Scribd
Upload
5 views40 pages

PyTorch Image Classification Guide

The document provides an overview of binary and multi-class image classification using deep learning with PyTorch, including prerequisites and model architecture. It discusses convolutional neural networks (CNNs), activation functions, and the process of training models with image datasets. Additionally, it covers leveraging pre-trained models and the steps for saving, loading, and making predictions with these models.

Uploaded by

Gianpipi Pacheco
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
5 views40 pages

PyTorch Image Classification Guide

The document provides an overview of binary and multi-class image classification using deep learning with PyTorch, including prerequisites and model architecture. It discusses convolutional neural networks (CNNs), activation functions, and the process of training models with image datasets. Additionally, it covers leveraging pre-trained models and the steps for saving, loading, and making predictions with these models.

Uploaded by

Gianpipi Pacheco
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Binary and multi-

class image
classification
DEEP LEARNING FOR IMAGES WITH PYTORCH

Michal Oleszak
Machine Learning Engineer
What will we learn with PyTorch?

DEEP LEARNING FOR IMAGES WITH PYTORCH


What will we learn with PyTorch?

DEEP LEARNING FOR IMAGES WITH PYTORCH


What will we learn with PyTorch?

DEEP LEARNING FOR IMAGES WITH PYTORCH


What will we learn with PyTorch?

DEEP LEARNING FOR IMAGES WITH PYTORCH


Prerequisites
Convolutional Neural Networks
Model training in PyTorch

Prerequisite course: Intermediate Deep Learning with PyTorch

DEEP LEARNING FOR IMAGES WITH PYTORCH


PyTorch library

DEEP LEARNING FOR IMAGES WITH PYTORCH


PyTorch library

DEEP LEARNING FOR IMAGES WITH PYTORCH


PyTorch library

DEEP LEARNING FOR IMAGES WITH PYTORCH


PyTorch library

DEEP LEARNING FOR IMAGES WITH PYTORCH


Image classification
Binary classification Multi-class classification

Two distinct classes (cats, dogs) Multiple classes (boat, train, car)

Activation function: Sigmoid Activation function: Softmax

Highest probability is the prediction

DEEP LEARNING FOR IMAGES WITH PYTORCH


Convolutional Neural Network model

DEEP LEARNING FOR IMAGES WITH PYTORCH


Convolutional Neural Network model

DEEP LEARNING FOR IMAGES WITH PYTORCH


Convolutional Neural Network model

DEEP LEARNING FOR IMAGES WITH PYTORCH


Convolutional Neural Network model

DEEP LEARNING FOR IMAGES WITH PYTORCH


Convolutional Neural Network model

DEEP LEARNING FOR IMAGES WITH PYTORCH


Datasets: class labels
classes = train_dataset.classes
print(classes)

['cat', 'dog']

print(train_dataset.class_to_idx)

{'cat': 0, 'dog': 1}

from torchvision import datasets


import [Link] as transforms

train_dir = '/data/train'
train_dataset = ImageFolder(root=train_dir,
transform=[Link]())

DEEP LEARNING FOR IMAGES WITH PYTORCH


Binary image classification: convolutional layer
Conv2d() : class BinaryCNN([Link]):

Input: 3 RGB channels (red, green, blue) def __init__(self):


super(BinaryCNN, self).__init__()
Output: 16 channels self.conv1 = nn.Conv2d(3, 16,
kernel_size=3, stride=1, padding=1)
Kernel: 3 x 3 matrix [Link] = [Link]()
[Link] = nn.MaxPool2d(kernel_size=2,
Stride = 1: the kernel moves 1 step
stride=2)

Padding = 1: 1 pixel around the border

ReLU() :
def forward(self, x):
A non-linear activation function

MaxPool2d() :
Kernel: 2×2
return x

Stride: 2 steps

DEEP LEARNING FOR IMAGES WITH PYTORCH


Binary image classification: fully connected layer
Flatten() : class BinaryCNN([Link]):

Tensors flattened into 1-D vector def __init__(self):


super(BinaryCNN, self).__init__()
Linear() : self.conv1 = nn.Conv2d(3, 16,
kernel_size=3, stride=1, padding=1)
Input: feature maps x height x width
[Link] = [Link]()
Output: a single class [Link] = nn.MaxPool2d(kernel_size=2,
stride=2)
Sigmoid() : [Link] = [Link]()
[0,1] self.fc1 = [Link](16 * 112 * 112, 1)
[Link] = [Link]()
def forward(self, x):
x = [Link]([Link](self.conv1(x)))
x = self.fc1([Link](x))
x = [Link](x)]
return x

DEEP LEARNING FOR IMAGES WITH PYTORCH


Multi-class image classification with CNN
class MultiClassCNN([Link]):
def __init__(self, num_classes):
super(MultiClassCNN, self).__init__()
...
[Link] = [Link](16 * 112 * 112, num_classes)
[Link] = [Link](dim=1)
def forward(self, x):
...
x = [Link](x)
return x

DEEP LEARNING FOR IMAGES WITH PYTORCH


Let's practice!
DEEP LEARNING FOR IMAGES WITH PYTORCH
Convolutional layers
for images
DEEP LEARNING FOR IMAGES WITH PYTORCH

