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) if self.local != "." else dir_name
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 = tf_loader.from_checkpoint(
model_path, input_names, outputs)
elif self.model_type in ["saved_model"]:
loaded = tf_loader.from_saved_model(
model_path,
input_names,
outputs,
self.tag,
self.signatures,
self.concrete_function,
self.large_model,
return_concrete_func=self.large_model)
if self.large_model:
# Must maintain ref to imported since concrete_func uses weak refs
# pylint: disable=unused-variable
graph_def, input_names, outputs, concrete_func, imported = loaded
else:
graph_def, input_names, outputs = loaded
elif self.model_type in ["keras"]:
graph_def, input_names, outputs = tf_loader.from_keras(
model_path, input_names, outputs)
else:
graph_def, input_names, outputs = tf_loader.from_graphdef(
model_path, input_names, outputs)
if utils.is_debug_mode():
utils.save_protobuf(
os.path.join(TEMP_DIR, name + "_after_tf_optimize.pb"),
graph_def)
if self.large_model:
inputs = {}
for k in input_names:
v = self.input_names[k]
inputs[k.split(":")[0]] = tf.constant(self.make_input(v))
tf_func = tf.function(concrete_func)
logger.info("Running TF")
tf_results_d = tf_func(**inputs)
if self.structured_outputs is None:
tf_results = list(tf_results_d.values())
else:
tf_results = [
tf_results_d[output] for output in self.structured_outputs
]
if self.perf:
logger.info("Running TF perf")
start = time.time()
for _ in range(PERFITER):
_ = concrete_func(**inputs)
self.tf_runtime = time.time() - start
logger.info("TensorFlow OK")
inputs = {}
shape_override = {}
tf_reset_default_graph()
from tf2onnx.tf_utils import compress_graph_def
const_node_values = None
with tf.Graph().as_default() as tf_graph:
if self.large_model:
const_node_values = compress_graph_def(graph_def)
tf.import_graph_def(graph_def, name='')
with tf_session(graph=tf_graph) 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")
elif not self.large_model:
tf_results = self.run_tensorflow(sess, inputs)
logger.info("TensorFlow OK")
model_proto = None
if self.skip_conversion:
if self.large_model:
external_tensor_storage = ExternalTensorStorage()
model_proto = utils.model_proto_from_zip(
self.converted_model, external_tensor_storage)
else:
external_tensor_storage = None
model_proto = utils.model_proto_from_file(self.converted_model)
logger.info("ONNX loaded from file")
else:
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(),
const_node_values=const_node_values)
onnx_graph = optimizer.optimize_graph(onnx_graph)
print("ONNX", onnx_graph.dump_node_statistics())
external_tensor_storage = ExternalTensorStorage(
) if self.large_model else None
model_proto = onnx_graph.make_model(
"converted from tf2onnx",
external_tensor_storage=external_tensor_storage)
logger.info("To_ONNX, OK")
if onnx_file:
self.create_onnx_file(name, model_proto, inputs, onnx_file,
external_tensor_storage)
if self.converted_model:
if self.large_model:
utils.save_onnx_zip(self.converted_model, model_proto,
external_tensor_storage)
else:
utils.save_protobuf(self.converted_model, model_proto)
logger.info("Created %s", self.converted_model)
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,
external_tensor_storage)
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="onnxruntime",
onnx_file=None,
opset=None,
extra_opset=None,
perf=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) if self.local != "." else dir_name
else:
model_path = self.local
logger.info("Load model from %s", model_path)
input_names = list(self.input_names.keys())
initialized_tables = {}
outputs = self.output_names
tflite_path = None
to_rename = None
if self.model_type in ["checkpoint"]:
graph_def, input_names, outputs = tf_loader.from_checkpoint(
model_path, input_names, outputs)
elif self.model_type in ["saved_model"]:
loaded = tf_loader.from_saved_model(
model_path,
None,
None,
self.tag,
self.signatures,
self.concrete_function,
self.large_model,
return_concrete_func=not self.run_tf_frozen,
return_initialized_tables=True,
return_tensors_to_rename=True)
if not self.run_tf_frozen:
# Must maintain ref to imported since concrete_func uses weak refs
# pylint: disable=unused-variable
graph_def, input_names, outputs, concrete_func, imported, initialized_tables, to_rename = loaded
else:
graph_def, input_names, outputs, initialized_tables, to_rename = loaded
elif self.model_type in ["keras"]:
graph_def, input_names, outputs = tf_loader.from_keras(
model_path, input_names, outputs)
elif self.model_type in ["tflite"]:
tflite_path = model_path
graph_def = None
else:
graph_def, input_names, outputs = tf_loader.from_graphdef(
model_path, input_names, outputs)
if utils.is_debug_mode():
utils.save_protobuf(
os.path.join(TEMP_DIR, name + "_after_tf_optimize.pb"),
graph_def)
