Merge pull request #1690 from lissyx/convert-toco

Convert toco

.install

@@ -3,7 +3,7 @@
 virtualenv -p python3 ../tmp/venv
 source ../tmp/venv/bin/activate
 pip install -r <(grep -v tensorflow requirements.txt)
-pip install tensorflow-gpu==1.11.0
+pip install tensorflow-gpu==1.12.0rc2
 
 python3 util/taskcluster.py --arch gpu --target ../tmp/native_client
 

DeepSpeech.py

@@ -18,10 +18,12 @@ import time
 import traceback
 import inspect
 import progressbar
+import tempfile
 
 from functools import partial
 from six.moves import zip, range, filter, urllib, BaseHTTPServer
 from tensorflow.python.tools import freeze_graph
+from tensorflow.contrib.lite.python import tflite_convert
 from threading import Thread, Lock
 from util.audio import audiofile_to_input_vector
 from util.feeding import DataSet, ModelFeeder
@@ -1831,9 +1833,8 @@ def create_inference_graph(batch_size=1, n_steps=16, use_new_decoder=False, tfli
         return (
             {
                 'input': input_tensor,
-                'input_lengths': seq_length,
+                'previous_state_c': previous_state_c,
-                'new_state_c': new_state_c,
+                'previous_state_h': previous_state_h,
-                'new_state_h': new_state_h,
             },
             {
                 'outputs': logits,
@@ -1849,11 +1850,17 @@ def export():
     '''
     log_info('Exporting the model...')
     with tf.device('/cpu:0'):
+        from tensorflow.python.framework.ops import Tensor, Operation
 
         tf.reset_default_graph()
         session = tf.Session(config=session_config)
 
         inputs, outputs = create_inference_graph(batch_size=1, n_steps=FLAGS.n_steps, tflite=FLAGS.export_tflite)
+        input_names = ",".join(tensor.op.name for tensor in inputs.values())
+        output_names_tensors = [ tensor.op.name for tensor in outputs.values() if isinstance(tensor, Tensor) ]
+        output_names_ops = [ tensor.name for tensor in outputs.values() if isinstance(tensor, Operation) ]
+        output_names = ",".join(output_names_tensors + output_names_ops)
+        input_shapes = ":".join(",".join(map(str, tensor.shape)) for tensor in inputs.values())
 
         if not FLAGS.export_tflite:
             mapping = {v.op.name: v for v in tf.global_variables() if not v.op.name.startswith('previous_state_')}
@@ -1872,11 +1879,7 @@ def export():
         checkpoint = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
         checkpoint_path = checkpoint.model_checkpoint_path
 
-        if not FLAGS.export_tflite:
         output_filename = 'output_graph.pb'
-        else:
-            output_filename = 'output_graph.fb'
-
         if FLAGS.remove_export:
             if os.path.isdir(FLAGS.export_dir):
                 log_info('Removing old export')
@@ -1887,14 +1890,7 @@ def export():
             if not os.path.isdir(FLAGS.export_dir):
                 os.makedirs(FLAGS.export_dir)
 
-            if not FLAGS.export_tflite:
+            def do_graph_freeze(output_file=None, output_node_names=None, variables_blacklist=None):
-                output_node_names   = 'logits,initialize_state'
-                variables_blacklist = 'previous_state_c,previous_state_h'
-            else:
-                output_node_names   = 'logits,new_state_c,new_state_h'
-                variables_blacklist = ''
-
-            # Freeze graph
                 freeze_graph.freeze_graph_with_def_protos(
                     input_graph_def=session.graph_def,
                     input_saver_def=saver.as_saver_def(),
@@ -1902,16 +1898,53 @@ def export():
                     output_node_names=output_node_names,
                     restore_op_name=None,
                     filename_tensor_name=None,
-                output_graph=output_graph_path,
+                    output_graph=output_file,
                     clear_devices=False,
                     variable_names_blacklist=variables_blacklist,
                     initializer_nodes='')
 
