def convert_to_onnx(model_filepath, onnx_filepath, output_node_names):
"""
Convert the model to an ONNX file, which can in turn be used for TensorRT inference
Arguments:
model_filepath: the path to the frozen .pb file
onnx_filepath: the path where the ONNX file should be saved
output_node_names: list of output node names
"""
# tf2onnx expects the node names in the format "input/output_node_name:port_id".
# Hence, we should provide the port ID before conversion.
input_node_names = [kInputName + ":0"]
output_node_names = list(map(lambda x: x + ":0", output_node_names))
# Use in-built function from tf2onnx to import the graph and optimize for conversion
graph_def, inputs, outputs = loader.from_graphdef(model_filepath,
input_node_names,
output_node_names)
graph_def = tf_optimize(input_node_names, output_node_names, graph_def,
False)
with tf.Graph().as_default() as default_graph:
tf.import_graph_def(graph_def, name='')
# Convert to ONNX
with tf.Session(graph=default_graph):
onnx_graph = process_tf_graph(default_graph,
opset=8,
input_names=inputs,
output_names=outputs)
onnx_graph = optimizer.optimize_graph(onnx_graph)
onnx_model = onnx_graph.make_model("segmentation_onnx_model")
# Save the ONNX model to disk
utils.save_protobuf(onnx_filepath, onnx_model)
def main():
args = get_args()
logging.basicConfig(level=logging.get_verbosity_level(args.verbose))
if args.debug:
utils.set_debug_mode(True)
logger = logging.getLogger(constants.TF2ONNX_PACKAGE_NAME)
extra_opset = args.extra_opset or []
custom_ops = {}
if args.custom_ops:
# default custom ops for tensorflow-onnx are in the "tf" namespace
custom_ops = {op: (default_custom_op_handler, []) for op in args.custom_ops.split(",")}
extra_opset.append(constants.TENSORFLOW_OPSET)
# get the frozen tensorflow model from graphdef, checkpoint or saved_model.
if args.graphdef:
graph_def, inputs, outputs = loader.from_graphdef(args.graphdef, args.inputs, args.outputs)
model_path = args.graphdef
if args.checkpoint:
graph_def, inputs, outputs = loader.from_checkpoint(args.checkpoint, args.inputs, args.outputs)
model_path = args.checkpoint
if args.saved_model:
graph_def, inputs, outputs = loader.from_saved_model(
args.saved_model, args.inputs, args.outputs, args.signature_def)
model_path = args.saved_model
if args.verbose:
logger.info("inputs: %s", inputs)
logger.info("outputs: %s", outputs)
# todo: consider to enable const folding by default?
graph_def = tf_optimize(inputs, outputs, graph_def, args.fold_const)
with tf.Graph().as_default() as tf_graph:
tf.import_graph_def(graph_def, name='')
with tf.Session(graph=tf_graph):
g = process_tf_graph(tf_graph,
continue_on_error=args.continue_on_error,
target=args.target,
opset=args.opset,
custom_op_handlers=custom_ops,
extra_opset=extra_opset,
shape_override=args.shape_override,
input_names=inputs,
output_names=outputs,
inputs_as_nchw=args.inputs_as_nchw)
onnx_graph = optimizer.optimize_graph(g)
model_proto = onnx_graph.make_model("converted from {}".format(model_path))
# write onnx graph
logger.info("")
logger.info("Successfully converted TensorFlow model %s to ONNX", model_path)
if args.output:
utils.save_protobuf(args.output, model_proto)
logger.info("ONNX model is saved at %s", args.output)
else:
logger.info("To export ONNX model to file, please run with `--output` option")
def convert_tf2onnx(model,
output,
inputs,
outputs,
signature_def=None,
opset=7):
import tensorflow as tf
from tf2onnx.tfonnx import process_tf_graph, tf_optimize
from tf2onnx import constants, loader, logging, utils, optimizer
logger = logging.getLogger(constants.TF2ONNX_PACKAGE_NAME)
if "pb" in model:
graph_def, inputs, outputs = loader.from_graphdef(
model, inputs, outputs)
elif "meta" in model:
graph_def, inputs, outputs = loader.from_checkpoint(
model, inputs, outputs)
elif "saved_model" in model:
graph_def, inputs, outputs = loader.from_saved_model(
model, inputs, outputs, signature_def)
graph_def = tf_optimize(inputs, outputs, graph_def, None)
with tf.Graph().as_default() as tf_graph:
tf.import_graph_def(graph_def, name='')
with tf.Session(graph=tf_graph):
g = process_tf_graph(tf_graph,
opset=opset,
input_names=inputs,
output_names=outputs)
onnx_graph = optimizer.optimize_graph(g)
model_proto = onnx_graph.make_model("converted from {}".format(model))
# write onnx graph
logger.info("")
logger.info("Successfully converted TensorFlow model %s to ONNX", model)
utils.save_protobuf(output, model_proto)
logger.info("ONNX model is saved at %s", output)
def run_test(self, name, backend="caffe2", onnx_file=None, opset=None, extra_opset=None,
perf=None, fold_const=None):
"""Run complete test against backend."""
