/tensorflow-onnx

Convert TensorFlow models to ONNX

Primary LanguageJupyter NotebookMIT LicenseMIT

tf2onnx - Convert TensorFlow models to ONNX.

Build Type OS Python Tensorflow Onnx opset Status
Unit Test - Basic Linux, MacOS*, Windows* 3.5, 3.6 1.5-1.14 7-10 Build Status
Unit Test - Full Linux, MacOS, Windows 3.5, 3.6, 3.7 1.5-1.14 7-10 Build Status

* Only test on python3.6, TF1.14.

Supported ONNX version

tensorflow-onnx will use the ONNX version installed on your system and installs the latest ONNX version if none is found.

We support opset 6 to 10. By default we use opset 7 for the resulting ONNX graph since most runtimes will support opset 7. Support for future opsets add added as they are released.

If you want the graph to be generated with a specific opset, use --opset in the command line, for example --opset 10.

Status

We support many TensorFlow models. Support for Fully Connected, Convolutional and dynamic LSTM networks is mature. A list of models that we use for testing can be found here.

Supported RNN classes and APIs: LSTMCell, BasicLSTMCell, GRUCell, GRUBlockCell, MultiRNNCell, and user defined RNN cells inheriting rnn_cell_impl.RNNCell, used along with DropoutWrapper, BahdanauAttention, AttentionWrapper. Check tips when converting RNN models.

You find a list of supported Tensorflow ops and their mapping to ONNX here.

Tensorflow has broad functionality and occacional mapping it to ONNX creates issues. The common issues we run into we try to document here Troubleshooting Guide.

Prerequisites

Install TensorFlow

If you don't have tensorflow installed already, install the desired tensorflow build, for example:

pip install tensorflow
or
pip install tensorflow-gpu

Install runtime

If you want to run tests, install a runtime that can run ONNX models. For example:

ONNX Runtime (available for Linux, Windows, and Mac):

pip install onnxruntime

For pytorch/caffe2, follow the instructions here:

https://pytorch.org/

We tested with pytorch/caffe2 and onnxruntime and unit tests are passing for those.

Supported Tensorflow and Python Versions

We are testing with tensorflow 1.5-1.14 and anaconda 3.5,3.6,3.7.

Installation

From pypi

pip install -U tf2onnx

From Source

Once dependencies are installed, from the tensorflow-onnx folder call:

python setup.py install
or 
python setup.py develop

tensorflow-onnx requires onnx-1.5 or better and will install/upgrade onnx if needed.

To create a distribution:

python setup.py bdist_wheel

Usage

You find a end to end tutorial for ssd-mobilenet here.

To convert a TensorFlow model, tf2onnx supports saved_model, checkpoint or frozen graph formats. We recommend the saved_model format. If checkpoint or frozen graph formats are used, the user needs to specify inputs and outputs for the graph by passing the input and output names with --inputs INPUTS and --outputs OUTPUTS.

python -m tf2onnx.convert 
    [--input SOURCE_GRAPHDEF_PB]
    [--graphdef SOURCE_GRAPHDEF_PB]
    [--checkpoint SOURCE_CHECKPOINT]
    [--saved-model SOURCE_SAVED_MODEL]
    [--output TARGET_ONNX_MODEL]
    [--inputs GRAPH_INPUTS]
    [--outputs GRAPH_OUTPUS]
    [--inputs-as-nchw inputs_provided_as_nchw]
    [--opset OPSET]
    [--target TARGET]
    [--custom-ops list-of-custom-ops]
    [--fold_const]
    [--continue_on_error]
    [--verbose]

Parameters

--input or --graphdef

TensorFlow model as graphdef file. If not already frozen we'll try to freeze the model. More information about freezing can be found here: freeze graph tool.

--checkpoint

TensorFlow model as checkpoint. We expect the path to the .meta file. tf2onnx will try to freeze the graph.

