Convert GNMT* Model to the Intermediate Representation (IR)

This tutorial explains how to convert Google* Neural Machine Translation (GNMT) model to the Intermediate Representation (IR).

On GitHub*, you can find several public versions of TensorFlow* GNMT model implementation. This tutorial explains how to convert the GNMT model from the TensorFlow* Neural Machine Translation (NMT) repository to the IR.

Create a Patch File

Before converting the model, you need to create a patch file for the repository. The patch modifies the framework code by adding a special command-line argument to the framework options that enables inference graph dumping:

  1. Go to a writable directory and create a GNMT_inference.patch file.

  2. Copy the following diff code to the file:

    diff --git a/nmt/ b/nmt/
    index 2cbef07..e185490 100644
    --- a/nmt/
    +++ b/nmt/
    @@ -17,9 +17,11 @@
     from __future__ import print_function
     import codecs
    +import os
     import time
     import tensorflow as tf
    +from tensorflow.python.framework import graph_io
     from . import attention_model
     from . import gnmt_model
    @@ -105,6 +107,29 @@ def start_sess_and_load_model(infer_model, ckpt_path):
       return sess, loaded_infer_model
    +def inference_dump_graph(ckpt_path, path_to_dump, hparams, scope=None):
    +    model_creator = get_model_creator(hparams)
    +    infer_model = model_helper.create_infer_model(model_creator, hparams, scope)
    +    sess = tf.Session(
    +        graph=infer_model.graph, config=utils.get_config_proto())
    +    with infer_model.graph.as_default():
    +        loaded_infer_model = model_helper.load_model(
    +            infer_model.model, ckpt_path, sess, "infer")
    +    utils.print_out("Dumping inference graph to {}".format(path_to_dump))
    +        sess,
    +        os.path.join(path_to_dump + 'inference_GNMT_graph')
    +        )
    +    utils.print_out("Dumping done!")
    +    output_node_name = 'index_to_string_Lookup'
    +    utils.print_out("Freezing GNMT graph with output node {}...".format(output_node_name))
    +    frozen = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def,
    +                                                          [output_node_name])
    +    graph_io.write_graph(frozen, '.', os.path.join(path_to_dump, 'frozen_GNMT_inference_graph.pb'), as_text=False)
    +    utils.print_out("Freezing done. Freezed model frozen_GNMT_inference_graph.pb saved to {}".format(path_to_dump))
     def inference(ckpt_path,
    diff --git a/nmt/ b/nmt/
    index f5823d8..a733748 100644
    --- a/nmt/
    +++ b/nmt/
    @@ -310,6 +310,13 @@ def add_arguments(parser):
       parser.add_argument("--num_intra_threads", type=int, default=0,
                           help="number of intra_op_parallelism_threads")
    +  # Special argument for inference model dumping without inference
    +  parser.add_argument("--dump_inference_model", type="bool", nargs="?",
    +                      const=True, default=False,
    +                      help="Argument for dump inference graph for specified trained ckpt")
    +  parser.add_argument("--path_to_dump", type=str, default="",
    +                      help="Path to dump inference graph.")
     def create_hparams(flags):
       """Create training hparams."""
    @@ -396,6 +403,9 @@ def create_hparams(flags):
    +      dump_inference_model=flags.dump_inference_model,
    +      path_to_dump=flags.path_to_dump,
    @@ -613,7 +623,7 @@ def create_or_load_hparams(
       return hparams
    -def run_main(flags, default_hparams, train_fn, inference_fn, target_session=""):
    +def run_main(flags, default_hparams, train_fn, inference_fn, inference_dump, target_session=""):
       """Run main."""
       # Job
       jobid = flags.jobid
    @@ -653,8 +663,26 @@ def run_main(flags, default_hparams, train_fn, inference_fn, target_session=""):
             out_dir, default_hparams, flags.hparams_path,
             save_hparams=(jobid == 0))
    -  ## Train / Decode
    -  if flags.inference_input_file:
    +  #  Dumping inference model
    +  if flags.dump_inference_model:
    +      # Inference indices
    +      hparams.inference_indices = None
    +      if flags.inference_list:
    +          (hparams.inference_indices) = (
    +              [int(token) for token in flags.inference_list.split(",")])
    +      # Ckpt
    +      ckpt = flags.ckpt
    +      if not ckpt:
    +          ckpt = tf.train.latest_checkpoint(out_dir)
    +      # Path to dump graph
    +      assert flags.path_to_dump != "", "Please, specify path_to_dump model."
    +      path_to_dump = flags.path_to_dump
    +      if not tf.gfile.Exists(path_to_dump): tf.gfile.MakeDirs(path_to_dump)
    +      inference_dump(ckpt, path_to_dump, hparams)
    +  elif flags.inference_input_file:
         # Inference output directory
         trans_file = flags.inference_output_file
         assert trans_file
    @@ -693,7 +721,8 @@ def main(unused_argv):
       default_hparams = create_hparams(FLAGS)
       train_fn = train.train
       inference_fn = inference.inference
    -  run_main(FLAGS, default_hparams, train_fn, inference_fn)
    +  inference_dump = inference.inference_dump_graph
    +  run_main(FLAGS, default_hparams, train_fn, inference_fn, inference_dump)
     if __name__ == "__main__":
  3. Save and close the file.

Convert GNMT Model to IR


Please, use TensorFlow version 1.13 or lower.

