/tfkg

Create, train, and save Tensorflow Keras models all in Golang

Primary LanguageGoMIT LicenseMIT

TFKG - A Tensorflow and Keras Golang port

This is experimental and quite nasty under the hood*

Summary

TFKG is a library for defining, training, saving, and running Tensorflow/Keras models with single GPU acceleration all in Golang.

The future of this project

See ideas-todo.md for what's in store

Tested Platforms

Platform OS CPU GPU Env CPU Support GPU Acceleration
Linux Ubuntu 18.04 Intel RTX 3090 Docker Yes Yes
Linux Ubuntu 18.04 Intel RTX 3090 Binary Yes Yes
Windows 11 AMD RTX 3080 Docker Yes Yes
Windows 11 AMD RTX 3080 Binary Yes Yes
Mac macOS 12 Intel AMD 5500m Docker Yes No
Mac macOS 12 M1 M1 Docker Yes No

Find your version

Versions starting with v0 are liable to change radically.

  • Tensorflow 2.6 experimental support: go get github.com/codingbeard/tfkg v0.26.28

Requirements

Docker

Linux environments are recommended, no GPU support on macOS and docker volumes are slow on macOS/Windows

Raw binary

Make sure to install the correct versions to match the version of this library

Features

  • Faster than typical python training
  • Define, train, evaluate, save, load, and infer Tensorflow compatible models all in Golang
  • Nvidia CUDA support on applicable platforms during Golang training/evaluation due to using the Tensorflow C library
  • Web interface for inspecting model training metrics. Use make web to start it
  • Load, shuffle, and preprocess csv datasets efficiently, even very large ones (tested on 300+GB csv file on a nvme ssd)
    • String Tokenizer
    • Float/Int normalization to between 0-1
    • Image loading and preprocessing
  • Automatic or custom class weighting for imbalanced datasets
  • Transfer learning between TFKG models

Keras model types supported

  • tensorflow.keras.Sequential (Single input)
  • tensorflow.keras.Model (Multiple input)

Keras Layers supported

Note that while the layers exist in the codebase, they were autogenerated and most have not been tested yet.

Keras Optimizers supported

Note that while the optimizers exist in the codebse, they were autogenerated and most have not been tested yet.

  • SGD
  • RMSprop
  • Adam
  • Adadelta
  • Adagrad
  • Adamax
  • Nadam
  • Ftrl

Keras Losses supported

  • Sparse categorical crossentropy
  • Binary crossentropy
  • Mean Squared Error
  • More coming soon

Metrics

  • Accuracy
  • False positive rate at true positive rate (Specificity at Sensitivity)
  • True positive rate at false positive rate (Sensitivity at Specificity)

Limitations

  • Python Tensorflow Libraries are still required to use this library, though the docker container has it all
  • This is an incomplete port of Tensorflow/Keras: There are many metrics and losses not yet ported
  • There is no community support or documentation. You must understand Tensorflow/Keras and Golang well to have a chance of getting this working on a new project
  • Using multiple GPU training is not supported

Examples:

Model Type Dataset Type Dataset Problem type Layers Location
Sequential Csv - Floats Iris Categorical Classification Input, Dense ./examples/iris
Functional Csv - Floats Iris Categorical Classification Input, Dense, Concatenate ./examples/multiple_inputs
Functional Csv - Strings Fraudulent Job Specs Binary Classification Input, Embedding, LSTM, Concatenate, Dense ./examples/jobs
Sequential Raw - Floats Random imbalanced Categorical Classification Input, Dense ./examples/class_weights
Sequential Images Sign Language Images Categorical Classification Input, Conv2D, MaxPooling2D, GlobalMaxPooling2D, Dense ./examples/sign
Sequential Csv - Floats Iris + Transferring Categorical Classification Input, Dense ./examples/transfer_learning
Sequential Csv - Floats Iris + loading vanilla keras model Categorical Classification - ./examples/vanilla
Functional Csv - Strings Fraudulent Job Specs Binary Classification Input, Embedding, CuDNNLSTM, Concatenate, Dense ./examples/gpu_train_cpu_infer

To test it out run the following then head to the web interface on http://localhost:8082

make init-docker
make web
make examples-iris

Define a model:

m := model.NewSequentialModel(
    logger,
    errorHandler,
    layer.Input().SetInputShape(tf.MakeShape(-1, 4)).SetDtype(layer.Float32),
    layer.Dense(100).SetActivation("swish"),
    layer.Dense(100).SetActivation("swish"),
    layer.Dense(float64(dataset.NumCategoricalClasses())).SetActivation("softmax"),
)

e = m.CompileAndLoad(model.LossSparseCategoricalCrossentropy, optimizer.NewAdam(), saveDir)
if e != nil {
    return
}

Load a dataset:

dataset, e := data.NewSingleFileDataset(
    logger,
    errorHandler,
    data.SingleFileDatasetConfig{
        FilePath:          "data/iris.data",
        CacheDir:          cacheDir,
        TrainPercent:      0.8,
        ValPercent:        0.1,
        TestPercent:       0.1,
        IgnoreParseErrors: true,
    },
    preprocessor.NewSparseCategoricalTokenizingYProcessor(
        errorHandler,
        cacheDir,
        4,
    ),
    preprocessor.NewProcessor(
        errorHandler,
        "petal_sizes",
        preprocessor.ProcessorConfig{
            CacheDir:    cacheDir,
            LineOffset:  0,
            DataLength:  4,
            RequiresFit: true,
            Divisor:     preprocessor.NewDivisor(errorHandler),
            Reader:      preprocessor.ReadCsvFloat32s,
            Converter:   preprocessor.ConvertDivisorToFloat32SliceTensor,
        },
    ),
)
if e != nil {
  errorHandler.Error(e)
  return
}

