def export():
with tf.Graph().as_default():
# Build Aquila model.
# Please refer to Tensorflow inception model for details.
flat_image_size = 3 * FLAGS.image_size**2
input_data = tf.placeholder(tf.float32, shape=(None, flat_image_size))
# reshape the images appropriately
images = tf.reshape(input_data,
(-1, FLAGS.image_size, FLAGS.image_size, 3))
# Run inference.
logits, _ = aquila_model.inference(images,
for_training=False,
restore_logits=True)
# Restore variables from training checkpoint.
variable_averages = tf.train.ExponentialMovingAverage(
aquila_model.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
with tf.Session() as sess:
# Restore variables from training checkpoints.
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/imagenet_train/model.ckpt-0,
# extract global_step from it.
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
print('Successfully loaded model from %s at step=%s.' %
(ckpt.model_checkpoint_path, global_step))
else:
print('No checkpoint file found at %s' % FLAGS.checkpoint_dir)
return
# Export inference model.
model_exporter = exporter.Exporter(saver)
signature = exporter.regression_signature(jpegs, logits)
model_exporter.init(default_graph_signature=signature)
model_exporter.export(FLAGS.export_dir, tf.constant(global_step),
sess)
print('Successfully exported model to %s' % FLAGS.export_dir)
def doBasicsOneExportPath(self,
export_path,
clear_devices=False,
global_step=GLOBAL_STEP):
# Build a graph with 2 parameter nodes on different devices.
tf.reset_default_graph()
with tf.Session(target="",
config=config_pb2.ConfigProto(
device_count={"CPU": 2})) as sess:
# v2 is an unsaved variable derived from v0 and v1. It is used to
# exercise the ability to run an init op when restoring a graph.
with sess.graph.device("/cpu:0"):
v0 = tf.Variable(10, name="v0")
with sess.graph.device("/cpu:1"):
v1 = tf.Variable(20, name="v1")
v2 = tf.Variable(1, name="v2", trainable=False, collections=[])
assign_v2 = tf.assign(v2, tf.add(v0, v1))
init_op = tf.group(assign_v2, name="init_op")
tf.add_to_collection("v", v0)
tf.add_to_collection("v", v1)
tf.add_to_collection("v", v2)
global_step_tensor = tf.Variable(global_step, name="global_step")
named_tensor_bindings = {
"logical_input_A": v0,
"logical_input_B": v1
}
signatures = {
"foo":
exporter.regression_signature(input_tensor=v0,
output_tensor=v1),
"generic":
exporter.generic_signature(named_tensor_bindings)
}
def write_asset(path):
file_path = os.path.join(path, "file.txt")
with gfile.FastGFile(file_path, "w") as f:
f.write("your data here")
asset_file = tf.Variable("hello42.txt", name="filename42")
assets = {("hello42.txt", asset_file)}
tf.initialize_all_variables().run()
# Run an export.
save = tf.train.Saver({
"v0": v0,
"v1": v1
},
restore_sequentially=True,
sharded=True)
export = exporter.Exporter(save)
export.init(
sess.graph.as_graph_def(),
init_op=init_op,
clear_devices=clear_devices,
default_graph_signature=exporter.classification_signature(
input_tensor=v0),
named_graph_signatures=signatures,
assets=assets,
assets_callback=write_asset)
export.export(export_path,
global_step_tensor,
sess,
exports_to_keep=gc.largest_export_versions(2))
# Restore graph.
compare_def = tf.get_default_graph().as_graph_def()
tf.reset_default_graph()
with tf.Session(target="",
config=config_pb2.ConfigProto(
device_count={"CPU": 2})) as sess:
save = tf.train.import_meta_graph(
os.path.join(export_path,
exporter.VERSION_FORMAT_SPECIFIER % global_step,
exporter.META_GRAPH_DEF_FILENAME))
self.assertIsNotNone(save)
meta_graph_def = save.export_meta_graph()
collection_def = meta_graph_def.collection_def
# Validate custom graph_def.
graph_def_any = collection_def[exporter.GRAPH_KEY].any_list.value
self.assertEquals(len(graph_def_any), 1)
graph_def = tf.GraphDef()
graph_def_any[0].Unpack(graph_def)
if clear_devices:
for node in compare_def.node:
node.device = ""
self.assertProtoEquals(compare_def, graph_def)
