Pytorch Tutorial

Chongruo Wu
Agenda

1. Popular Frameworks

2. Pytorch, Basics

3. Helpful skills
Popular Deep Learning Frameworks

Gluon: new MXNet interface to accelerate research

Popular Deep Learning Frameworks

Imperative: Imperative-style programs perform computation as you run them

Symbolic: define the function first, then compile them

Gluon: new MXNet interface to accelerate research

Popular Deep Learning Frameworks

Gluon: new MXNet interface to accelerate research

Popular Deep Learning Frameworks

lua
C++, Python,
R, Julia, Perl
Scala

Stanford cs231n.
MxNet Tutorial, CVPR 2017
MxNet Tutorial, CVPR 2017
MxNet Tutorial, CVPR 2017
Stanford cs231n.
MxNet Tutorial, CVPR 2017
MxNet Tutorial, CVPR 2017
MxNet Online Document, https://goo.gl/UZ2byD
Stanford cs231n.
Stanford cs231n.
Pytorch
Stanford cs231n.
Stanford cs231n.
Pytorch Tensors

https://transfer.d2.mpi-inf.mpg.de/rshetty/hlcv/Pytorch_tutorial.pdf
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Variable
The autograd package provides automatic differentiation for all operations on Tensors.

“ autograd.Variable is the central class of the package. It wraps a Tensor, and supports
nearly all of operations defined on it.

Once you finish your computation you can call .backward() and have all the gradients
computed automatically. “

Pytorch Tutorial. www.pytorch.org

Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Stanford cs231n.
Module, single layer

Other layers:
Dropout, Linear,
Normalization Layer
Module, network

Pytorch, zero to all. HKUST

Module, sub-network

http://book.paddlepaddle.org/03.image_classification/
Module, sub-network

https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
Module
Stanford cs231n.
When starting a new project

1. Data preparation ( processing, format )

2. Model Design ( pretrained, design your own Model)

3. Training Strategy
Train a simple Network

Stanford cs231n.
Train a simple Network

1. Forward: compute output of each layer

2. Backward: compute gradient
3. Update: update the parameters with computed gradient
Train a simple Network

Stanford cs231n.
Train a simple Network

Stanford cs231n.
Train a simple Network

Stanford cs231n.
Train a simple Network

Stanford cs231n.
Train a simple Network

Stanford cs231n.
Train a simple Network

Stanford cs231n.
MNIST example
MNIST example
MNIST example

https://goo.gl/mQEw15
MNIST example

Data Loading

https://goo.gl/mQEw15
MNIST example

Define Network

https://goo.gl/mQEw15
MNIST example

Training

https://goo.gl/mQEw15
MNIST example

Inference

eval() mode:

*Dropout Layer
*Batchnorm Layer

https://goo.gl/mQEw15
When starting a new project

1. Data preparation ( processing, format )

2. Model Design ( pretrained, design your own model)

3. Training Strategy
Data Loading
Data Loading

Pytorch, zero to all. HKUST

Data Loading

Pytorch, zero to all. HKUST

Data Loading

https://goo.gl/mQEw15
Data Loading

Pytorch, zero to all. HKUST

Data Loading

Pytorch, zero to all. HKUST

Data Loading

Stanford cs231n.
Data Loading

Pytorch, zero to all. HKUST

Data Processing

https://goo.gl/mQEw15
Data Processing
Pix2pix Code

https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
When starting a new project

1. Data preparation ( processing, format )

2. Model Design ( pretrained, design your own model)

3. Training Strategy ( learning rate )

Learning Rate Scheduler

Stanford cs231n.
Learning Rate Scheduler

torch.optim.lr_scheuler
● StepLR: LR is delayed by gamma every step_size epochs
● MultiStepLR: LR is delayed by gamma once the number of epoch reaches milestones.
● ExponentialLR
● CosineAnnealingLR
● ReduceLROnPlateau

https://github.com/Jiaming-Liu/pytorch-lr-scheduler
http://pytorch.org/docs/master/optim.html#how-to-adjust-learning-rate
Pretrained Model
Load Model

args.resume is the path to the trained model

Define the model

before loading parameters

https://gist.github.com/panovr/2977d9f26866b05583b0c40d88a315bf
Weights Initialization

net.apply( weights_init_normal)

https://goo.gl/bqeW1K
Weights Initialization

https://goo.gl/bqeW1K
Hooks

It is used to inspect or modify the output and grad of a layer

Need to write register function

Hooks ( Forward )

http://pytorch.org/tutorials/beginner/former_torchies/nn_tutorial.html
Hooks ( Backward )

http://pytorch.org/tutorials/beginner/former_torchies/nn_tutorial.html
Cudnn.benchmark flag

https://goo.gl/5gzj8F
Visualization, Pytorch Visdom
https://github.com/facebookresearch/visdom
Visualization, TensorBoard

https://www.tensorflow.org/get_started/summaries_and_tensorboard
Other Resources
Official examples. https://goo.gl/Q6Z2k8
Other Resources
Official documents. https://goo.gl/gecKC4
Other Resources
Pix2pix code https://github.com/phillipi/pix2pix
Other Resources
Pytorch, Zero to All (HKUST) https://goo.gl/S3vEUN
Thank You