Changing an input image in model view

Hi! Sorry if this has been asked, couldn’t anything on this topic. I am using perceptilabs to develop a CNN for license plate recognition. Is there a way that I can scroll through images in my dataset in the model view? I want to be able to see what each convolutional filter does so I can tweak sizes, strides etc. accordingly, but right now the input is a fixed image that I have no use for (I am classifying 1 class only, so half the images are just garbage data, and the one that perceptilabs chooses to show me is garbage). It would be great to be able to see what the filters do at different feature sizes, with slightly stretched or compressed images. I think I’ve seen this feature in a video where they used an older version (0.11 I think) but I can’t any way to change the image in 0.13.5.

Hi @Dodecahedron and welcome to the forum!

We currently don’t have the functionality to change what image you are viewing in the model view. Note also that everything you see in the model view is untrained, so looking too deeply into how features are extracted in there might not be too helpful outside the standard dimensionality and health check.
We will however get an upgrade to the Evaluate view early next year that lets you do just what you said but for the trained model.
With all of this in mind, do you think it would still be useful to go through images in the model view as well? In that case I would be happy to add it as a feature request to our list :slight_smile:

All the best,
Robert

Hi, Robert, thanks for answering. I understand, I thought that the model view updated when you’ve already trained your model. If so, being able to go through images in the model view will just be an interesting addition but not essential. Being able to do that in the evaluate view is of course important. Good luck with the project.

Hi @Dodecahedron

We seem to be in the middle of a little surge in new users - which is great! Welcome to the forums!

I see @robertl answered your direct question, but if you feel like sharing more about how you were structuring your model I’m sure there are now quite a few others here he could add some useful input - even if you can’t see exactly what you want right now. For example - I’ve seen license plate NN based on CNN and LSTM… maybe a transformer is the way to go now? There was a post recently on how to do something with a Vision Transformer that might be useful to you.

Keep in touch!

UPDATE

Here’s the paper I was referring to : https://arxiv.org/pdf/1601.05610.pdf

1 Like