Clarifications for "Training your own Object Detector"

[chasemaier]

Regarding: https://github.com/justadudewhohacks/tfjs-tiny-yolov2#training-your-own-object-detector

Hi! Thanks for putting this together. I wanted to see if I was able to train a custom model given your existing instructions but I was hoping you might be able to update the documentation with some more explicit instructions if they were readily available.

For example, is the /train folder supposed to contain scripts to help with the training process? I see npm start references a server.js which doesn't seem to be checked in.

Ran into some similar stumbling points trying to reproduce the face training example -- e.g. train.html in that folder references scripts that aren't defined or in the location that is specified as far as I could tell.

I'd love to have some way to have a stand-alone baseline training process that works end-to-end easily from just checking out the code as-is (small set of images/bounding boxes as a sample training set) -- then i could swap out the images and bounding boxes used as input -- especially if there was a script to help with the Computing Box Anchors step. Not sure how far the current state is from that though. I'm having a hard time getting the training process started without wiring up the missing bits.

Another clarification: Will the same process work for sizes < 32 if that value is changed (e.g. to 8)?

[chenqing]

@chasemaier Have you train your dataset success?

[ziyiwang]

hi similarly, the self training module seems to be missing files, any chance we can fill those in ?