logger.info("===================================")
logger.info("Running %s", name)
self.perf = perf
# get the model
if self.url:
_, dir_name = self.download_file()
model_path = os.path.join(dir_name, self.local)
else:
model_path = self.local
dir_name = os.path.dirname(self.local)
logger.info("Downloaded to %s", model_path)
input_names = list(self.input_names.keys())
outputs = self.output_names
if self.model_type in ["checkpoint"]:
graph_def, input_names, outputs = loader.from_checkpoint(model_path, input_names, outputs)
elif self.model_type in ["saved_model"]:
graph_def, input_names, outputs = loader.from_saved_model(model_path, input_names, outputs)
else:
graph_def, input_names, outputs = loader.from_graphdef(model_path, input_names, outputs)
# create the input data
inputs = {}
for k, v in self.input_names.items():
if k not in input_names:
continue
if isinstance(v, six.text_type) and v.startswith("np."):
inputs[k] = eval(v) # pylint: disable=eval-used
else:
inputs[k] = self.make_input(v)
if self.more_inputs:
for k, v in self.more_inputs.items():
inputs[k] = v
graph_def = tf2onnx.tfonnx.tf_optimize(inputs.keys(), self.output_names, graph_def, fold_const)
if utils.is_debug_mode():
utils.save_protobuf(os.path.join(TEMP_DIR, name + "_after_tf_optimize.pb"), graph_def)
shape_override = {}
g = tf.import_graph_def(graph_def, name='')
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True), graph=g) as sess:
# fix inputs if needed
for k in inputs.keys(): # pylint: disable=consider-iterating-dictionary
t = sess.graph.get_tensor_by_name(k)
dtype = tf.as_dtype(t.dtype).name
v = inputs[k]
if dtype != v.dtype:
logger.warning("input dtype doesn't match tensorflow's")
inputs[k] = np.array(v, dtype=dtype)
if self.force_input_shape:
for k, v in inputs.items():
shape_override[k] = list(v.shape)
# run the model with tensorflow
if self.skip_tensorflow:
logger.info("TensorFlow SKIPPED")
else:
tf_results = self.run_tensorflow(sess, inputs)
logger.info("TensorFlow OK")
model_proto = None
try:
# convert model to onnx
onnx_graph = self.to_onnx(sess.graph, opset=opset, extra_opset=extra_opset,
shape_override=shape_override, input_names=inputs.keys())
model_proto = onnx_graph.make_model("converted from tf2onnx")
model_proto = optimizer.optimize_graph(onnx_graph).make_model("optimized")
logger.info("To_ONNX, OK")
if utils.is_debug_mode():
onnx_graph.dump_graph()
if onnx_file:
self.create_onnx_file(name, model_proto, inputs, onnx_file)
except Exception:
logger.error("To_ONNX FAIL", exc_info=1)
try:
onnx_results = None
if backend == "caffe2":
onnx_results = self.run_caffe2(name, model_proto, inputs)
elif backend == "onnxmsrtnext":
onnx_results = self.run_onnxmsrtnext(name, model_proto, inputs)
elif backend == "onnxruntime":
onnx_results = self.run_onnxruntime(name, model_proto, inputs)
else:
raise ValueError("unknown backend")
logger.info("Run_ONNX OK")
try:
if self.skip_tensorflow:
logger.info("Results: skipped tensorflow")
else:
if self.check_only_shape:
for tf_res, onnx_res in zip(tf_results, onnx_results):
np.testing.assert_array_equal(tf_res.shape, onnx_res.shape)
else:
for tf_res, onnx_res in zip(tf_results, onnx_results):
np.testing.assert_allclose(tf_res, onnx_res, rtol=self.rtol, atol=self.atol)
logger.info("Results: OK")
return True
except Exception:
logger.error("Results", exc_info=1)
except Exception:
logger.error("Run_ONNX FAIL", exc_info=1)
return False
def run_test(self,
name,
backend="caffe2",
debug=False,
onnx_file=None,
opset=None,
extra_opset=None,
perf=None,
fold_const=None):
"""Run complete test against backend."""
