Deep Learning

Is it possible to use deeplearning4j to train a network from start and/or do transfer learning and retrain the last couple of layers (not only the output layer)?

Hi!

I have been doing image classification using this framefowk for some time, and are going to move on to object detection (boxing in the object position also).
For the output layer, is the only choice to use Yolo2OutputLayer? And if so, in the notes in the java class I can read;
"Doesn't currently support simultaneous training on both detection and classification"...

Is it possible to use deeplearning4j to train a network from start and/or do transfer learning and retrain the last couple of layers (not only the output layer)?

Best Regards

You already invited:

2 Answers

Evren

Upvotes from:

you can do transfer learning
see https://deeplearning4j.org/doc ... rning
this is an example of transfer learning
https://github.com/deeplearnin ... vgg16

Fredrik

Upvotes from:

that's the only object detection achitecture we have built in - but it's one of the best options out there for object detection IMO
that's a reference to the training procedure used in the yolo 2 paper (link in javadoc), you should read it if you want to understand what that's about(tl;dr they trained on classification and object detection datasets simultaneously)
yes, of course - the yolo house number of example does this (with a pretrained model from the zoo, but it's the same thing)

If you wanna answer this question please Login or Register

Is it possible to use deeplearning4j to train a network from start and/or do transfer learning and retrain the last couple of layers (not only the output layer)?

2 Answers

The Author

Similar Questions

This question condition

Is it possible to use deeplearning4j to train a network from start and/or do transfer learning and retrain the last couple of layers (not only the output layer)?

Links

2 Answers

The Author

Similar Questions

This question condition