+            if not FLAGS.export_tflite:
+                do_graph_freeze(output_file=output_graph_path, output_node_names=output_names, variables_blacklist='previous_state_c,previous_state_h')
+            else:
+                temp_fd, temp_freeze = tempfile.mkstemp(dir=FLAGS.export_dir)
+                os.close(temp_fd)
+                do_graph_freeze(output_file=temp_freeze, output_node_names=output_names, variables_blacklist='')
+                output_tflite_path = os.path.join(FLAGS.export_dir, output_filename.replace('.pb', '.tflite'))
+                class TFLiteFlags():
+                    def __init__(self):
+                        self.graph_def_file = temp_freeze
+                        self.inference_type = 'FLOAT'
+                        self.input_arrays   = input_names
+                        self.input_shapes   = input_shapes
+                        self.output_arrays  = output_names
+                        self.output_file    = output_tflite_path
+                        self.output_format  = 'TFLITE'
+
+                        default_empty = [
+                            'inference_input_type',
+                            'mean_values',
+                            'default_ranges_min', 'default_ranges_max',
+                            'drop_control_dependency',
+                            'reorder_across_fake_quant',
+                            'change_concat_input_ranges',
+                            'allow_custom_ops',
+                            'converter_mode',
+                            'post_training_quantize',
+                            'dump_graphviz_dir',
+                            'dump_graphviz_video'
+                        ]
+                        for e in default_empty:
+                            self.__dict__[e] = None
+
+                flags = TFLiteFlags()
+                tflite_convert._convert_model(flags)
+                os.unlink(temp_freeze)
+                log_info('Exported model for TF Lite engine as {}'.format(os.path.basename(output_tflite_path)))
+
             log_info('Models exported at %s' % (FLAGS.export_dir))
         except RuntimeError as e:
             log_error(str(e))
 
-
 def do_single_file_inference(input_file_path):
     with tf.Session(config=session_config) as session:
         inputs, outputs = create_inference_graph(batch_size=1, use_new_decoder=True)

Dockerfile

@@ -62,7 +62,7 @@ RUN wget https://bootstrap.pypa.io/get-pip.py && \
 # Clone TensoFlow from Mozilla repo
 RUN git clone https://github.com/mozilla/tensorflow/
 WORKDIR /tensorflow
-RUN git checkout r1.11
+RUN git checkout r1.12
 
 
 # GPU Environment Setup
@@ -190,7 +190,7 @@ RUN cp /tensorflow/bazel-bin/native_client/libctc_decoder_with_kenlm.so /DeepSpe
 
 # Install TensorFlow
 WORKDIR /DeepSpeech/
-RUN pip install tensorflow-gpu==1.11.0
+RUN pip install tensorflow-gpu==1.12.0rc2
 
 
 # Make DeepSpeech and install Python bindings

README.md

@@ -45,6 +45,7 @@ See the output of `deepspeech -h` for more information on the use of `deepspeech
   - [Training a model](#training-a-model)
   - [Checkpointing](#checkpointing)
   - [Exporting a model for inference](#exporting-a-model-for-inference)
+  - [Exporting a model for TFLite](#exporting-a-model-for-tflite)
   - [Distributed computing across more than one machine](#distributed-training-across-more-than-one-machine)
   - [Continuing training from a release model](#continuing-training-from-a-release-model)
 - [Code documentation](#code-documentation)
@@ -226,7 +227,7 @@ If you have a capable (Nvidia, at least 8GB of VRAM) GPU, it is highly recommend
 
 ```bash
 pip3 uninstall tensorflow
-pip3 install 'tensorflow-gpu==1.11.0'
+pip3 install 'tensorflow-gpu==1.12.0rc2'
 ```
 
 ### Common Voice training data
@@ -317,6 +318,10 @@ Be aware however that checkpoints are only valid for the same model geometry the
 If the `--export_dir` parameter is provided, a model will have been exported to this directory during training.
 Refer to the corresponding [README.md](native_client/README.md) for information on building and running a client that can use the exported model.
 
+### Exporting a model for TFLite
+
+If you want to experiment with the TF Lite engine, you need to export a model that is compatible with it, then use the `--export_tflite` flag. If you already have a trained model, you can re-export it for TFLite by running `DeepSpeech.py` again and specifying the same `checkpoint_dir` that you used for training, as well as passing `--notrain --notest --export_tflite --export_dir /model/export/destination`.
+
 ### Making a mmap-able model for inference
 
 The `output_graph.pb` model file generated in the above step will be loaded in memory to be dealt with when running inference.

native_client/README.md

@@ -52,7 +52,7 @@ Check the [main README](../README.md) for more details.
 If you'd like to build the binaries yourself, you'll need the following pre-requisites downloaded/installed:
 