--saved-model

TensorFlow model as saved_model. We expect the path to the saved_model directory. tf2onnx will try to freeze the graph.

--output

the target onnx file path.

--inputs, --outputs

Tensorflow model's input/output names, which can be found with summarize graph tool. Those names typically end on :0, for example --inputs input0:0,input1:0. inputs and outputs are not needed for models in saved-model format.

--inputs-as-nchw

By default we preserve the image format of inputs (nchw or nhwc) as given in the TensorFlow model. If your hosts (for example windows) native format nchw and the model is written for nhwc, --inputs-as-nchw tensorflow-onnx will transpose the input. Doing so is convinient for the application and the converter in many cases can optimize the transpose away. For example --inputs input0:0,input1:0 --inputs-as-nchw input0:0 assumes that images are passed into input0:0 as nchw while the TensorFlow model given uses nhwc.

--opset

By default we use the opset 7 to generate the graph. By specifying --opset the user can override the default to generate a graph with the desired opset. For example --opset 5 would create a onnx graph that uses only ops available in opset 5. Because older opsets have in most cases fewer ops, some models might not convert on a older opset.

--target

Some models require special handling to run on some runtimes. In particular, the model may use unsupported data types. Workarounds are activated with --target TARGET. Currently supported values are listed on this wiki. If your model will be run on Windows ML, you should specify the appropriate target value.

--custom-ops

the runtime may support custom ops that are not defined in onnx. A user can asked the converter to map to custom ops by listing them with the --custom-ops option. Tensorflow ops listed here will be mapped to a custom op with the same name as the tensorflow op but in the onnx domain ai.onnx.converters.tensorflow. For example: --custom-ops Print will insert a op Print in the onnx domain ai.onnx.converters.tensorflow into the graph. We also support a python api for custom ops documented later in this readme.

--fold_const

when set, TensorFlow fold_constants transformation will be applied before conversion. This will benefit features including Transpose optimization (e.g. Transpose operations introduced during tf-graph-to-onnx-graph conversion will be removed), and RNN unit conversion (for example LSTM). Older TensorFlow version might run into issues with this option depending on the model.

Usage example (run following commands in tensorflow-onnx root directory):

python -m tf2onnx.convert\
    --input tests/models/fc-layers/frozen.pb\
    --inputs X:0\
    --outputs output:0\
    --output tests/models/fc-layers/model.onnx\
    --verbose

Some models specify placeholders with unknown ranks and dims which can not be mapped to onnx. In those cases one can add the shape behind the input name in [], for example --inputs X:0[1,28,28,3]

Tool to get Graph Inputs & Outputs

To find the inputs and outputs for the TensorFlow graph the model developer will know or you can consult TensorFlow's summarize_graph tool, for example:

summarize_graph --in_graph=tests/models/fc-layers/frozen.pb

Tool to Freeze Graph

The TensorFlow tool to freeze the graph is here.

For example:

python -m tensorflow.python.tools.freeze_graph \
    --input_graph=my_checkpoint_dir/graphdef.pb \
    --input_binary=true \
    --output_node_names=output \
    --input_checkpoint=my_checkpoint_dir \
    --output_graph=tests/models/fc-layers/frozen.pb

Testing

There are 2 types of tests.

Unit test

python setup.py test

Validate pre-trained TensorFlow models

python tests/run_pretrained_models.py
usage: run_pretrained_models.py [-h] [--cache CACHE] [--tests TESTS] [--backend BACKEND] [--verbose] [--debug] [--config yaml-config]

optional arguments:
  -h, --help         show this help message and exit
  --cache CACHE      pre-trained models cache dir
  --tests TESTS      tests to run
  --backend BACKEND  backend to use
  --config           yaml config file
  --verbose          verbose output, option is additive
  --opset OPSET      target opset to use
  --perf csv-file    capture performance numbers for tensorflow and onnx runtime
  --debug            dump generated graph with shape info
  --fold_const when set, TensorFlow fold_constants transformation will be applied before conversion. This will benefit features including Transpose optimization (e.g. Transpose operations introduced during tf-graph-to-onnx-graph conversion will be removed), and RNN unit conversion (for example LSTM).

run_pretrained_models.py will run the TensorFlow model, captures the TensorFlow output and runs the same test against the specified ONNX backend after converting the model.

