def set_output_shape(self, source_node, IR_node):
sym_group = self.model.symbol.get_internals()
for sym in sym_group:
if source_node.name == sym.name:
arg_shape, output_shape, aux_shape = sym.infer_shape(
data=self.data_shape)
for idx in range(len(output_shape)):
output_list = list(output_shape[idx])
# transpose to channel last
if not self.data_format in MXNetParser.channels_last:
channel = output_list.pop(1)
output_list.append(channel)
if IR_node.op == "DataInput":
MXNetParser._copy_shape(IR_node,
[-1] + output_list[1:])
shape = graph_pb2.TensorShape()
for dim in output_list:
new_dim = shape.dim.add()
if dim == None:
new_dim.size = -1
else:
new_dim.size = dim
IR_node.attr["_output_shapes"].list.shape.extend([shape])
break
def _set_output_shape(source_node, IR_node):
shape = graph_pb2.TensorShape()
for dim in source_node.layer.output_shape:
new_dim = shape.dim.add()
new_dim.size = dim if dim else -1
IR_node.attr["_output_shapes"].list.shape.extend([shape])
def _copy_and_reop(source_node, IR_node, new_op=None):
if new_op == None: new_op = source_node.type
IR_node.name = source_node.name
IR_node.op = new_op
if '_output_shape' in source_node.layer['attr'].keys():
output_list = source_node.layer['attr']['_output_shape']
shape = graph_pb2.TensorShape()
for dim in output_list:
new_dim = shape.dim.add()
if dim == None:
new_dim.size = -1
else:
new_dim.size = int(dim)
IR_node.attr["_output_shape"].list.shape.extend([shape])
if 'shape' in source_node.layer['attr'].keys():
shape_list = source_node.layer['attr']['shape']
if not output_list == None:
for dim in shape_list:
new_dim = IR_node.attr["shape"].shape.dim.add()
if dim == None:
new_dim.size = -1
else:
new_dim.size = int(dim)
else:
IR_node.attr["shape"].shape.unknown_rank = True
def _set_output_shape(self, source_node, IR_node):
shape = graph_pb2.TensorShape()
layer_name = source_node.name
shape_pytorch = self.shape_dict[layer_name]
new_dim = shape.dim.add()
# (batch, C, H, W) & NHWC
if len(shape_pytorch) == 4:
if shape_pytorch[0] == 1:
new_dim.size = -1
else:
new_dim.size = shape_pytorch[0]
for index in [2, 3, 1]:
new_dim = shape.dim.add()
dim = shape_pytorch[index]
new_dim.size = dim if dim else -1
elif len(shape_pytorch) == 2:
if shape_pytorch[0] == 1:
new_dim.size = -1
else:
new_dim.size = shape_pytorch[0]
for _ in range(2):
new_dim = shape.dim.add()
new_dim.size = 1
new_dim = shape.dim.add()
dim = shape_pytorch[1]
new_dim.size = dim if dim else -1
IR_node.attr["_output_shapes"].list.shape.extend([shape])
def rename_InputLayer(self, source_node):
# only for training
IR_node = self.IR_graph.node.add()
# name, op
IR_node.name = source_node.name
IR_node.op = "DataInput"
graph_shape = graph_pb2.TensorShape()
coreml_node_layer = source_node.layer
new_dim = graph_shape.dim.add()
new_dim.size = -1
new_dim = graph_shape.dim.add()
new_dim.size = coreml_node_layer.type.imageType.width
new_dim = graph_shape.dim.add()
new_dim.size = coreml_node_layer.type.imageType.height
new_dim = graph_shape.dim.add()
if coreml_node_layer.type.imageType.colorSpace == 10:
new_dim.size = 2
elif coreml_node_layer.type.imageType.colorSpace == 20:
new_dim.size = 3
elif coreml_node_layer.type.imageType.colorSpace == 30:
new_dim.size = 3
else:
assert False
IR_node.attr["_output_shapes"].list.shape.extend([graph_shape])
# input edge
self.convert_inedge(source_node, IR_node)
# shape
# NHWC channel last
# in fact, here is NWHC
new_dim = IR_node.attr['shape'].shape.dim.add()
new_dim.size = -1