# Validate init_op.
init_ops = collection_def[exporter.INIT_OP_KEY].node_list.value
self.assertEquals(len(init_ops), 1)
self.assertEquals(init_ops[0], "init_op")
# Validate signatures.
signatures_any = collection_def[
exporter.SIGNATURES_KEY].any_list.value
self.assertEquals(len(signatures_any), 1)
signatures = manifest_pb2.Signatures()
signatures_any[0].Unpack(signatures)
default_signature = signatures.default_signature
self.assertEqual(
default_signature.classification_signature.input.tensor_name,
"v0:0")
bindings = signatures.named_signatures[
"generic"].generic_signature.map
self.assertEquals(bindings["logical_input_A"].tensor_name, "v0:0")
self.assertEquals(bindings["logical_input_B"].tensor_name, "v1:0")
read_foo_signature = (
signatures.named_signatures["foo"].regression_signature)
self.assertEquals(read_foo_signature.input.tensor_name, "v0:0")
self.assertEquals(read_foo_signature.output.tensor_name, "v1:0")
# Validate the assets.
assets_any = collection_def[exporter.ASSETS_KEY].any_list.value
self.assertEquals(len(assets_any), 1)
asset = manifest_pb2.AssetFile()
assets_any[0].Unpack(asset)
assets_path = os.path.join(
export_path, exporter.VERSION_FORMAT_SPECIFIER % global_step,
exporter.ASSETS_DIRECTORY, "file.txt")
asset_contents = gfile.GFile(assets_path).read()
self.assertEqual(asset_contents, "your data here")
self.assertEquals("hello42.txt", asset.filename)
self.assertEquals("filename42:0", asset.tensor_binding.tensor_name)
# Validate graph restoration.
save.restore(
sess,
os.path.join(export_path,
exporter.VERSION_FORMAT_SPECIFIER % global_step,
exporter.VARIABLES_DIRECTORY))
self.assertEqual(10, tf.get_collection("v")[0].eval())
self.assertEqual(20, tf.get_collection("v")[1].eval())
tf.get_collection(exporter.INIT_OP_KEY)[0].run()
self.assertEqual(30, tf.get_collection("v")[2].eval())
def Export():
export_path = "/tmp/half_plus_two"
with tf.Session() as sess:
# Make model parameters a&b variables instead of constants to
# exercise the variable reloading mechanisms.
a = tf.Variable(0.5, name="a")
b = tf.Variable(2.0, name="b")
# Calculate, y = a*x + b
# here we use a placeholder 'x' which is fed at inference time.
x = tf.placeholder(tf.float32, name="x")
y = tf.add(tf.mul(a, x), b, name="y")
# Setup a standard Saver for our variables.
save = tf.train.Saver({"a": a, "b": b}, sharded=True)
# asset_path contains the base directory of assets used in training (e.g.
# vocabulary files).
original_asset_path = tf.constant("/tmp/original/export/assets")
# Ops reading asset files should reference the asset_path tensor
# which stores the original asset path at training time and the
# overridden assets directory at restore time.
asset_path = tf.Variable(original_asset_path,
name="asset_path",
trainable=False,
collections=[])
assign_asset_path = asset_path.assign(original_asset_path)
# CopyAssets is used as a callback during export to copy files to the
# given export directory.
def CopyAssets(export_path):
print("copying asset files to: %s" % export_path)
# Use a fixed global step number.
global_step_tensor = tf.Variable(123, name="global_step")
# Create a RegressionSignature for our input and output.
signature = exporter.regression_signature(input_tensor=x, output_tensor=y)
# Create two filename assets and corresponding tensors.
# TODO(b/26254158) Consider adding validation of file existance as well as
# hashes (e.g. sha1) for consistency.
original_filename1 = tf.constant("hello1.txt")
filename1 = tf.Variable(original_filename1,
name="filename1",
trainable=False,
collections=[])
assign_filename1 = filename1.assign(original_filename1)
original_filename2 = tf.constant("hello2.txt")
filename2 = tf.Variable(original_filename2,
name="filename2",
trainable=False,
collections=[])
assign_filename2 = filename2.assign(original_filename2)
assets = {("hello1.txt", original_filename1),
("hello2.txt", original_filename2)}
# Init op contains a group of all variables that we assign.
init_op = tf.group(assign_asset_path, assign_filename1, assign_filename2)