If the option --perf csv-file is specified, we'll capture the timeing for inferece of tensorflow and onnx runtime and write the result into the given csv file.

You call it for example with:

python tests/run_pretrained_models.py --backend onnxruntime --config tests/run_pretrained_models.yaml --perf perf.csv

Tool to save pre-trained model

We provide an utility to save pre-trained model along with its config. Put save_pretrained_model(sess, outputs, feed_inputs, save_dir, model_name) in your last testing epoch and the pre-trained model and config will be saved under save_dir/to_onnx. Please refer to the example in tools/save_pretrained_model.py for more information. Note the minimum required Tensorflow version is r1.6.

Using the Python API

TensorFlow to ONNX conversion

In some cases it will be useful to convert the models from TensorFlow to ONNX from a python script. You can use the following API:

import tf2onnx

tf2onnx.tfonnx.process_tf_graph(tf_graph, 
            continue_on_error=False, verbose=False, target=None,
            opset=None, custom_op_handlers=None,
            custom_rewriter=None, extra_opset=None,
            shape_override=None, inputs_as_nchw=None,
            input_names=None, output_names=None):
    """Convert tensorflow graph to onnx graph.
        Args:
            tf_graph: tensorflow graph
            continue_on_error: if an op can't be processed (aka there is no mapping), continue
            verbose: print summary stats (deprecated)
            target: list of workarounds applied to help certain platforms
            opset: the opset to be used (int, default is latest)
            custom_op_handlers: dictionary of custom ops handlers
            custom_rewriter: list of custom graph rewriters
            extra_opset: list of extra opset's, for example the opset's used by custom ops
            shape_override: dict with inputs that override the shapes given by tensorflow
            inputs_as_nchw: transpose inputs in list from nchw to nchw
            input_names: list of input node names in graph, input name format as node_name:port_id
            output_names: list of output node names in graph, output name format as node_name:port_id
        Return:
            onnx graph
    """

For example in examples/call_coverter_via_python.py:

import tensorflow as tf
import tf2onnx

with tf.Session() as sess:
    x = tf.placeholder(tf.float32, [2, 3], name="input")
    x_ = tf.add(x, x)
    _ = tf.identity(x_, name="output")
    onnx_graph = tf2onnx.tfonnx.process_tf_graph(sess.graph, input_names=["input:0"], output_names=["output:0"])
    model_proto = onnx_graph.make_model("test")
    with open("/tmp/model.onnx", "wb") as f:
        f.write(model_proto.SerializeToString())

Creating custom op mappings from python

For complex custom ops that require graph rewrites or input / attribute rewrites using the python interface to insert a custom op will be the eaiest way to accomplish the task. A dictionary of name->custom_op_handler can be passed to tf2onnx.tfonnx.process_tf_graph. If the op name is found in the graph the handler will have access to all internal structures and can rewrite that is needed. For example examples/custom_op_via_python.py:

import tensorflow as tf
import tf2onnx
from onnx import helper

_TENSORFLOW_DOMAIN = "ai.onnx.converters.tensorflow"


def print_handler(ctx, node, name, args):
    # replace tf.Print() with Identity
    #   T output = Print(T input, data, @list(type) U, @string message, @int first_n, @int summarize)
    # becomes:
    #   T output = Identity(T Input)
    node.domain = _TENSORFLOW_DOMAIN
    del node.input[1:]
    return node


