0]).sum().item() cause Python has stoped working.

    6. Calculate Accuracy

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

    When my code reach to this line, the widows kill the python kernel. image

    print('Accuracy of the Network is : %d' %(100*correct/total))

A set of examples around pytorch in Vision, Text, Reinforcement Learning, etc.

PyTorch Examples WARNING: if you fork this repo, github actions will run daily on it. To disable this, go to /examples/settings/actions and Disable Ac

Nov 29, 2022
Deep Learning (with PyTorch)

Deep Learning (with PyTorch) This notebook repository now has a companion website, where all the course material can be found in video and textual for

Dec 2, 2022
Open source guides/codes for mastering deep learning to deploying deep learning in production in PyTorch, Python, C++ and more.
Open source guides/codes for mastering deep learning to deploying deep learning in production in PyTorch, Python, C++ and more.

Deep Learning Materials by Deep Learning Wizard Start Learning Now Please head to www.deeplearningwizard.com to start learning! It is mobile/tablet fr

Nov 21, 2022
C++ Implementation of PyTorch Tutorials for Everyone
C++ Implementation of PyTorch Tutorials for Everyone

C++ Implementation of PyTorch Tutorials for Everyone OS (Compiler)\LibTorch 1.9.0 macOS (clang 10.0, 11.0, 12.0) Linux (gcc 8, 9, 10, 11) Windows (msv

Dec 3, 2022
Simple examples to introduce PyTorch

This repository introduces the fundamental concepts of PyTorch through self-contained examples. At its core, PyTorch provides two main features: An n-

Dec 4, 2022
Minimal tutorials for PyTorch

Minimal tutorials for PyTorch adapted from Alec Radford's Theano tutorials. Tensor multiplication Linear Regression Logistic Regression Neural Network

Oct 25, 2022
PyTorch Tutorial for Deep Learning Researchers

This repository provides tutorial code for deep learning researchers to learn PyTorch. In the tutorial, most of the models were implemented with less

Dec 2, 2022
PyTorch Implementation of Fully Convolutional Networks. (Training code to reproduce the original result is available.)
PyTorch Implementation of Fully Convolutional Networks. (Training code to reproduce the original result is available.)

pytorch-fcn PyTorch implementation of Fully Convolutional Networks. Requirements pytorch >= 0.2.0 torchvision >= 0.1.8 fcn >= 6.1.5 Pillow scipy tqdm

Dec 1, 2022
Simple PyTorch Tutorials Zero to ALL!

PyTorchZeroToAll Quick 3~4 day lecture materials for HKUST students. Video Lectures: (RNN TBA) Youtube Bilibili Slides Lecture Slides @GoogleDrive If

Nov 25, 2022
Pytorch implementations of various Deep NLP models in cs-224n(Stanford Univ)

DeepNLP-models-Pytorch Pytorch implementations of various Deep NLP models in cs-224n(Stanford Univ: NLP with Deep Learning) This is not for Pytorch be

Dec 2, 2022