print(name)
self.perf = perf
# get the model
if self.url:
_, dir_name = self.download_file()
model_path = os.path.join(dir_name, self.local)
else:
model_path = self.local
dir_name = os.path.dirname(self.local)
print("\tdownloaded", model_path)
inputs = list(self.input_names.keys())
outputs = self.output_names
if self.model_type in ["checkpoint"]:
graph_def, inputs, outputs = loader.from_checkpoint(
model_path, inputs, outputs)
elif self.model_type in ["saved_model"]:
graph_def, inputs, outputs = loader.from_saved_model(
model_path, inputs, outputs)
else:
graph_def, inputs, outputs = loader.from_graphdef(
model_path, inputs, outputs)
# create the input data
inputs = {}
for k, v in self.input_names.items():
if isinstance(v, six.text_type) and v.startswith("np."):
inputs[k] = eval(v) # pylint: disable=eval-used
else:
inputs[k] = self.make_input(v)
if self.more_inputs:
for k, v in self.more_inputs.items():
inputs[k] = v
graph_def = tf2onnx.tfonnx.tf_optimize(inputs.keys(),
self.output_names, graph_def,
fold_const)
shape_override = {}
g = tf.import_graph_def(graph_def, name='')
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True),
graph=g) as sess:
# fix inputs if needed
for k in inputs.keys(): # pylint: disable=consider-iterating-dictionary
t = sess.graph.get_tensor_by_name(k)
dtype = tf.as_dtype(t.dtype).name
v = inputs[k]
if dtype != v.dtype:
log.warning("input dtype doesn't match tensorflow's")
inputs[k] = np.array(v, dtype=dtype)
if self.force_input_shape:
for k, v in inputs.items():
shape_override[k] = list(v.shape)
# run the model with tensorflow
if self.skip_tensorflow:
print("\ttensorflow", "SKIPPED")
else:
tf_results = self.run_tensorflow(sess, inputs)
print("\ttensorflow", "OK")
model_proto = None
try:
# convert model to onnx
onnx_graph = self.to_onnx(sess.graph,
opset=opset,
extra_opset=extra_opset,
shape_override=shape_override,
input_names=inputs.keys())
model_proto = onnx_graph.make_model("converted from tf2onnx")
new_model_proto = optimizer.optimize_graph(
onnx_graph, debug=debug).make_model("optimized")
if new_model_proto:
model_proto = new_model_proto
else:
print(
"\tNON-CRITICAL, optimizers are not applied successfully"
)
print("\tto_onnx", "OK")
if debug:
onnx_graph.dump_graph()
if onnx_file:
self.create_onnx_file(name, model_proto, inputs, onnx_file)
except Exception as ex:
tb = traceback.format_exc()
print("\tto_onnx", "FAIL", ex, tb)
try:
onnx_results = None
if backend == "caffe2":
onnx_results = self.run_caffe2(name, model_proto, inputs)
elif backend == "onnxmsrtnext":
onnx_results = self.run_onnxmsrtnext(name, model_proto, inputs)
elif backend == "onnxruntime":
onnx_results = self.run_onnxruntime(name, model_proto, inputs)
else:
raise ValueError("unknown backend")
print("\trun_onnx OK")
try:
if self.skip_tensorflow:
print("\tResults: skipped tensorflow")
else:
if self.check_only_shape:
for tf_res, onnx_res in zip(tf_results, onnx_results):
np.testing.assert_array_equal(
tf_res.shape, onnx_res.shape)
else:
for tf_res, onnx_res in zip(tf_results, onnx_results):
np.testing.assert_allclose(tf_res,
onnx_res,
rtol=self.rtol,
atol=self.atol)
print("\tResults: OK")
return True
except Exception as ex:
print("\tResults: ", ex)
except Exception as ex:
print("\trun_onnx", "FAIL", ex)
return False
def main():
args = get_args()
# override unknown dimensions from -1 to 1 (aka batchsize 1) since not every runtime does
# support unknown dimensions.
utils.ONNX_UNKNOWN_DIMENSION = args.unknown_dim
if args.custom_ops:
# default custom ops for tensorflow-onnx are in the "tf" namespace
custom_ops = {
op: (default_custom_op_handler, [])
for op in args.custom_ops.split(",")
}
extra_opset = [helper.make_opsetid(_TENSORFLOW_DOMAIN, 1)]
else:
custom_ops = {}
extra_opset = None
# get the frozen tensorflow model from graphdef, checkpoint or saved_model.
if args.graphdef:
graph_def, inputs, outputs = loader.from_graphdef(
args.graphdef, args.inputs, args.outputs)
model_path = args.graphdef
if args.checkpoint:
graph_def, inputs, outputs = loader.from_checkpoint(
args.checkpoint, args.inputs, args.outputs)
model_path = args.checkpoint
if args.saved_model:
graph_def, inputs, outputs = loader.from_saved_model(
args.saved_model, args.inputs, args.outputs)