# Run an export.
tf.initialize_all_variables().run()
export = exporter.Exporter(save)
export.init(sess.graph.as_graph_def(),
init_op=init_op,
default_graph_signature=signature,
assets=assets,
assets_callback=CopyAssets)
export.export(export_path, global_step_tensor, sess)
def doBasicsOneExportPath(self,
export_path,
clear_devices=False,
global_step=GLOBAL_STEP,
sharded=True):
# Build a graph with 2 parameter nodes on different devices.
tf.reset_default_graph()
with tf.Session(
target="",
config=config_pb2.ConfigProto(device_count={"CPU": 2})) as sess:
# v2 is an unsaved variable derived from v0 and v1. It is used to
# exercise the ability to run an init op when restoring a graph.
with sess.graph.device("/cpu:0"):
v0 = tf.Variable(10, name="v0")
with sess.graph.device("/cpu:1"):
v1 = tf.Variable(20, name="v1")
v2 = tf.Variable(1, name="v2", trainable=False, collections=[])
assign_v2 = tf.assign(v2, tf.add(v0, v1))
init_op = tf.group(assign_v2, name="init_op")
tf.add_to_collection("v", v0)
tf.add_to_collection("v", v1)
tf.add_to_collection("v", v2)
global_step_tensor = tf.Variable(global_step, name="global_step")
named_tensor_bindings = {"logical_input_A": v0, "logical_input_B": v1}
signatures = {
"foo": exporter.regression_signature(input_tensor=v0,
output_tensor=v1),
"generic": exporter.generic_signature(named_tensor_bindings)
}
asset_filepath_orig = os.path.join(tf.test.get_temp_dir(), "hello42.txt")
asset_file = tf.constant(asset_filepath_orig, name="filename42")
tf.add_to_collection(tf.GraphKeys.ASSET_FILEPATHS, asset_file)
with gfile.FastGFile(asset_filepath_orig, "w") as f:
f.write("your data here")
assets_collection = tf.get_collection(tf.GraphKeys.ASSET_FILEPATHS)
ignored_asset = os.path.join(tf.test.get_temp_dir(), "ignored.txt")
with gfile.FastGFile(ignored_asset, "w") as f:
f.write("additional data here")
tf.initialize_all_variables().run()
# Run an export.
save = tf.train.Saver({"v0": v0,
"v1": v1},
restore_sequentially=True,
sharded=sharded)
export = exporter.Exporter(save)
export.init(sess.graph.as_graph_def(),
init_op=init_op,
clear_devices=clear_devices,
default_graph_signature=exporter.classification_signature(
input_tensor=v0),
named_graph_signatures=signatures,
assets_collection=assets_collection)
export.export(export_path,
global_step_tensor,
sess,
exports_to_keep=gc.largest_export_versions(2))
# Restore graph.
compare_def = tf.get_default_graph().as_graph_def()
tf.reset_default_graph()
with tf.Session(
target="",
config=config_pb2.ConfigProto(device_count={"CPU": 2})) as sess:
save = tf.train.import_meta_graph(
os.path.join(export_path, constants.VERSION_FORMAT_SPECIFIER %
global_step, constants.META_GRAPH_DEF_FILENAME))
self.assertIsNotNone(save)
meta_graph_def = save.export_meta_graph()
collection_def = meta_graph_def.collection_def
# Validate custom graph_def.
graph_def_any = collection_def[constants.GRAPH_KEY].any_list.value
self.assertEquals(len(graph_def_any), 1)
graph_def = tf.GraphDef()
graph_def_any[0].Unpack(graph_def)
if clear_devices:
for node in compare_def.node:
node.device = ""
self.assertProtoEquals(compare_def, graph_def)
# Validate init_op.
init_ops = collection_def[constants.INIT_OP_KEY].node_list.value
self.assertEquals(len(init_ops), 1)
self.assertEquals(init_ops[0], "init_op")
# Validate signatures.
signatures_any = collection_def[constants.SIGNATURES_KEY].any_list.value
self.assertEquals(len(signatures_any), 1)
signatures = manifest_pb2.Signatures()
signatures_any[0].Unpack(signatures)
default_signature = signatures.default_signature
self.assertEqual(
default_signature.classification_signature.input.tensor_name, "v0:0")
bindings = signatures.named_signatures["generic"].generic_signature.map
self.assertEquals(bindings["logical_input_A"].tensor_name, "v0:0")
self.assertEquals(bindings["logical_input_B"].tensor_name, "v1:0")
read_foo_signature = (
signatures.named_signatures["foo"].regression_signature)
self.assertEquals(read_foo_signature.input.tensor_name, "v0:0")
self.assertEquals(read_foo_signature.output.tensor_name, "v1:0")
