def convert_to_mace_pb(model_file, input_node, output_node, dsp_mode):
"""
nnlib does not have batch norm, so use tensorflow optimizer to fold
batch norm with convolution. The fold optimization reorders ops, so
we sort ops first by topology.
"""
input_graph_def = tf.GraphDef()
with gfile.Open(model_file, "rb") as f:
data = f.read()
input_graph_def.ParseFromString(data)
input_graph_def = graph_util.sort_tf_graph(input_graph_def)
net_def = mace_pb2.NetDef()
with tf.Session() as session:
with session.graph.as_default() as graph:
tf.import_graph_def(input_graph_def, name="")
ops = graph.get_operations()
dsp_ops = DspOps()
resolved_ops = set()
# convert const node
unresolved_ops = [op for op in ops if op.type == 'Const']
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, net_def, output_node,
dsp_ops)
# convert op node
unresolved_ops = [op for op in ops if op.type != 'Const']
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, net_def, output_node,
dsp_ops)
add_output_node(net_def, output_node)
net_def = reverse_batch_to_space_and_biasadd(net_def)
net_def = fuse_quantize(net_def, input_node, output_node)
sorted_net_def = graph_util.sort_mace_graph(net_def, '__output__')
net_def_with_node_id = add_node_id(sorted_net_def)
dtype = mace_pb2.DT_FLOAT
final_net_def = add_input_output_info(net_def_with_node_id,
input_node, output_node,
graph, dtype)
arg = final_net_def.arg.add()
arg.name = 'dsp_mode'
arg.i = dsp_mode
return final_net_def
def convert_to_mace_pb(model_file, input_node, output_node, dsp_mode):
"""
nnlib does not have batch norm, so use tensorflow optimizer to fold
batch norm with convolution. The fold optimization reorders ops, so
we sort ops first by topology.
"""
input_graph_def = tf.GraphDef()
with gfile.Open(model_file, "rb") as f:
data = f.read()
input_graph_def.ParseFromString(data)
input_graph_def = graph_util.sort_tf_graph(input_graph_def)
net_def = mace_pb2.NetDef()
with tf.Session() as session:
with session.graph.as_default() as graph:
tf.import_graph_def(input_graph_def, name="")
ops = graph.get_operations()
dsp_ops = DspOps()
resolved_ops = set()
# convert const node
unresolved_ops = [op for op in ops if op.type == 'Const']
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, net_def, output_node,
dsp_ops)
# convert op node
unresolved_ops = [op for op in ops if op.type != 'Const']
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, net_def, output_node,
dsp_ops)
add_output_node(net_def, output_node)
net_def = reverse_batch_to_space_and_biasadd(net_def)
net_def = fuse_quantize(net_def, input_node, output_node)
sorted_net_def = graph_util.sort_mace_graph(net_def, '__output__')
net_def_with_node_id = add_node_id(sorted_net_def)
dtype = mace_pb2.DT_FLOAT
final_net_def = add_input_output_info(
net_def_with_node_id, input_node, output_node, graph, dtype)
arg = final_net_def.arg.add()
arg.name = 'dsp_mode'
arg.i = dsp_mode
return final_net_def
def run(self):
mace_check(
len(self._option.input_nodes) == 1
and len(self._option.output_nodes) == 1,
'dsp only support single input and output')
for tensor in self._model.tensors:
self._consts[tensor.name] = tensor
# convert op node
self.convert_ops()
self.add_input_output_node()
self._model = graph_util.sort_mace_graph(self._model, '__output__')
self.add_node_id()
return self._model
def run(self):
mace_check(len(self._option.input_nodes) == 1
and len(self._option.output_nodes) == 1,
'dsp only support single input and output')
for tensor in self._model.tensors:
self._consts[tensor.name] = tensor
# convert op node
self.convert_ops()
self.add_input_output_node()
if not self._option.check_nodes:
output_name = self._option.output_nodes.values()[0].name
else:
output_name = self._option.check_nodes.values()[0].name
output_name = normalize_name(output_name)
self._model = graph_util.sort_mace_graph(self._model, output_name)
self.add_node_id()
return self._model
def run(self):
ops = self._tf_graph.get_operations()
dsp_ops = DspOps()
resolved_ops = set()
mace_check(
len(self._option.input_nodes) == 1
and len(self._option.output_nodes) == 1,
'dsp only support single input and output')
input_node = self._option.input_nodes.values()[0].name
output_node = self._option.output_nodes.values()[0].name
# convert const node
unresolved_ops = [op for op in ops if op.type == 'Const']
with tf.Session() as session:
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, self._mace_net_def,
dsp_ops)
# convert op node
unresolved_ops = [op for op in ops if op.type != 'Const']
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, self._mace_net_def,
dsp_ops)
add_output_node(self._mace_net_def, output_node)
net_def = reverse_batch_to_space_and_biasadd(self._mace_net_def)
net_def = fuse_quantize(net_def)
sorted_net_def = graph_util.sort_mace_graph(net_def, '__output__')
net_def_with_node_id = add_node_id(sorted_net_def)
dtype = mace_pb2.DT_FLOAT
final_net_def = add_input_output_info(net_def_with_node_id,
input_node, output_node,
self._tf_graph, dtype)
return final_net_def
def run(self):
ops = self._tf_graph.get_operations()
dsp_ops = DspOps()
resolved_ops = set()
mace_check(len(self._option.input_nodes) == 1
and len(self._option.output_nodes) == 1,
'dsp only support single input and output')
input_node = self._option.input_nodes.values()[0].name
output_node = self._option.output_nodes.values()[0].name
# convert const node
unresolved_ops = [op for op in ops if op.type == 'Const']
with tf.Session() as session:
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, self._mace_net_def,
dsp_ops)
# convert op node
unresolved_ops = [op for op in ops if op.type != 'Const']
while len(unresolved_ops) > 0:
convert_ops(unresolved_ops, resolved_ops, self._mace_net_def,
dsp_ops)
add_output_node(self._mace_net_def, output_node)
net_def = reverse_batch_to_space_and_biasadd(self._mace_net_def)
net_def = fuse_quantize(net_def)
sorted_net_def = graph_util.sort_mace_graph(net_def, '__output__')
net_def_with_node_id = add_node_id(sorted_net_def)
dtype = mace_pb2.DT_FLOAT
final_net_def = add_input_output_info(
net_def_with_node_id, input_node, output_node,
self._tf_graph, dtype)
return final_net_def
