You will need to be more specific, a trained neural network is heavily dependent on implementation - for example a trained TensorFlow network is not interchangeable with a trained Caffe network. There are tools to convert between the two, but they're limited in scope.
Tensorflow is preferred but Caffe works too. What I'm looking for are pretrained image recognition, audio recognition, etc algorithms that can be embedded into a Nifi processor, so that popular ML techniques can be accessed with a drag and drop interface :)
obviously there are some interesting details to get this right but sounds like a really cool idea and look forward to seeing this pull request come into the apache nifi community.
4 comments
[ 154 ms ] story [ 348 ms ] threadDo you have a target implementation in mind?
I dont know of an authorative list, but it might be good to start one!
Usually the models have a name, after the project or paper that was used,
To get you started, here's the "Inception v3" model: https://storage.googleapis.com/download.tensorflow.org/model...
Maybe making a list of "Inception v3", "ImageNet" etc, then letting the user select? Most ML engineers know these by their names.
One for caffe: You can download BVLC CaffeNet Model from: http://dl.caffe.berkeleyvision.org/bvlc_reference_caffenet.c...