model_path = args.saved_model
# todo: consider to enable const folding by default?
graph_def = tf_optimize(inputs, outputs, graph_def, args.fold_const)
with tf.Graph().as_default() as tf_graph:
tf.import_graph_def(graph_def, name='')
with tf.Session(graph=tf_graph):
g = process_tf_graph(tf_graph,
continue_on_error=args.continue_on_error,
verbose=args.verbose,
target=args.target,
opset=args.opset,
custom_op_handlers=custom_ops,
extra_opset=extra_opset,
shape_override=args.shape_override,
input_names=inputs,
output_names=outputs,
inputs_as_nchw=args.inputs_as_nchw)
model_proto = g.make_model("converted from {}".format(model_path))
new_model_proto = GraphUtil.optimize_model_proto(model_proto)
if new_model_proto:
model_proto = new_model_proto
else:
print("NON-CRITICAL, optimizers are not applied successfully")
# write onnx graph
if args.output:
utils.save_protobuf(args.output, model_proto)
print("\nComplete successfully, the onnx model is generated at " +
args.output)
def run_test(self, name, backend="caffe2", onnx_file=None, opset=None, extra_opset=None,
perf=None, fold_const=None):
"""Run complete test against backend."""
self.perf = perf
# get the model
if self.url:
_, dir_name = self.download_model()
logger.info("Downloaded to %s", dir_name)
model_path = os.path.join(dir_name, self.local)
else:
model_path = self.local
logger.info("Load model from %s", model_path)
input_names = list(self.input_names.keys())
outputs = self.output_names
if self.model_type in ["checkpoint"]:
graph_def, input_names, outputs = loader.from_checkpoint(model_path, input_names, outputs)
elif self.model_type in ["saved_model"]:
graph_def, input_names, outputs = loader.from_saved_model(model_path, input_names, outputs)
else:
graph_def, input_names, outputs = loader.from_graphdef(model_path, input_names, outputs)
# remove unused input names
input_names = list(set(input_names).intersection(self.input_names.keys()))
graph_def = tf2onnx.tfonnx.tf_optimize(input_names, self.output_names, graph_def, fold_const)
if utils.is_debug_mode():
utils.save_protobuf(os.path.join(TEMP_DIR, name + "_after_tf_optimize.pb"), graph_def)
inputs = {}
shape_override = {}
g = tf.import_graph_def(graph_def, name='')
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True), graph=g) as sess:
# create the input data
for k in input_names:
v = self.input_names[k]
t = sess.graph.get_tensor_by_name(k)
expected_dtype = tf.as_dtype(t.dtype).name
if isinstance(v, six.text_type) and v.startswith("np."):
np_value = eval(v) # pylint: disable=eval-used
if expected_dtype != np_value.dtype:
logger.warning("dtype mismatch for input %s: expected=%s, actual=%s", k, expected_dtype,
np_value.dtype)
inputs[k] = np_value.astype(expected_dtype)
else:
inputs[k] = self.make_input(v).astype(expected_dtype)
if self.force_input_shape:
for k, v in inputs.items():
shape_override[k] = list(v.shape)
# run the model with tensorflow
if self.skip_tensorflow:
logger.info("TensorFlow SKIPPED")
else:
tf_results = self.run_tensorflow(sess, inputs)
logger.info("TensorFlow OK")
model_proto = None
try:
# convert model to onnx
onnx_graph = self.to_onnx(sess.graph, opset=opset, extra_opset=extra_opset,
shape_override=shape_override, input_names=inputs.keys())
onnx_graph = optimizer.optimize_graph(onnx_graph)
model_proto = onnx_graph.make_model("converted from tf2onnx")
logger.info("To_ONNX, OK")
if onnx_file:
self.create_onnx_file(name, model_proto, inputs, onnx_file)
except Exception:
logger.error("To_ONNX FAIL", exc_info=1)
return False
try:
onnx_results = None
if backend == "caffe2":
onnx_results = self.run_caffe2(name, model_proto, inputs)
elif backend == "onnxruntime":
onnx_results = self.run_onnxruntime(name, model_proto, inputs)
else:
raise ValueError("unknown backend")
logger.info("Run_ONNX OK")
try:
if self.skip_tensorflow:
logger.info("Results: skipped tensorflow")
else:
if self.check_only_shape:
for tf_res, onnx_res in zip(tf_results, onnx_results):
np.testing.assert_array_equal(tf_res.shape, onnx_res.shape)
else:
for tf_res, onnx_res in zip(tf_results, onnx_results):
np.testing.assert_allclose(tf_res, onnx_res, rtol=self.rtol, atol=self.atol)
logger.info("Results: OK")
return True
except Exception:
logger.error("Results", exc_info=1)
except Exception:
logger.error("Run_ONNX FAIL", exc_info=1)
return False