new_dim = IR_node.attr['shape'].shape.dim.add()
new_dim.size = coreml_node_layer.type.imageType.width
new_dim = IR_node.attr['shape'].shape.dim.add()
new_dim.size = coreml_node_layer.type.imageType.height
new_dim = IR_node.attr['shape'].shape.dim.add()
if coreml_node_layer.type.imageType.colorSpace == 10:
new_dim.size = 2
elif coreml_node_layer.type.imageType.colorSpace == 20:
new_dim.size = 3
elif coreml_node_layer.type.imageType.colorSpace == 30:
new_dim.size = 3
else:
assert False
def gen_Input(self):
IR_node = self.IR_graph.node.add()
IR_node.name = 'input'
IR_node.op = "DataInput"
for node in self.IR_graph.node:
if node.name in self.src_graph.input_layers:
node.input.append('input')
assert len(self.input_shape) == 4
new_dim = IR_node.attr["shape"].shape.dim.add()
if self.input_shape[0] == 1:
new_dim.size = -1
else:
new_dim.size = self.input_shape[0]
for index in [2, 3, 1]:
new_dim = IR_node.attr["shape"].shape.dim.add()
new_dim.size = self.input_shape[index]
shape = graph_pb2.TensorShape()
new_dim = shape.dim.add()
shape_pytorch = self.input_shape
if len(shape_pytorch) == 4:
if shape_pytorch[0] == 1:
new_dim.size = -1
else:
new_dim.size = shape_pytorch[0]
for index in [2, 3, 1]:
new_dim = shape.dim.add()
dim = shape_pytorch[index]
new_dim.size = dim if dim else -1
elif len(shape_pytorch) == 2:
if shape_pytorch[0] == 1:
new_dim.size = -1
else:
new_dim.size = shape_pytorch[0]
for _ in range(2):
new_dim = shape.dim.add()
new_dim.size = 1
new_dim = shape.dim.add()
dim = shape_pytorch[1]
new_dim.size = dim if dim else -1
IR_node.attr["_output_shapes"].list.shape.extend([shape])
def _set_output_shape(self, source_node, IR_node):
shape = graph_pb2.TensorShape()
layer_name = source_node.name
shape_pytorch = self.shape_dict[layer_name]
new_dim = shape.dim.add()
if not shape_pytorch:
print(
"Warning: Pytorch cannot inference outputshape of \"{}\" with operator \"{}\". Setting outputshape manually in json file is alternative ."
.format(source_node.name, source_node.type))
IR_node.attr["_output_shapes"].list.shape.extend([shape])
return
# (batch, C, H, W) & NHWC
if len(shape_pytorch) == 4:
if shape_pytorch[0] == 1:
new_dim.size = -1
else:
new_dim.size = shape_pytorch[0]
for index in [2, 3, 1]:
new_dim = shape.dim.add()
dim = shape_pytorch[index]
new_dim.size = dim if dim else -1
elif len(shape_pytorch) == 2:
if shape_pytorch[0] == 1:
new_dim.size = -1
else:
new_dim.size = shape_pytorch[0]
for _ in range(2):
new_dim = shape.dim.add()
new_dim.size = 1
new_dim = shape.dim.add()
dim = shape_pytorch[1]
new_dim.size = dim if dim else -1
IR_node.attr["_output_shapes"].list.shape.extend([shape])
def _set_output_shape(source_node, IR_node):
shape = graph_pb2.TensorShape()
source_node_layer = source_node.layer
layer_name = source_node_layer.output[0]
shape_coreml = CoremlParser.shape_dict[layer_name]
# (seq, batch, C, H, W) & NHWC
new_dim = shape.dim.add()
if shape_coreml[1] == 1:
new_dim.size = -1
else:
new_dim.size = shape_coreml[1]
for index in [3, 4, 2]:
new_dim = shape.dim.add()
dim = shape_coreml[index]
new_dim.size = dim if dim else -1
IR_node.attr["_output_shapes"].list.shape.extend([shape])
def list_to_shape(shape):
ret = graph_pb2.TensorShape()
for dim in shape:
new_dim = ret.dim.add()
new_dim.size = dim
return ret
def _set_output_shape(source_node, IR_node, output_shapes):
shape = graph_pb2.TensorShape()
for output_shape in output_shapes:
new_dim = shape.dim.add()
new_dim.size = output_shape
IR_node.attr["_output_shapes"].list.shape.extend([shape])