if tflite_path is not None:
inputs = {}
for k in input_names:
v = self.input_names[k]
inputs[k] = self.make_input(v)
interpreter = tf.lite.Interpreter(tflite_path)
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
input_name_to_index = {
n['name'].split(':')[0]: n['index']
for n in input_details
}
for k, v in inputs.items():
interpreter.resize_tensor_input(input_name_to_index[k],
v.shape)
interpreter.allocate_tensors()
def run_tflite():
for k, v in inputs.items():
interpreter.set_tensor(input_name_to_index[k], v)
interpreter.invoke()
result = [
interpreter.get_tensor(output['index'])
for output in output_details
]
return result
tf_results = run_tflite()
if self.perf:
logger.info("Running TFLite perf")
n = 0
start = time.time()
stop = start + PERF_TIME
while time.time() < stop:
for _ in range(PERF_STEP):
_ = run_tflite()
n += PERF_STEP
self.tf_runtime = 1000 * (time.time() - start) / n
logger.info("TFLite perf {:.2f}ms/inference, n={}".format(
self.tf_runtime, n))
logger.info("TFLite OK")
if not self.run_tf_frozen:
inputs = {}
for k in input_names:
v = self.input_names[k]
inputs[k.split(":")[0]] = tf.constant(self.make_input(v))
tf_func = tf.function(concrete_func)
logger.info("Running TF")
tf_results_d = tf_func(**inputs)
# If there is only a single output a dict might not be returned
if isinstance(tf_results_d, tf.Tensor):
tf_results = [tf_results_d]
else:
tf_results = [
tf_results_d[k] for k in sorted(tf_results_d.keys())
]
tf_results = [tf_res.numpy() for tf_res in tf_results]
if self.perf:
logger.info("Running TF perf")
n = 0
start = time.time()
stop = start + PERF_TIME
if self.tf_profile is not None:
tf.profiler.experimental.start(self.tf_profile)
while time.time() < stop:
for _ in range(PERF_STEP):
_ = concrete_func(**inputs)
n += PERF_STEP
if self.tf_profile is not None:
tf.profiler.experimental.stop()
self.tf_runtime = 1000 * (time.time() - start) / n
logger.info("TF perf {:.2f}ms/inference, n={}".format(
self.tf_runtime, n))
logger.info("TensorFlow OK")
shape_override = {}
const_node_values = None
tf_graph = None
if graph_def is not None:
inputs = {}
tf_reset_default_graph()
with tf.Graph().as_default() as tf_graph:
from tf2onnx.tf_utils import compress_graph_def
if self.large_model:
const_node_values = compress_graph_def(graph_def)
tf.import_graph_def(graph_def, name='')
with tf_session(graph=tf_graph) 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:
if expected_dtype == "string":
inputs[k] = self.make_input(v).astype(
np.str).astype(np.object)
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")
elif self.run_tf_frozen:
if self.tf_profile is not None:
tf.profiler.experimental.start(self.tf_profile)
tf_results = self.run_tensorflow(sess, inputs)
if self.tf_profile is not None:
tf.profiler.experimental.stop()
logger.info("TensorFlow OK")
tf_graph = sess.graph
model_proto = None
if self.skip_conversion:
if self.large_model:
external_tensor_storage = ExternalTensorStorage()
model_proto = utils.model_proto_from_zip(
self.converted_model, external_tensor_storage)
else:
external_tensor_storage = None
model_proto = utils.model_proto_from_file(self.converted_model)
logger.info("ONNX loaded from file")
else:
try:
# convert model to onnx
onnx_graph = self.to_onnx(
tf_graph,
opset=opset,
extra_opset=extra_opset,
shape_override=shape_override,
input_names=inputs.keys(),
const_node_values=const_node_values,
initialized_tables=initialized_tables,
tflite_path=tflite_path,
tensors_to_rename=to_rename)
onnx_graph = optimizer.optimize_graph(onnx_graph)
print("ONNX", onnx_graph.dump_node_statistics())
external_tensor_storage = ExternalTensorStorage(
) if self.large_model else None
model_proto = onnx_graph.make_model(
"converted from tf2onnx",
external_tensor_storage=external_tensor_storage)
logger.info("To_ONNX, OK")
if onnx_file:
self.create_onnx_file(name, model_proto, inputs, onnx_file,
external_tensor_storage)
if self.converted_model:
if self.large_model:
utils.save_onnx_zip(self.converted_model, model_proto,
external_tensor_storage)
else:
utils.save_protobuf(self.converted_model, model_proto)
logger.info("Created %s", self.converted_model)
except Exception:
logger.error("To_ONNX FAIL", exc_info=1)
return False
try:
onnx_results = None
if backend == "onnxruntime":
if to_rename is None:
struc_outputs = self.output_names
else:
struc_outputs = [
to_rename.get(k, k) for k in self.output_names
]
onnx_results = self.run_onnxruntime(name, model_proto, inputs,
struc_outputs,
external_tensor_storage)
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):
good_cnt = np.count_nonzero(
np.isclose(tf_res,
onnx_res,
rtol=self.rtol,
atol=self.atol))
bad_cnt = tf_res.size - good_cnt
if bad_cnt > self.ptol / 100 * tf_res.size:
# Prints a nice error message with stats
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