How to set the DeepLearning architecture ?

Hi
I have two-class data that each point is an image. The data consist of 200 pos and 800 neg.
The size of each image is 200,200. What is the optimal architecture of the DeepLearning WF ?

Best
Malik

Hi @malik,

this heavily depends on your task, data, data quality etc. Can you clarify?

Best,

Christian

Hi @christian.dietz
The data is as i said images of objects that are so similar and would like to find the right architecture that allow me to capture the differences between the two classes.
The original size of the images are variable that force me to resize it and make it small images. The original size is about 600,600.

Best
Malik

@malik image analysis / object recognition is really, really data dependent and there are tons of factors influencing the success of your training process with DL. One thing I can recommend is that you check our example workflows for DL4J (e.g. face recognition) and start from there… AlexNet, Inception, VGG etc can all be rebuild with KNIME nodes. In case you use the new KNIME Deep Learning integration with Keras / Tensorflow you can also take advantage of pre-trained networks which you simply fine-tune (“transfer learning”). See

https://www.knime.com/nodeguide/analytics/deep-learning/keras/cats-and-dogs/fine-tune-vgg16
https://www.knime.com/nodeguide/analytics/deep-learning/keras/classify-images-using-inceptionv3
https://www.knime.com/nodeguide/analytics/deep-learning/keras/classify-images-using-resnet50

and DL4J

https://www.knime.com/nodeguide/analytics/deep-learning/dl4j/celebrity-detection-using-alex-net

Cheers,

Christian

2 Likes