Say I have two saved models one from tensorflow 1.8 and the other from tensorflow 2.2. Serving both of those could run into compatibility issues.
Would it be possible to serve both of those in the same tensorflow/serving binary ?
My intuition suggests NO one cannot, at least not easily.
I am not an expert in bazel files but I presume compiling tensorflow/serving needs to build and link the tensorflow core lib. I am not sure whether one could link together two different versions of the tensorflow core library together.
I guess one could compile the tensorflow/serving binary in two different release points 1.8.0 and also 2.2.0 and deploy both of those binaries in your infrastructure separately. Then one needs to manage at the model discovery layer and request routing layer about which model needs to be loaded in which tensorflow/serving binary and also which predict request should talk to which tensorflow/serving endpoint.
I'm definitely not an expert on the deep inner workings of TensorFlow, so take this with a grain of salt. But I think what you want to do may actually be pretty easy.
My very approximate (and possibly completely incorrect) understanding is that the TensorFlow APIs are a sort of wrapper that creates a graph representing whatever computation you'd like to do, and that the compiled graph is cross-compatible between at least some versions, even if the APIs used to create and manipulate it aren't.
Empirically, I've been able to take models built with TensorFlow 1.15.x and put them into TensorFlow Serving on 2.3.0 with absolutely no problems at all.
Related
TensorFlow Federated (TFF) is an open-source framework for ML and other computations on decentralized data.
As per Stack overflow link
TFF only provides a simulation environment for use in Federated
Learning (FL) research. There is not yet a "real world" FL deployment
platform.
But, tensorFlow release history shows that now there are many release versions for TF 2.x as well.
https://github.com/tensorflow/federated/releases
Can anybody comment, if TFF is still or simulation environment or can be used as "real world" FL deployment platform?
At this time, TensorFlow Federated does not have out-of-the-box support for what would generally be considered production federated learning. A production-level deployment runtime still needs to be built.
For different flavors of federated learning this maybe easier or harder.
It may be possible to create a system for cross-silo FL using the executor components already available in TensorFlow Federated. Particularly it maybe possible to extend and build something on top of the remote execution stack (e.g. tff.framework.RemoteExecutor)
However for cross-device FL it maybe signifcantly more work, as there are no integrations or components for deploying and execution computations on mobile operating systems (iOS, Android, etc) yet.
I want to use Tensorflow 1.4 for my ML modeling needs. My use case requires:
Training the model on GPU <--- I know how to do this with TF
Deploying the trained model on an ordinary box - as an .exe on CPU running Windows (for inference) <----
I don't know how to do this.
Can somebody tell me if TF 1.4 supports this and if so, point me to a guide or explain how its done ?
This is a little late but this video on youtube covers it pretty well.
He uses pyinstaller which grabs everything needed and puts it all either into one executable without anything else, or a folder with the exe in there and other stuff.
I've tried this myself and it works pretty well, although since pyinstaller smashes everything needed into one folder which gets really huge, it includes the entire tensorflow library, the python interpreter and if you use tensorflow-gpu, it also includes the cudnn files as well which are like 600mb, effectively leaving you with over a 1gb worth of files in the end.
That can be reduced by excluding modules that you don't need, I recommend creating a virtual environment and start with a clean installation of python.
Hope this helps in anyway.
I was wondering if it was possible to deploy Tensorflow custom ops or custom reader written in C++ inside cloud-ml.
It looks like cloud-ml does not accept running native code in its standard mode (I'm not really interested in using a virtualized environment), at least for Python package they only accept pure python with no C dependency.
Likely the easiest way to do this is to include as an extra package the build of the entire custom Tensorflow Wheel that includes the op. For specifying extra packages see: https://cloud.google.com/ml-engine/docs/how-tos/packaging-trainer#to_include_custom_dependencies_with_your_package
For building a TF wheel from source see: https://www.tensorflow.org/install/install_sources#build_the_pip_package
You could also try to download/install just the .so file for the new op, but that would require either downloading it inside the setup.py of your training package or inside the training python code itself.
Note that you can currently only upload custom packages during Training, and not during Batch or Online Prediction, so a model trained using a custom TF version may not work with the prediction service.
TensorFlow has a separate project for its production usage, as noted here, called TensorFlow Serving.
How should I use CNTK in a production environment, and how should I handle it's deployment? Hopefully one could deploy trained models in to a server/cluster that serves the model with RPC or HTTP REST API.
If no such tool exists, what should be the first steps to develop it and a good architecture to roll out on my own?
We do support serving CNTK models in a production environment. You can find information about model evaluation/inference: https://github.com/Microsoft/CNTK/wiki/CNTK-Evaluation-Overview. For deployment, you need deploy the dlls specified here. A tutorial for deploying CNTK in Azure is available here.
No such tool exists from the CNTK team, but the new APIs in C++ or python should make the task pretty easy once you have a trained model.
I have built syntaxnet, and tensorflow-serving using bazel. Both embed their own (partial?) copy of tensorflow itself. I already have the problem where I'd like to "import" some parts of tensorflow-serving in a script that "lives" in the syntaxnet tree which I can't figure out (without doing some VERY ugly things).
Now I'd like "tensorboard", but that apparently doesn't get built as part of the embedded tensorflow inside of syntaxnet or tensorflow-serving.
So now I'm sure "I'm doing it wrong". How am I supposed to be combining the artifacts built by various separate bazel workspaces?
In particular, how can I build tensorflow (with tensorboard) AND syntaxnet AND tensorflow-serving and have them "installed" for use so I can start writing my own scripts in a completely separate directory/repository?
Is "./bazel-bin/blah" really the end-game with bazel? There is no "make install" equivalent?
You're right, currently Tensorboard targets are only exposed in the Tensorflow repo, and not the other two that use it. That means that to actually bring up Tensorboard, you'll need to checkout Tensorflow on its own and compile/run Tensorboard there (pointing it to the generated logdir).
Actually generating the training summary data in a log directory is done during training, in your case in the tensorflow/models repo. It looks like SummaryWriter is used in inception_train.py, so perhaps you can add something similar to syntaxnet. If that doesn't work and you're not able to link Tensorboard, I'd recommend filing an issue in tensorflow/models to add support for Tensorboard there. You shouldn't need Tensorboard in Tensorflow Serving.
Importing parts of Tensorflow Serving in syntaxnet would require you to add this new depedency as a submodule (like is done with tensorflow) or possibly a git_repository in the WORKSPACE file if that works. We've never tried this, so it's possible that something is broken for this untested use case. Please file issues if you encounter a problem with this.
As for just installing and running, Tensorflow Serving doesn't support that right now. It's a set of libraries that you link in directly into your server binary and compile (the repo offers some example servers and clients), but right now there is no simple "installed server". Tensorflow along with Tensorboard however can be installed and linked from anywhere.