[WIP] Fix feature extractor examples #1086
Conversation
| @@ -71,7 +71,7 @@ function setupButtons() { | |||
| // When the Dog button is pressed, add the current frame | |||
| // from the video with a label of "dog" to the classifier | |||
| document.querySelector("#addSample").addEventListener("click", function() { | |||
| regressor.addImage(slider.value); | |||
| regressor.addImage(parseFloat(slider.value)); | |||
There was a problem hiding this comment.
Should we handle this internally in ml5? I don't feel strongly but we could do this conversion there since passing in a string label is not compatible with regression!
There was a problem hiding this comment.
I was wondering about this too. The best logic I could consider is changing any values that seem numeric into a numeric value instead of a string label.
Are string labels really not compatible? It seems like there might be some code in the model for associating string levels to some of the regression outputs. I'll open an issue to come back to this!
There was a problem hiding this comment.
Great question! The concept of regression is quite broad, but in the case of ml5 and the way ml5.neuralNetwork() was designed, the only valid outputs are numeric. If it's a string then by definition (at least in the case of ml5) it's a classification. I'm open to rethinking this of course!
|
Looks like there may be some possibly valid test failures here. I'll take a closer look to see what may be happening. |
|
Test failures here we were related to #1093. Now that that has been temporarily addressed, I feel comfortable merging this in! |
Our FeatureExtractor examples were breaking with the following error when I'd try to add samples for the model to use for training:
#191 introduced some logic that would make videos avoid resizing. Marking this as WIP because I'm not sure why the videos were ignored for resizing in that PR.
I also noticed that it seems that we get some odd behavior when we pass in numerical values as strings into the model. Our JS example passed in strings that broke the mapping the example was expecting .
After the changes, both the p5.js and JS versions of FeatureExtractor_Image_Classification of FeatureExtractor_Image_Regression seem to be working well!