# Validate the assets.
assets_any = collection_def[constants.ASSETS_KEY].any_list.value
self.assertEquals(len(assets_any), 1)
asset = manifest_pb2.AssetFile()
assets_any[0].Unpack(asset)
assets_path = os.path.join(export_path,
constants.VERSION_FORMAT_SPECIFIER %
global_step, constants.ASSETS_DIRECTORY,
"hello42.txt")
asset_contents = gfile.GFile(assets_path).read()
self.assertEqual(asset_contents, "your data here")
self.assertEquals("hello42.txt", asset.filename)
self.assertEquals("filename42:0", asset.tensor_binding.tensor_name)
ignored_asset_path = os.path.join(export_path,
constants.VERSION_FORMAT_SPECIFIER %
global_step, constants.ASSETS_DIRECTORY,
"ignored.txt")
self.assertFalse(gfile.Exists(ignored_asset_path))
# Validate graph restoration.
if sharded:
save.restore(sess,
os.path.join(
export_path, constants.VERSION_FORMAT_SPECIFIER %
global_step, constants.VARIABLES_FILENAME_PATTERN))
else:
save.restore(sess,
os.path.join(
export_path, constants.VERSION_FORMAT_SPECIFIER %
global_step, constants.VARIABLES_FILENAME))
self.assertEqual(10, tf.get_collection("v")[0].eval())
self.assertEqual(20, tf.get_collection("v")[1].eval())
tf.get_collection(constants.INIT_OP_KEY)[0].run()
self.assertEqual(30, tf.get_collection("v")[2].eval())
def Export():
export_path = "/tmp/half_plus_two"
with tf.Session() as sess:
# Make model parameters a&b variables instead of constants to
# exercise the variable reloading mechanisms.
a = tf.Variable(0.5, name="a")
b = tf.Variable(2.0, name="b")
# Calculate, y = a*x + b
# here we use a placeholder 'x' which is fed at inference time.
x = tf.placeholder(tf.float32, name="x")
y = tf.add(tf.mul(a, x), b, name="y")
# Setup a standard Saver for our variables.
save = tf.train.Saver({"a": a, "b": b}, sharded=True)
# asset_path contains the base directory of assets used in training (e.g.
# vocabulary files).
original_asset_path = tf.constant("/tmp/original/export/assets")
# Ops reading asset files should reference the asset_path tensor
# which stores the original asset path at training time and the
# overridden assets directory at restore time.
asset_path = tf.Variable(original_asset_path,
name="asset_path",
trainable=False,
collections=[])
assign_asset_path = asset_path.assign(original_asset_path)
# CopyAssets is used as a callback during export to copy files to the
# given export directory.
def CopyAssets(export_path):
print "copying asset files to: %s" % export_path
# Use a fixed global step number.
global_step_tensor = tf.Variable(123, name="global_step")
# Create a RegressionSignature for our input and output.
signature = exporter.regression_signature(input_tensor=x, output_tensor=y)
# Create two filename assets and corresponding tensors.
# TODO(b/26254158) Consider adding validation of file existance as well as
# hashes (e.g. sha1) for consistency.
original_filename1 = tf.constant("hello1.txt")
filename1 = tf.Variable(original_filename1,
name="filename1",
trainable=False,
collections=[])
assign_filename1 = filename1.assign(original_filename1)
original_filename2 = tf.constant("hello2.txt")
filename2 = tf.Variable(original_filename2,
name="filename2",
trainable=False,
collections=[])
assign_filename2 = filename2.assign(original_filename2)
assets = {("hello1.txt", original_filename1),
("hello2.txt", original_filename2)}
# Init op contains a group of all variables that we assign.
init_op = tf.group(assign_asset_path, assign_filename1, assign_filename2)
# Run an export.
tf.initialize_all_variables().run()
export = exporter.Exporter(save)
export.init(sess.graph.as_graph_def(),
init_op=init_op,
default_graph_signature=signature,
assets=assets,
assets_callback=CopyAssets)
export.export(export_path, global_step_tensor, sess)
