Tensorflow Support for Go

<p>Anyone have experience using tensorflow in go? Is it still being worked on? I would hate to have to resort to using Python...</p> <hr/>**评论：**<br/><br/>pgaleone: <pre><p>Yes, I&#39;m working with go + tensorflow.</p> <p>The Go APIs are just bindings for the C++ API: you have to manually define the graph, deal with the scopes (that without the context managers of python it&#39;s a pain) and differently from the C++ API the type system is a complete mess.</p> <p>Since every method accepts a tf.Output type as input and the tf.Output has it&#39;s Tensorflow type, you have to correctly define the tf.Output structure with the right (and thus implemented) type. You have no compile time check and thus if you, for instance, pass a int64 instead of an int32 when you create a parameter and the function you&#39;re using has no a registered kernel for the int64, your program will compile but when you execute the graph within a session you&#39;ll get a segfault and debugging is extremely hard.</p> <p>Moreover you don&#39;t have the Variable (that&#39;s a Python class) and thus you can&#39;t train model (no tensorboard support, ecc...). The Go API (and the C++ API too) it&#39;s extremely good to load a model defined and trained using python and use it in a production env.</p> <p>If you want to define and train models: use python. If you want to use trained models or use the Go bindings to execute simple operations within the tensorflow environment (for instance execute a RGB to HSV conversion) you can use Go (but keep in mind that&#39;s hard to define and debug because of lack of the compile time support)</p></pre>loganjspears: <pre><p>That&#39;s interesting. For my use case I don&#39;t care where I train, but would like to use go for predictions. Thank you for sharing your experience. </p> <p>Also do you think Go will ever get first class support?</p></pre>pgaleone: <pre><p>I see very few updates in the repo folder dedicated to go ( <a href="https://github.com/tensorflow/tensorflow/commits/master/tensorflow/go" rel="nofollow">https://github.com/tensorflow/tensorflow/commits/master/tensorflow/go</a> ): the great majority of the commits are just the updates of the generated wrappers (commits made by a bot, really).</p> <p>Thus I guess that&#39;s something the tensorflow developers are not very interested in, they just maintain this automatic wrapper and focus their efforts on the C++ (and python) API.</p> <p>However, I&#39;m writing a wrapper around the Go API (something like an additional abstraction layer that hides the complexity of scopes management and type issues): this won&#39;t add the support for training or any higher level application (that&#39;s python for this), but it can be used to build computational graphs and executes operation into GPUs and then use the results directly in Go (in reality I&#39;m building an image processing library around the tensorflow Go bindings, but the same approach can be used to build a general tensor-processing library and use it like we do in python)</p></pre>chewxy: <pre><p>Self promotion here - if you want to use a Go equivalent, there is <a href="https://github.com/chewxy/gorgonia">gorgonia</a></p> <p>There will be a few API breakings (3 last I checked) in the coming weeks though.</p></pre>loganjspears: <pre><p>Checked it out and it looks like a serious alternative. Thanks for sharing.</p></pre>chewxy: <pre><p>And it did dynamic graphs out of the box too (in the coming weeks it&#39;ll be CUDA accelerated too!)</p></pre>Damien0: <pre><p>Yes, there is an official Go API: <a href="https://godoc.org/github.com/tensorflow/tensorflow/tensorflow/go" rel="nofollow">https://godoc.org/github.com/tensorflow/tensorflow/tensorflow/go</a></p></pre>loganjspears: <pre><p>Thanks! I found getting started info here: <a href="https://github.com/tensorflow/tensorflow/blob/master/tensorflow/go/README.md">https://github.com/tensorflow/tensorflow/blob/master/tensorflow/go/README.md</a> </p> <p>Also there is a short example in the repo. Looks like the API is not stable tho :(</p></pre>