Step 1. Clone the GitHub repository and check out the commit:

  1. Clone the NMT reposirory:

    git clone
  2. Check out the necessary commit:

    git checkout b278487980832417ad8ac701c672b5c3dc7fa553

Step 2. Get a trained model. You have two options:

  • Train the model with the GNMT wmt16_gnmt_4_layer.json or wmt16_gnmt_8_layer.json configuration file using the NMT framework.

  • Do not use the pre-trained checkpoints provided in the NMT repository, as they are outdated and can be incompatible with the current repository version.

This tutorial assumes the use of the trained GNMT model from wmt16_gnmt_4_layer.json config, German to English translation.

Step 3. Create an inference graph:

The OpenVINO assumes that a model is used for inference only. Hence, before converting the model into the IR, you need to transform the training graph into the inference graph. For the GNMT model, the training graph and the inference graph have different decoders: the training graph uses a greedy search decoding algorithm, while the inference graph uses a beam search decoding algorithm.

  1. Apply the GNMT_inference.patch patch to the repository. Refer to the Create a Patch File instructions if you do not have it:

    git apply /path/to/patch/GNMT_inference.patch
  2. Run the NMT framework to dump the inference model:

python -m nmt.nmt
    --infer_mode beam_search
    --path_to_dump /path/to/dump/model/

If you use different checkpoints, use the corresponding values for the src, tgt, ckpt, hparams_path, and vocab_prefix parameters. Inference checkpoint inference_GNMT_graph and frozen inference graph frozen_GNMT_inference_graph.pb will appear in the /path/to/dump/model/ folder.

To generate vocab.bpe.32000, execute the nmt/scripts/ script. If you face an issue of a size mismatch between the checkpoint graph’s embedding layer and vocabulary (both src and target), we recommend you to add the following code to the file to the extend_hparams function after the line 508 (after initialization of the src_vocab_size and tgt_vocab_size variables):

src_vocab_size -= 1
tgt_vocab_size -= 1

Step 4. Convert the model to the IR:

python3 path/to/model_optimizer/
--input_model /path/to/dump/model/frozen_GNMT_inference_graph.pb
--input "IteratorGetNext:1{i32}[1],IteratorGetNext:0{i32}[1 50],dynamic_seq2seq/hash_table_Lookup_1:0[1]->[2],dynamic_seq2seq/hash_table_Lookup:0[1]->[1]"
--output dynamic_seq2seq/decoder/decoder/GatherTree
--output_dir /path/to/output/IR/

Input and output cutting with the --input and --output options is required since OpenVINO does not support IteratorGetNext and LookupTableFindV2 operations.

Input cutting:

  • IteratorGetNext operation iterates over a dataset. It is cut by output ports: port 0 contains data tensor with shape [batch_size, max_sequence_length], port 1 contains sequence_length for every batch with shape [batch_size].

  • LookupTableFindV2 operations (dynamic_seq2seq/hash_table_Lookup_1 and dynamic_seq2seq/hash_table_Lookup nodes in the graph) are cut with constant values).

Output cutting:

  • LookupTableFindV2 operation is cut from the output and the dynamic_seq2seq/decoder/decoder/GatherTree node is treated as a new exit point.

For more information about model cutting, refer to Cutting Off Parts of a Model.

How to Use GNMT Model


This step assumes you have converted a model to the Intermediate Representation.

Inputs of the model:

  • IteratorGetNext/placeholder_out_port_0 input with shape [batch_size, max_sequence_length] contains batch_size decoded input sentences. Every sentence is decoded the same way as indices of sentence elements in vocabulary and padded with index of eos (end of sentence symbol). If the length of the sentence is less than max_sequence_length, remaining elements are filled with index of eos token.

  • IteratorGetNext/placeholder_out_port_1 input with shape [batch_size] contains sequence lengths for every sentence from the first input. For example, if max_sequence_length = 50, batch_size = 1 and the sentence has only 30 elements, then the input tensor for IteratorGetNext/placeholder_out_port_1 should be [30].

Outputs of the model:

  • dynamic_seq2seq/decoder/decoder/GatherTree tensor with shape [max_sequence_length * 2, batch, beam_size], that contains beam_size best translations for every sentence from input (also decoded as indices of words in vocabulary).


    Shape of this tensor in TensorFlow* can be different: instead of max_sequence_length * 2, it can be any value less than that, because OpenVINO does not support dynamic shapes of outputs, while TensorFlow can stop decoding iterations when eos symbol is generated.*

  1. With benchmark app:

    python3 -m <path to the generated GNMT IR> -d CPU
  2. With Inference Engine Python API:


Before running the example, insert a path to your GNMT .xml and .bin files into MODEL_PATH and WEIGHTS_PATH, and fill input_data_tensor and seq_lengths tensors according to your input data.

from openvino.inference_engine import IENetwork, IECore

MODEL_PATH = '/path/to/IR/frozen_GNMT_inference_graph.xml'
WEIGHTS_PATH = '/path/to/IR/frozen_GNMT_inference_graph.bin'

# Creating network
net = IENetwork(

# Creating input data
input_data = {'IteratorGetNext/placeholder_out_port_0': input_data_tensor,
              'IteratorGetNext/placeholder_out_port_1': seq_lengths}

# Creating plugin and loading extensions
ie = IECore()
ie.add_extension(extension_path="", device_name="CPU")

# Loading network
exec_net = ie.load_network(network=net, device_name="CPU")

# Run inference
result_ie = exec_net.infer(input_data)

For more information about Python API, refer to Inference Engine Python API Overview.