http://googleresearch.blogspot.ru/2015/07/deepdream-code-example-for-visualizing.html
interesting introduction panel to deep learning
this was in the shoutbox some months ago
neural talk and walk implementation
will this kind of library be compatible with the automatic importer from vl ?
another one from kyle mcdonald’s feed
text prediction
examples http://rodarmor.com/artnet/
https://github.com/karpathy/char-rnn
for reference: this could be interesting to get started on the general topic: http://ml4a.github.io/
One of the annoying things about machine learning at the moment is that the better known and documented frameworks tend to be built specifically for use with Linux (and to a lesser extent OSX). I came across an interesting discovery regarding Tensorflow models.
TFDeploy allows Python programmers to deploy their Tensorflow models using plain Numpy and Scikit. ^quote:TFDeploy:
Deploy tensorflow graphs for fast evaluation and export to tensorflow-less environments running numpy.
^
For the training aspect of machine learning Tensorflow’s GPU accelerated libraries are really important, but Numpy should be fast enough to run fairly complex ML models.
It is possible to install Numpy and Scipy with Iron Python, and, as demonstrated by MicroD, Python code can be incorporated into vvvv. So is it possible that Tensorflow-trained models could be made to run inside vvvv?
if i understand this correctly, you can also interface with gRTP to the running tensorflow application and going pure c# from there.
just an alternative
meanwhile there is also a .net binding for tensorflow: https://github.com/migueldeicaza/TensorFlowSharp
1 Like
any skillled person interested in slowly starting to bring this fidgets to the mmmmasses ?
Many Crashes later…
Using the TensorflowSharp library and Alpha x64 (its a 64 bit library).
It should be possible to import models to use but I’m a while away from that.
A visual programming language and a machine learning library should be a good fit, what, with the the visible overview of the graph, but a vl library is going to need some thoughtful design choices, There are quite a lot of ways to muck things up at the moment and I haven’t tried anything fancy.
The documentation for the C# wrapper is a little lacking, there are still some bugs and seemingly redundant nodes. Oh and there’s a pretty big memory leak somewhere, its a pretty barebones c# port of the c library; no stablizers.
Like the mainline python, chages happen quite frequently, and do break things.
4 Likes
I have managed to import a model into vl. Again, it was a little tricky to do, but this time because TFOutputs from the model are not explicitly visible and require prior knowledge, placeholders were not intuitive, and tensors themselves require an array construction to cast the value of the TFTensor object into meaningful numbers.
In general, c#/f# code is at times easy to understand but tricky to translate into vl; my vl doesn’t ‘read’ as well it might but its good exercise.
I’ll continue to try out different models now I understand the practice a little better, and training is possible, but keep an eye on a c# implementation of Keras which uses TensorflowSharp as a backend, It will probably be better than raw Tensorflow for creating and training models.
Anyone wanna try this out?
1 Like