 * [TensorFlow requirements](https://www.tensorflow.org/install/install_sources)
-* [TensorFlow `r1.11` sources](https://github.com/mozilla/tensorflow/tree/r1.11)
+* [TensorFlow `r1.12` sources](https://github.com/mozilla/tensorflow/tree/r1.12)
 * [libsox](https://sourceforge.net/projects/sox/)
 
 It is required to use our fork of TensorFlow since it includes fixes for common problems encountered when building the native client files.

requirements.txt

@@ -1,7 +1,7 @@
 pandas
 progressbar2
 python-utils
-tensorflow == 1.11.0
+tensorflow == 1.12.0rc2
 numpy
 matplotlib
 scipy

taskcluster/darwin-amd64-cpu-opt.yml

@@ -6,8 +6,7 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.osx.${event.head.sha}"
     - "notify.irc-channel.${notifications.irc}.on-exception"
     - "notify.irc-channel.${notifications.irc}.on-failed"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.osx/artifacts/public/home.tar.xz"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.osx/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.osx/artifacts/public/summarize_graph"
   scripts:
     build: "taskcluster/host-build.sh"
     package: "taskcluster/package.sh"

taskcluster/darwin-opt-base.tyml

@@ -39,7 +39,6 @@ payload:
       training: { $eval: as_slugid("test-training_upstream-linux-amd64-py27mu-opt") }
     in:
       TENSORFLOW_BUILD_ARTIFACT: ${build.tensorflow}
-      SUMMARIZE_GRAPH_BINARY: ${build.summarize_graph}
       DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
 
   # There is no VM yet running tasks on OSX

taskcluster/linux-amd64-cpu-opt.yml

@@ -14,8 +14,7 @@ build:
   system_config:
     >
       ${swig.patch_nodejs.linux}
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/home.tar.xz"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/summarize_graph"
   scripts:
     build: "taskcluster/host-build.sh"
     package: "taskcluster/package.sh"

taskcluster/linux-amd64-ctc-opt.yml

@@ -4,8 +4,7 @@ build:
     - "pull_request.synchronize"
     - "pull_request.reopened"
   template_file: linux-opt-base.tyml
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/home.tar.xz"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/summarize_graph"
   scripts:
     build: 'taskcluster/decoder-build.sh'
     package: 'taskcluster/decoder-package.sh'

taskcluster/linux-amd64-gpu-opt.yml

@@ -12,8 +12,7 @@ build:
   system_config:
     >
       ${swig.patch_nodejs.linux}
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.gpu/artifacts/public/home.tar.xz"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.gpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.gpu/artifacts/public/summarize_graph"
   maxRunTime: 14400
   scripts:
     build: "taskcluster/cuda-build.sh"

taskcluster/linux-arm64-cpu-opt.yml

@@ -4,8 +4,7 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branchortag}.arm64"
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branchortag}.${event.head.sha}.arm64"
     - "index.project.deepspeech.deepspeech.native_client.arm64.${event.head.sha}"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.arm64/artifacts/public/home.tar.xz"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.arm64/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/summarize_graph"
   ## multistrap 2.2.0-ubuntu1 is broken in 14.04: https://bugs.launchpad.net/ubuntu/+source/multistrap/+bug/1313787
   system_setup:
     >

taskcluster/linux-opt-base.tyml

@@ -36,7 +36,6 @@ then:
         training: { $eval: as_slugid("test-training_upstream-linux-amd64-py27mu-opt") }
       in:
         TENSORFLOW_BUILD_ARTIFACT: ${build.tensorflow}
-        SUMMARIZE_GRAPH_BINARY: ${build.summarize_graph}
         DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
 
     command:

taskcluster/linux-rpi3-cpu-opt.yml

@@ -4,8 +4,7 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branchortag}.arm"
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branchortag}.${event.head.sha}.arm"
     - "index.project.deepspeech.deepspeech.native_client.arm.${event.head.sha}"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.arm/artifacts/public/home.tar.xz"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.arm/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/summarize_graph"
   ## multistrap 2.2.0-ubuntu1 is broken in 14.04: https://bugs.launchpad.net/ubuntu/+source/multistrap/+bug/1313787
   system_setup:
     >

taskcluster/node-package.yml

@@ -16,8 +16,7 @@ build:
   system_config:
     >
       ${swig.patch_nodejs.linux}
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/home.tar.xz"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/summarize_graph"
   scripts:
     build: "taskcluster/node-build.sh"
     package: "taskcluster/node-package.sh"