Michal Oleszak
Machine Learning Engineer
Convolutional layers for images
Apply convolutional layers to image data
Access and add convolutional layers

Create convolutional blocks

Used to adapt models to a specific task

DEEP LEARNING FOR IMAGES WITH PYTORCH


Conv2d: input channels

Grayscale image: in_channels=1

RGB image (red, green, blue): in_channels=3

Transparency includes alpha channel: in_channels=4

from [Link] import functional


image = [Link]("[Link]")
num_channels = functional.get_image_num_channels(image)
print("Number of channels: ", num_channels)

Number of channels: 3

DEEP LEARNING FOR IMAGES WITH PYTORCH


Conv2d: kernel

Input tensor Kernel Output tensor (feature map)

Kernel (colored in green) moves from left to right, top to bottom of the image1

1 Thevenot, Axel. 2020. A visual and mathematical explanation of the 2D convolution layer.

DEEP LEARNING FOR IMAGES WITH PYTORCH


Kernel sizes

The most common kernel sizes: 3×3 ( Conv2d ) and 2×2 ( MaxPool2d )

Convolution is a dot product of the kernel (green) and the image region (pink)
The sum of the dot product creates a feature map (blue)

DEEP LEARNING FOR IMAGES WITH PYTORCH


Kernel is a filter
Capture image patterns

DEEP LEARNING FOR IMAGES WITH PYTORCH


Conv2d: output channels

Input channel Kernel filters Output channels

The number of output channels determines how many filters are applied

Each output channel corresponds to a distinct filter

A higher number of output channels allows the layer to learn more complex features

Output channel numbers are commonly chosen as powers of 2 (16, 32, 64, 128)
It simplifies the process of combining and dividing channels in subsequent layers

DEEP LEARNING FOR IMAGES WITH PYTORCH


Adding convolutional layers
import torch
import [Link] as nn

class Net([Link]):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3, padding=1)

conv2 = nn.Conv2d(in_channels=16, out_channels=32, kernel_size=3, padding=1)

model = Net()
model.add_module('conv2', conv2)

DEEP LEARNING FOR IMAGES WITH PYTORCH


Accessing convolutional layers
print(model)

Net(
(conv1): Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv2): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)

model.conv2

Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))

DEEP LEARNING FOR IMAGES WITH PYTORCH


Creating convolutional blocks
Stacking convolutional layers in a block with [Link]()

class BinaryImageClassification([Link]):
def __init__(self):
super(BinaryImageClassification, self).__init__()
self.conv_block = [Link](
nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1),
[Link](),
nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1),
[Link](),
nn.MaxPool2d(kernel_size=2, stride=2)
)
def forward(self, x):
x = self.conv_block(x)

DEEP LEARNING FOR IMAGES WITH PYTORCH


Let's practice!
DEEP LEARNING FOR IMAGES WITH PYTORCH
Working with pre-
trained models
DEEP LEARNING FOR IMAGES WITH PYTORCH

Michal Oleszak
Machine Learning Engineer
Leveraging pre-trained models
Training models from scratch:
Long process

Requires lots of data

Pre-trained models - models already trained on a task


Directly reusable on a new task

Require adjustment to the new task (transfer learning)

Steps to leveraging pre-trained models:


Saving & loading models locally

Downloading torchvision models

DEEP LEARNING FOR IMAGES WITH PYTORCH


Saving a complete PyTorch model
[Link]()

Model extension: .pt or .pth

Save model weights with .state_dict()

[Link](model.state_dict(), "[Link]")

DEEP LEARNING FOR IMAGES WITH PYTORCH


Loading PyTorch models
Instantiate a new model

new_model = BinaryCNN()

Load saved parameters

new_model.load_state_dict([Link]('[Link]'))

DEEP LEARNING FOR IMAGES WITH PYTORCH


Downloading torchvision models
Import resnet architecture and weights
from [Link] import (
resnet18, ResNet18_Weights Extract weights
)
Instantiate a model passing it weights

weights = ResNet18_Weights.DEFAULT
Store required data transforms

model = resnet18(weights=weights)
transforms = [Link]()

DEEP LEARNING FOR IMAGES WITH PYTORCH


Prepare new input images
from PIL import Image Load image
Transform image
image = [Link]("[Link]")
image_tensor = transform(image) Reshape image
image_reshaped = image_tensors.unsqueeze(0)

DEEP LEARNING FOR IMAGES WITH PYTORCH


Generating a new prediction
[Link]() Evaluation mode for inference
Disable gradients
with torch.no_grad():
pred = model(image_reshaped).squeeze(0) Pass image to model and remove batch
dimension
pred_cls = [Link](0)
cls_id = pred_cls.argmax().item()
Apply softmax
cls_name = [Link]["categories"][cls_id] Select the highest-probability class and
extract its index
print(cls_name)
Map class index to label
Egyptian cat Print class label

DEEP LEARNING FOR IMAGES WITH PYTORCH


Let's practice
DEEP LEARNING FOR IMAGES WITH PYTORCH

You might also like