riga/tfdeploy

Deploy tensorflow graphs for fast evaluation and export to tensorflow-less environments running numpy.

Now with tensorflow 1.0 support.

Evaluation usage

import tfdeploy as td
import numpy as np

model = td.Model("/path/to/model.pkl")
inp, outp = model.get("input", "output")

batch = np.random.rand(10000, 784)
result = outp.eval({inp: batch})

Installation and dependencies

Via pip

pip install tfdeploy

or by simply copying the file into your project.

Numpy ≥ 1.10 should be installed on your system. Scipy is optional. See optimization for more info on optional packages.

By design, tensorflow is required when creating a model. All versions ≥ 1.0.1 are supported.

Content

• Why?
• How?
  • Convert your graph
  • Load the model and evaluate
  • Write your own operation
• Ensembles
• Optimization
• Performance
• Contributing
• Development
• Authors
• License

Why?

Working with tensorflow is awesome. Model definition and training is simple yet powerful, and the range of built-in features is just striking.

However, when it comes down to model deployment and evaluation, things get a bit more cumbersome than they should be. You either export your graph to a new file and save your trained variables in a separate file, or you make use of tensorflow's serving system. Wouldn't it be great if you could just export your model to a simple numpy-based callable? Of course it would. And this is exactly what tfdeploy does for you.

To boil it down, tfdeploy

• is lightweight. A single file with < 150 lines of core code. Just copy it to your project.
• faster than using tensorflow's Tensor.eval.
• does not need tensorflow during evaluation.
• only depends on numpy.
• can load one or more models from a single file.
• does not support GPUs (maybe gnumpy is worth a try here).

How?

The central class is tfdeploy.Model. The following two examples demonstrate how a model can be created from a tensorflow graph, saved to and loaded from disk, and eventually evaluated.

Convert your graph

import tensorflow as tf
import tfdeploy as td

# setup tfdeploy (only when creating models)
td.setup(tf)

# build your graph
sess = tf.Session()

# use names for input and output layers
x = tf.placeholder("float", shape=[None, 784], name="input")
W = tf.Variable(tf.truncated_normal([784, 100], stddev=0.05))
b = tf.Variable(tf.zeros([100]))
y = tf.nn.softmax(tf.matmul(x, W) + b, name="output")

sess.run(tf.global_variables_initializer())

# ... training ...

# create a tfdeploy model and save it to disk
model = td.Model()
model.add(y, sess) # y and all its ops and related tensors are added recursively
model.save("model.pkl")

Load the model and evaluate

import numpy as np
import tfdeploy as td

model = td.Model("model.pkl")

# shorthand to x and y
x, y = model.get("input", "output")

# evaluate
batch = np.random.rand(10000, 784)
result = y.eval({x: batch})

Write your own Operation

tfdeploy supports most of the Operation's implemented in tensorflow. However, if you miss one (in that case, submit a PR or an issue ;) ) or if you're using custom ops, you might want to extend tfdeploy by defining a new class op that inherits from tfdeploy.Operation:

import tensorflow as tf
import tfdeploy as td
import numpy as np

# setup tfdeploy (only when creating models)
td.setup(tf)

# ... write you model here ...

# let's assume your final tensor "y" relies on an op of type "InvertedSoftmax"
# before creating the td.Model, you should add that op to tfdeploy

class InvertedSoftmax(td.Operation):
    @staticmethod
    def func(a):
        e = np.exp(-a)
        # ops should return a tuple
        return np.divide(e, np.sum(e, axis=-1, keepdims=True)),

# this is equivalent to
# @td.Operation.factory
# def InvertedSoftmax(a):
#     e = np.exp(-a)
#     return np.divide(e, np.sum(e, axis=-1, keepdims=True)),

# now we're good to go
model = td.Model()
model.add(y, sess)
model.save("model.pkl")

When writing new ops, three things are important:

• Try to avoid loops, prefer numpy vectorization.
• Return a tuple.
• Don't change incoming tensors/arrays in-place, always work on and return copies.

