Add Converter Tools

tools/convert_tfjs.py

@@ -0,0 +1,102 @@
+# coding: utf-8
+# Copyright (c) 2018 Arata Furukawa
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+import tensorflow as tf
+from tensorflowjs import quantization
+from tensorflowjs.converters import tf_saved_model_conversion
+
+DEFAULT_TAGS = [tf.saved_model.tag_constants.SERVING]
+DEFAULT_SIGNATURE = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
+DEFAULT_INPUTS = ['image']
+DEFAULT_OUTPUTS = ['classes','probabilities']
+
+def convert_to_tfjs(
+    savedmodel_dir,
+    output_dir,
+    tags,
+    signature,
+    inputs,
+    outputs,
+    quantization_dtype,
+    skip_op_check,
+    strip_debug_ops
+  ):
+  with tf.Graph().as_default() as graph, tf.Session(graph=graph) as sess:
+    meta_graph = tf.saved_model.loader.load(sess, tags, savedmodel_dir)
+    meta = meta_graph.signature_def[signature]
+
+    output_node_names = [
+      meta.outputs[key].name for key in outputs
+    ]
+
+    # Getting input/output tensor name from signature for information.
+    input_tensors = [
+      graph.get_tensor_by_name(meta.inputs[key].name) for key in inputs
+    ]
+    output_tensors = [
+      graph.get_tensor_by_name(name) for name in output_node_names
+    ]
+    print('input tensors:')
+    for i, t in enumerate(input_tensors):
+      print('  {}: "{}" {}'.format(i, t.name, t.shape.as_list()))
+    print('output tensors:')
+    for i, t in enumerate(output_tensors):
+      print('  {}: "{}" {}'.format(i, t.name, t.shape.as_list()))
+
+    tf_saved_model_conversion.convert_tf_saved_model(
+      savedmodel_dir,
+      output_node_names=','.join([n.split(':')[0] for n in output_node_names]),
+      output_dir=output_dir,
+      saved_model_tags=','.join(tags),
+      quantization_dtype=quantization_dtype,
+      skip_op_check=skip_op_check,
+      strip_debug_ops=strip_debug_ops)
+
+if __name__ == '__main__':
+  import argparse
+  p = argparse.ArgumentParser()
+  p.add_argument('savedmodel_dir')
+  p.add_argument('output_dir')
+  p.add_argument('--tags', default=DEFAULT_TAGS)
+  p.add_argument('--signature', default=DEFAULT_SIGNATURE)
+  p.add_argument('--inputs', default=DEFAULT_INPUTS)
+  p.add_argument('--outputs', default=DEFAULT_OUTPUTS)
+  p.add_argument('--quantization_bytes', type=int, choices=set(quantization.QUANTIZATION_BYTES_TO_DTYPES.keys()))
+  p.add_argument('--skip_op_check', default=False)
+  p.add_argument('--strip_debug_ops', default=True)
+  args = p.parse_args()
+
+  quantization_dtype = (
+    quantization.QUANTIZATION_BYTES_TO_DTYPES[args.quantization_bytes]
+    if args.quantization_bytes else None)
+  convert_to_tfjs(
+    args.savedmodel_dir,
+    args.output_dir,
+    args.tags,
+    args.signature,
+    args.inputs,
+    args.outputs,
+    quantization_dtype,
+    args.skip_op_check,
+    args.strip_debug_ops
+  )
+
+

tools/convert_tflite.py

@@ -0,0 +1,106 @@
+# coding: utf-8
+# Copyright (c) 2018 Arata Furukawa
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+import tensorflow as tf
+
+# Workaround for bugs. Details refer below:
+# https://github.com/tensorflow/tensorflow/issues/15410
+import tempfile
+import subprocess
+tf.contrib.lite.tempfile = tempfile
+tf.contrib.lite.subprocess = subprocess
+
+DEFAULT_TAGS = [tf.saved_model.tag_constants.SERVING]
+DEFAULT_SIGNATURE = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
+DEFAULT_INPUTS = ['image']
+# Not using `classes` because TFLite not support ArgMax.
+DEFAULT_OUTPUTS = ['probabilities']
+
+def convert_to_tflite(
+    savedmodel_dir,
+    output_path,
+    tags,
+    signature,
+    inputs,
+    outputs
+  ):
+  with tf.Graph().as_default() as graph, tf.Session(graph=graph) as sess:
+    meta_graph = tf.saved_model.loader.load(sess, tags, savedmodel_dir)
+    meta = meta_graph.signature_def[signature]
+
+    # Freeze variables.
+    output_node_names = [
+      meta.outputs[key].name for key in outputs
+    ]
+    frozen_graph_def = tf.graph_util.convert_variables_to_constants(
+      sess, sess.graph_def, [
+        n.split(':')[0] for n in output_node_names
+      ])
+
+    def fix_shape(t):
+      if not t.shape.is_fully_defined():
+        t.set_shape([
+          d if d is not None else 1 for d in t.shape.as_list()
+        ])
+      return t
+
+    # Getting input/output tensor name from signature.
+    input_tensors = map(fix_shape, [
+      graph.get_tensor_by_name(meta.inputs[key].name) for key in inputs
+    ])
+    output_tensors = [
+      graph.get_tensor_by_name(name) for name in output_node_names
+    ]
+
+    print('input tensors:')
+    for i, t in enumerate(input_tensors):
+      print('  {}: "{}" {}'.format(i, t.name, t.shape.as_list()))
+    print('output tensors:')
+    for i, t in enumerate(output_tensors):
+      print('  {}: "{}" {}'.format(i, t.name, t.shape.as_list()))
+
+    # Convert to FlatBuffers by TOCO.
+    tflite_model = tf.contrib.lite.toco_convert(
+      frozen_graph_def, input_tensors, output_tensors)
+
+    # Save
+    with open(output_path, 'wb') as f:
+      f.write(tflite_model)
+
+if __name__ == '__main__':
+  import argparse
+  p = argparse.ArgumentParser()
+  p.add_argument('savedmodel_dir')
+  p.add_argument('output_path')
+  p.add_argument('--tags', default=DEFAULT_TAGS)
+  p.add_argument('--signature', default=DEFAULT_SIGNATURE)
+  p.add_argument('--inputs', default=DEFAULT_INPUTS)
+  p.add_argument('--outputs', default=DEFAULT_OUTPUTS)
+  args = p.parse_args()
+  convert_to_tflite(
+    args.savedmodel_dir,
+    args.output_path,
+    args.tags,
+    args.signature,
+    args.inputs,
+    args.outputs
+  )
+