Merge pull request #127 from Cwiiis/serving-client

Write a Tensorflow Serving client, fixes #21

client/BUILD

@@ -0,0 +1,15 @@
+# Description: Deepspeech Serving Client.
+
+load("//tensorflow_serving:serving.bzl", "serving_proto_library")
+
+py_binary(
+    name = "deepspeech_client",
+    srcs = [
+        "deepspeech_client.py",
+    ],
+    deps = [
+        "//tensorflow_serving/apis:predict_proto_py_pb2",
+        "//tensorflow_serving/apis:prediction_service_proto_py_pb2",
+        "@org_tensorflow//tensorflow:tensorflow_py",
+    ],
+)

client/README.md

@@ -0,0 +1,42 @@
+# DeepSpeech client
+
+A client for running queries on an exported DeepSpeech model.
+
+## Requirements
+
+* [Tensorflow Serving](https://tensorflow.github.io/serving/setup)
+
+## Building
+
+Create a symbolic link in the Tensorflow Serving checkout to the deepspeech client directory.
+
+```
+cd serving
+ln -s ../DeepSpeech/deepspeech_client ./
+```
+
+If you haven't already, you'll need to build the Tensorflow Server.
+
+```
+bazel build -c opt //tensorflow_serving/model_servers:tensorflow_model_server
+```
+
+Then you can build the DeepSpeech client.
+
+```
+bazel build -c opt //deepspeech_client
+```
+
+## Running
+
+Start a server running an exported DeepSpeech model.
+
+```
+bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server --port=9000 --model_name=deepspeech --model_base_path=/path/to/deepspeech/export
+```
+
+Now run the client.
+
+```
+bazel-bin/deepspeech_client/deepspeech_client --server=localhost:9000 --file=/path/to/audio.wav
+```

client/deepspeech_client.py

@@ -0,0 +1,75 @@
+#!/usr/bin/env python2.7
+
+"""A client that talks to tensorflow_model_server loaded with deepspeech model.
+
+The client queries the service with the given audio and prints a ranked list
+of decoded outputs to the standard output, one per line.
+
+Typical usage example:
+
+    deepspeech_client.py --server=localhost:9000 --file audio.wav
+"""
+
+import os
+import sys
+import threading
+from grpc.beta import implementations
+import numpy as np
+import tensorflow as tf
+
+from tensorflow_serving.apis import predict_pb2
+from tensorflow_serving.apis import prediction_service_pb2
+
+sys.path.append(os.path.join(os.path.dirname(os.path.realpath(__file__)), '..'))
+from util.text import ndarray_to_text
+from util.audio import audiofile_to_input_vector
+
+tf.app.flags.DEFINE_string('server', '', 'PredictionService host:port')
+tf.app.flags.DEFINE_string('file', '', 'Wave audio file')
+# These need to match the constants used when training the deepspeech model
+tf.app.flags.DEFINE_string('n_input', 26, 'Number of MFCC features')
+tf.app.flags.DEFINE_string('n_context', 9, 'Number of frames of context')
+FLAGS = tf.app.flags.FLAGS
+
+def _create_rpc_callback(event):
+    def _callback(result_future):
+        exception = result_future.exception()
+        if exception:
+            print exception
+        else:
+            results = tf.contrib.util.make_ndarray(result_future.result().outputs['outputs'])
+            for result in results[0]:
+                print ndarray_to_text(result)
+        event.set()
+    return _callback
+
+def do_inference(hostport, audio):
+    host, port = hostport.split(':')
+    channel = implementations.insecure_channel(host, int(port))
+    stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
+
+    request = predict_pb2.PredictRequest()
+    request.model_spec.name = 'deepspeech'
+    request.inputs['input'].CopyFrom(tf.contrib.util.make_tensor_proto(audio))
+
+    event = threading.Event()
+    result_future = stub.Predict.future(request, 5.0)  # 5 seconds
+    result_future.add_done_callback(_create_rpc_callback(event))
+    if event.is_set() != True:
+        event.wait()
+
+def main(_):
+    if not FLAGS.server:
+        print 'please specify server host:port'
+        return
+    if not FLAGS.file:
+        print 'pleace specify an audio file'
+        return
+
+    audio_waves = audiofile_to_input_vector(
+                  FLAGS.file, FLAGS.n_input, FLAGS.n_context)
+    audio = np.array([ audio_waves ])
+    do_inference(FLAGS.server, audio)
+
+if __name__ == '__main__':
+    tf.app.run()

util/text.py

@@ -67,6 +67,12 @@ def sparse_tuple_to_texts(tuple):
     # List of strings
     return results
 
+def ndarray_to_text(value):
+    results = ''
+    for i in range(len(value)):
+        results += chr(value[i] + FIRST_INDEX)
+    return results.replace('`', ' ')
+
 def wer(original, result):
     """
     The WER is defined as the editing/Levenshtein distance on word level