Ensembles

tfdeploy provides a helper class to evaluate an ensemble of models: Ensemble. It can load multiple models, evaluate them and combine their output values using different methods.

# create the ensemble
ensemble = td.Ensemble(["model1.pkl", "model2.pkl", ...], method=td.METHOD_MEAN)

# get input and output tensors (which actually are TensorEnsemble instances)
input, output = ensemble.get("input", "output")

# evaluate the ensemble just like a normal model
batch = ...
value = output.eval({input: batch})

The return value of get() is a TensorEnsemble istance. It is basically a wrapper around multiple tensors and should be used as keys in the feed_dict of the eval() call.

You can choose between METHOD_MEAN (the default), METHOD_MAX and METHOD_MIN. If you want to use a custom ensembling method, use METHOD_CUSTOM and overwrite the static func_custom() method of the TensorEnsemble instance.

Optimization

Most ops are written using pure numpy. However, multiple implementations of the same op are allowed that may use additional third-party Python packages providing even faster functionality for some situations.

For example, numpy does not provide a vectorized lgamma function. Thus, the standard tfdeploy.Lgamma op uses math.lgamma that was previously vectorized using numpy.vectorize. For these situations, additional implementations of the same op are possible (the lgamma example is quite academic, but this definitely makes sense for more sophisticated ops like pooling). We can simply tell the op to use its scipy implementation instead:

td.Lgamma.use_impl(td.IMPL_SCIPY)

Currently, allowed implementation types are numpy (IMPL_NUMPY, the default) and scipy (IMPL_SCIPY).

Adding additional implementations

Additional implementations can be added by setting the impl attribute of the op factory or by using the add_impl decorator of existing operations. The first registered implementation will be the default one.

# create the default lgamma op with numpy implementation
lgamma_vec = np.vectorize(math.lgamma)

@td.Operation.factory
# equivalent to
# @td.Operation.factory(impl=td.IMPL_NUMPY)
def Lgamma(a):
    return lgamma_vec(a),

# add a scipy-based implementation
@Lgamma.add_impl(td.IMPL_SCIPY)
def Lgamma(a):
    return sp.special.gammaln(a),

Auto-optimization

If scipy is available on your system, it is reasonable to use all ops in their scipy implementation (if it exists, of course). This should be configured before you create any model from tensorflow objects using the second argument of the setup function:

td.setup(tf, td.IMPL_SCIPY)

Ops that do not implement IMPL_SCIPY stick with the numpy version (IMPL_NUMPY).

Performance

tfdeploy is lightweight (1 file, < 150 lines of core code) and fast. Internal evaluation calls have only very few overhead and tensor operations use numpy vectorization. The actual performance depends on the ops in your graph. While most of the tensorflow ops have a numpy equivalent or can be constructed from numpy functions, a few ops require additional Python-based loops (e.g. BatchMatMul). But in many cases it's potentially faster than using tensorflow's Tensor.eval.

This is a comparison for a basic graph where all ops are vectorized (basically Add, MatMul and Softmax):

> ipython -i tests/perf/simple.py

In [1]: %timeit -n 100 test_tf()
100 loops, best of 3: 109 ms per loop

In [2]: %timeit -n 100 test_td()
100 loops, best of 3: 60.5 ms per loop

Contributing

If you want to contribute with new ops and features, I'm happy to receive pull requests. Just make sure to add a new test case to tests/core.py or tests/ops.py and run them via:

> python -m unittest tests

Test grid

In general, tests should be run for different environments:

Docker

For testing purposes, it is convenient to use docker. Fortunately, the official tensorflow images contain all we need:

git clone https://github.com/riga/tfdeploy.git
cd tfdeploy

docker run --rm -v `pwd`:/root/tfdeploy -w /root/tfdeploy -e "TD_TEST_SCIPY=1" tensorflow/tensorflow:1.0.1 python -m unittest tests

Development

• Source hosted at GitHub
• Report issues, questions, feature requests on GitHub Issues

Authors

• Marcel R.