e = dataset.SaveProcessors(saveDir)
if e != nil {
    return
}

Train a model:

m.Fit(
    dataset,
    model.FitConfig{
        Epochs:     10,
        Validation: true,
        BatchSize:  batchSize,
        PreFetch:   10,
        Verbose:    1,
        Metrics: []metric.Metric{
            &metric.SparseCategoricalAccuracy{
                Name:       "acc",
                Confidence: 0.5,
                Average:    true,
            },
        },
        Callbacks: []callback.Callback{
            &callback.Logger{
                FileLogger: logger,
            },
            &callback.Checkpoint{
                OnEvent:    callback.EventEnd,
                OnMode:     callback.ModeVal,
                MetricName: "val_acc",
                Compare:    callback.CheckpointCompareMax,
                SaveDir:    saveDir,
            },
        },
    },
)

Load and predict using a saved TFKG model:

inference, e := data.NewInference(
    logger,
    errorHandler,
    saveDir,
    preprocessor.NewProcessor(
        errorHandler,
        "petal_sizes",
        preprocessor.ProcessorConfig{
            Converter: preprocessor.ConvertDivisorToFloat32SliceTensor,
        },
    ),
)
if e != nil {
    return
}

inputTensors, e := inference.GenerateInputs([][]float32{{6.0, 3.0, 4.8, 1.8}})
if e != nil {
    return
}

outputTensor, e := m.Predict(inputTensors...)
if e != nil {
    return
}

outputValues := outputTensor.Value().([][]float32)

logger.InfoF(
    "main",
    "Predicted classes: %s: %f, %s: %f, %s: %f",
    "Iris-setosa",
    outputValues[0][0],
    "Iris-versicolor",
    outputValues[0][1],
    "Iris-virginica",
    outputValues[0][2],
)

*Nasty under the hood

The Tensorflow/Keras python package saves a Graph (see more: https://www.tensorflow.org/guide/intro_to_graphs) which can be executed in other languages using their C library as long as there are C bindings.

The C library does not contain all the functionality of the python library when it comes to defining and saving models, it can only execute Graphs.

The Graph is calculated in python based on your model configuration, and a lot of clever code on the part of the developers in optimising the graph.

While possible, it is not currently feasible for me to generate the Graph in Golang, so I am relying on python to do so.

This means while the model is technically defined and trained in Golang, it just generates a json config string which static python code uses to configure the model and then saves it ready for loading in Golang for training. For the moment this is a needed evil.

If some kind soul wants to replicate Keras and Autograph to generate the Graph in Golang, feel free to make a pull request. I may eventually do it, but it is not likely. There is a branch origin/scratch which allows you to investigate the graph of a saved model.

Tensorflow C and Python library in a docker container on M1 Apple Silicon

See: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/lib_package/README.md

See: https://www.tensorflow.org/install/source#docker_linux_builds

Docker did not play nicely with the amd64 precompiled Tensorflow C library so I had to compile it from source with avx disabled on a different linux amd64 machine.

The compiled libraries and licenses can be found at: https://github.com/CodingBeard/tfkg/releases/tag/v0.2.6.5 and need to be placed in ./docker/tf-jupyter-golang-m1/

These are the steps I took to compile the library from sources to make it work:

// On a linux amd64 machine with docker installed:
git clone https://github.com/tensorflow/tensorflow
cd tensorflow
git checkout v2.6.0
docker run -it -w /tensorflow_src -v $PWD:/mnt -v $PWD:/tensorflow_src -e HOST_PERMS="$(id -u):$(id -g)" tensorflow/tensorflow:devel-gpu bash
> apt update && apt install apt-transport-https curl gnupg
> curl -fsSL https://bazel.build/bazel-release.pub.gpg | gpg --dearmor > bazel.gpg && \
    mv bazel.gpg /etc/apt/trusted.gpg.d/ && \
    echo "deb [arch=amd64] https://storage.googleapis.com/bazel-apt stable jdk1.8" | tee /etc/apt/sources.list.d/bazel.list
> apt update && apt install bazel-3.7.2 nano
> nano .bazelrc
// add the lines after the existing build:cuda lines:
build:cuda --linkopt=-lm
build:cuda --linkopt=-ldl
build:cuda --host_linkopt=-lm
build:cuda --host_linkopt=-ldl
> ./configure 
// take the defaults EXCEPT :
// ... "--config=opt" is specified [Default is -Wno-sign-compare]: -mno-avx
// The below will compile it for a specific GPU, find your gpu's compute capability and enter it twice separated by a comma (3000 series is 8.6)
// ... TensorFlow only supports compute capabilities >= 3.5 [Default is: 3.5,7.0]: 8.6,8.6
> bazel-3.7.2 build --config=cuda --config=opt //tensorflow/tools/lib_package:libtensorflow
> mkdir output
> cp bazel-bin/tensorflow/tools/lib_package/libtensorflow.tar.gz ./output/
> cp bazel-bin/tensorflow/tools/lib_package/clicenses.tar ./output/
> rm -r bazel-*
> bazel-3.7.2 build --config=cuda --config=opt //tensorflow/tools/pip_package:build_pip_package
> ./bazel-bin/tensorflow/tools/pip_package/build_pip_package ./output/tf-2.6.0-gpu-noavx
> quit
// copy the libs and wheel from ./output into the TFKG project under ./docker/tf-jupyter-golang-m1
...

Acknowledgements

Big shout out to github.com/galeone for their Tensorflow Golang fork for 2.6 and again for their article on how to train a model in golang which helped me figure out how to then save the trained variables: https://pgaleone.eu/tensorflow/go/2020/11/27/deploy-train-tesorflow-models-in-go-human-activity-recognition/