@liuliu, can libccv be used to make your own classifiers, like in Caffe, you can use a pre existing model and fine tune it to a new dataset? Can I use libccv to train a classifier of clouds vs birds, or can it only be used for the 1000 standard category recognition from ImageNet?
It includes a training program (./bin/image-net.c), however, it requires a bit of work to get it started (you need to initialize a new convnet with the pre-trained parameters, and then fine-tuning the whole network). It is on my todo list to make a more friendly training program that for these simple tasks, you should be able to do with CPU easily.
I remember when you first announced this project some years back. You've done such an incredible job! :)
Since you mention distribution of pre-trained models by participants in ImageNet 2014, I have to wonder if there would be a market for distributing specialized classifiers like open source software. Eg, an npm-like website you might be able to find a "dog breed classifier", "car make/model classifier", etc. If so, what would it take to make it happen?
Cool, I'd love to help out with adding that tutorial to libccv, do you and the libccv team use a communication tool like trello or skype to communicate? Or email?
That's when learning rate changed to a smaller number. The graph mainly shows that with different initialization scheme, the network starts descending initially faster.