with tf.Session() as sess:
    x = tf.placeholder(tf.float32, [2, 3], name="input")
    x_ = tf.add(x, x)
    x_ = tf.Print(x, [x], "hello")
    _ = tf.identity(x_, name="output")
    onnx_graph = tf2onnx.tfonnx.process_tf_graph(sess.graph,
                                                 custom_op_handlers={"Print": (print_handler, ["Identity", "mode"])},
                                                 extra_opset=[helper.make_opsetid(_TENSORFLOW_DOMAIN, 1)],
                                                 input_names=["input:0"],
                                                 output_names=["output:0"])
    model_proto = onnx_graph.make_model("test")
    with open("/tmp/model.onnx", "wb") as f:
        f.write(model_proto.SerializeToString())

How tf2onnx works

The converter needs to take care of a few things:

  1. Convert the protobuf format. Since the format is similar this step is straight forward.
  2. TensorFlow types need to be mapped to their ONNX equivalent.
  3. For many ops TensorFlow passes parameters like shapes as inputs where ONNX wants to see them as attributes. Since we use a frozen graph, the converter will fetch the input as constant, converts it to an attribute and remove the original input.
  4. TensorFlow in many cases composes ops out of multiple simpler ops. The converter will need to identify the subgraph for such ops, slice the subgraph out and replace it with the ONNX equivalent. This can become fairly complex so we use a graph matching library for it. A good example of this is the tensorflow transpose op.
  5. TensorFlow's default data format is NHWC where ONNX requires NCHW. The converter will insert transpose ops to deal with this.
  6. There are some ops like relu6 that are not supported in ONNX but the converter can be composed out of other ONNX ops.
  7. ONNX backends are new and their implementations are not complete yet. For some ops the converter generate ops with deal with issues in existing backends.

Step 1 - start with a frozen graph.

tf2onnx starts with a frozen graph. This is because of item 3 above.

Step 2 - 1:1 convertion of the protobuf from tensorflow to onnx

tf2onnx first does a simple convertion from the TensorFlow protobuf format to the ONNX protobuf format without looking at individual ops. We do this so we can use the ONNX graph as internal representation and write helper functions around it. The code that does the conversion is in tensorflow_to_onnx(). tensorflow_to_onnx() will return the ONNX graph and a dictionary with shape information from TensorFlow. The shape information is helpful in some cases when processing individual ops. The ONNX graph is wrapped in a Graph object and nodes in the graph are wrapped in a Node object to allow easier graph manipulations on the graph. All code that deals with nodes and graphs is in graph.py.

Step 3 - rewrite subgraphs

In the next step we apply graph matching code on the graph to re-write subgraphs for ops like transpose and lstm. For an example looks at rewrite_transpose().

Step 4 - process individual ops

In the fourth step we look at individual ops that need attention. The dictionary _OPS_MAPPING will map tensorflow op types to a method that is used to process the op. The simplest case is direct_op() where the op can be taken as is. Whenever possible we try to group ops into common processing, for example all ops that require dealing with broadcasting are mapped to broadcast_op(). For an op that composes the tensorflow op from multiple onnx ops, see relu6_op().

Step 5 - final processing

Once all ops are converted, we need to do a topological sort since ONNX requires it. process_tf_graph() is the method that takes care of all above steps.

Extending tf2onnx

If you like to contribute and add new conversions to tf2onnx, the process is something like:

  1. See if the op fits into one of the existing mappings. If so adding it to _OPS_MAPPING is all that is needed.
  2. If the new op needs extra procesing, start a new mapping function.
  3. If the tensorflow op is composed of multiple ops, consider using a graph re-write. While this might be a little harder initially, it works better for complex patterns.
  4. Add a unit test in tests/test_backend.py. The unit tests mostly create the tensorflow graph, run it and capture the output, than convert to onnx, run against a onnx backend and compare tensorflow and onnx results.
  5. If there are pre-trained models that use the new op, consider adding those to test/run_pretrained_models.py.

License

MIT License