Webinar: Break into Deep Learning for Image Data without Code - May 24, 2022

Resources Events
,
1

Join me for the webinar ‘Break into Deep Learning for Image Data without Code’ on Tuesday, May 24, 2022 at 5 PM - 6 PM UTC +2 (Berlin) which is 10 AM - 11 PM UTC -5 (Chicago).

In this webinar, you will get a basic introduction to the representation of images in Computer Vision. Next, you will learn about the key theoretical concepts behind Convolutional Neural Networks and their application for the classification of images. Finally, we will see how the performance of a simple CNN can be substantially improved thanks to Transfer Learning.

Throughout the webinar, we will rely on the KNIME Deep Learning - Keras Integration, which allows you to define, train and deploy your deep learning models in a fully codeless fashion.

All the solutions presented in this webinar are available for public use on the KNIME Hub.

3 Likes
2

Waiting for the webinar, you might find useful to…

See you there! :smiley:

3 Likes
3

Hi all,

Thanks for attending the Webinar yesterday. Was great to see so many of you!

Please find a few useful resources here:

Slides → Break into Deep Learning for Image Data without Code

Recordings → Break into Deep Learning for Image Data without Code

Workflows → Classifying Images of Cats and Dogs (simple CNN - training), Classifying Images of Cats and Dogs (simple CNN - deployment), Classifying Images of Cats and Dog with Transfer Learning (with Keras Freeze Layers node), and Classifying Images of Cats and Dog with Transfer Learning (fast execution)

Documentation → KNIME Deep Learning Integration Installation Guide

Additional resources (readings) → Introduction to Convolutional Neural Networks and Computer Vision and A Beginners Guide to Codeless Deep Learning: MNIST Digit classification.

Now it’s your turn
I’d like to challenge you to improve the performance of the simple CNN for image classification (without Transfer Learning) that I illustrated during the webinar.

Exercise → Classifying Images of Cats and Dogs - Training (Exercise)

Hint: Play around with hyperparameter optimization and/or the network architecture.

Share your solution in this thread and on the KNIME Hub! :smiley:

2 Likes
4

Ciao Roberto

Thanks for the great webinar and for following it up with this nice challenge :grinning: :+1: !

A have a first question :thinking:

image
image1027×296 56.9 KB

The dataset is made of 25,000 images but the workflow has limited in principle the number of images to use to only 4,000 which eventually is split into 80% training & 20% test. Shall this number remain the same for the challenge or can we use the whole amount of images whilst preserving the same training-testing 80%-20% ratio ?

Thanks in advance for your reply !

Best
Ael

1 Like
5

Hi @aworker, many thanks for your question :smiley:. Ideally, you should work only with the network architecture and/or hyperparameters since we focus mostly on those aspects yesterday.

Of course, in real life applications, increasing the amount of image data is definitely a feasible (and often necessary) alternative.

Have fun! :smiley:

Roberto

1 Like
6

Thanks Roberto :grinning: :+1:!

In bocca al lupo a tutti i partecipanti :crossed_fingers:

Cheers

1 Like