taskcluster/test-armbian-opt-base.tyml

@@ -35,7 +35,6 @@ then:
         linux_arm64_build: { $eval: as_slugid("linux-arm64-cpu-opt") }
         node_package: { $eval: as_slugid("node-package") }
       in:
-        CONVERT_GRAPHDEF_MEMMAPPED: ${build.convert_graphdef}
         DEEPSPEECH_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${linux_arm64_build}/artifacts/public
         DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package}/artifacts/public
         DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
@@ -44,7 +43,7 @@ then:
         PIP_DEFAULT_TIMEOUT: "60"
         PIP_EXTRA_INDEX_URL: "https://lissyx.github.io/deepspeech-python-wheels/"
         EXTRA_PYTHON_CONFIGURE_OPTS: "--with-fpectl" # Required by Debian Stretch
-        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.11.0-11-gbee8254"
+        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.12.0-rc2-5-g1c93ca2"
 
     command:
       - "/bin/bash"

taskcluster/test-darwin-opt-base.tyml

@@ -41,7 +41,7 @@ then:
         DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
         DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.2.0-prod-ctcdecode/output_graph.pb
         DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.2.0-prod-ctcdecode/output_graph.pbmm
-        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.11.0-11-gbee8254"
+        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.12.0-rc2-5-g1c93ca2"
 
     command:
       - - "/bin/bash"

taskcluster/test-linux-opt-base.tyml

@@ -44,7 +44,7 @@ then:
         DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.2.0-prod-ctcdecode/output_graph.pb
         DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.2.0-prod-ctcdecode/output_graph.pbmm
         PIP_DEFAULT_TIMEOUT: "60"
-        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.11.0-11-gbee8254"
+        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.12.0-rc2-5-g1c93ca2"
 
     command:
       - "/bin/bash"

taskcluster/test-raspbian-opt-base.tyml

@@ -35,7 +35,6 @@ then:
         linux_rpi3_build: { $eval: as_slugid("linux-rpi3-cpu-opt") }
         node_package: { $eval: as_slugid("node-package") }
       in:
-        CONVERT_GRAPHDEF_MEMMAPPED: ${build.convert_graphdef}
         DEEPSPEECH_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${linux_rpi3_build}/artifacts/public
         DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package}/artifacts/public
         DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
@@ -44,7 +43,7 @@ then:
         PIP_DEFAULT_TIMEOUT: "60"
         PIP_EXTRA_INDEX_URL: "https://www.piwheels.org/simple"
         EXTRA_PYTHON_CONFIGURE_OPTS: "--with-fpectl" # Required by Raspbian Stretch / PiWheels
-        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.11.0-11-gbee8254"
+        EXPECTED_TENSORFLOW_VERSION: "TensorFlow: v1.12.0-rc2-5-g1c93ca2"
 
     command:
       - "/bin/bash"

taskcluster/test-training_upstream-linux-amd64-py27mu-opt.yml

@@ -7,7 +7,7 @@ build:
       apt-get -qq -y install ${python.packages_trusty.apt}
   args:
     tests_cmdline: "${system.homedir.linux}/DeepSpeech/ds/tc-train-tests.sh 2.7.14:mu"
-  convert_graphdef: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.11.bee825492fcf830bd65a024bf859cbfc218e1473.cpu/artifacts/public/convert_graphdef_memmapped_format"
+  convert_graphdef: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.12.1c93ca24c99d7011ad639eea4cd96e4fe45e1a95.cpu/artifacts/public/convert_graphdef_memmapped_format"
   metadata:
     name: "DeepSpeech Linux AMD64 CPU upstream training Py2.7 mu"
     description: "Training a DeepSpeech LDC93S1 model for Linux/AMD64 using upstream TensorFlow Python 2.7 mu, CPU only, optimized version"

tc-train-tests.sh

@@ -66,7 +66,7 @@ pushd ${HOME}/DeepSpeech/ds/
 popd
 
 cp /tmp/train/output_graph.pb ${TASKCLUSTER_ARTIFACTS}
-cp /tmp/train/output_graph.fb ${TASKCLUSTER_ARTIFACTS}
+cp /tmp/train/output_graph.tflite ${TASKCLUSTER_ARTIFACTS}
 
 if [ ! -z "${CONVERT_GRAPHDEF_MEMMAPPED}" ]; then
   convert_graphdef=$(basename "${CONVERT_GRAPHDEF_MEMMAPPED}")