def update_param_to_nndct(self, node, param_name, param_data):
for param_type, tensor in node.op.params.items():
if tensor.name == param_name:
if node.op.type in [
NNDCT_OP.CONVTRANSPOSE2D, NNDCT_OP.CONVTRANSPOSE3D
] and param_type == node.op.ParamName.WEIGHTS:
param_data = np.copy(param_data).swapaxes(1, 0)
param_data = np.ascontiguousarray(param_data)
if node.op.type in [
NNDCT_OP.DEPTHWISE_CONV2D, NNDCT_OP.DEPTHWISE_CONV3D
] and param_type == node.op.ParamName.WEIGHTS:
in_channels = node.node_config("in_channels")
out_channels = node.node_config("out_channels")
kernel_size = node.node_config("kernel_size")
channel_mutiplier = int(out_channels / in_channels)
param_data = param_data.reshape(
(channel_mutiplier, in_channels, *kernel_size))
if node.op.type in [
NNDCT_OP.DEPTHWISE_CONVTRANSPOSE2D,
NNDCT_OP.DEPTHWISE_CONVTRANSPOSE3D
] and param_type == node.op.ParamName.WEIGHTS:
in_channels = node.node_config("in_channels")
out_channels = node.node_config("out_channels")
kernel_size = node.node_config("kernel_size")
channel_mutiplier = int(out_channels / in_channels)
param_data = param_data.reshape(
(in_channels, channel_mutiplier, *kernel_size))
param_data = np.copy(param_data).swapaxes(0, 1)
param_data = np.ascontiguousarray(param_data)
tensor.from_ndarray(param_data)
tensor_util.convert_parameter_tensor_format(
tensor, FrameworkType.TORCH, FrameworkType.NNDCT)
def update_param_to_nndct(self, node, param_name, param_data):
for param_type, tensor in node.op.params.items():
if tensor.name == param_name:
if node.op.type == NNDCT_OP.CONVTRANSPOSE2D:
if param_type == node.op.ParamName.WEIGHTS:
param_data = np.copy(param_data).transpose(1, 0, 2, 3)
if node.op.type == NNDCT_OP.DEPTHWISE_CONV2D and param_type == node.op.ParamName.WEIGHTS:
in_channels = node.node_config("in_channels")
out_channels = node.node_config("out_channels")
kernel_size = node.node_config("kernel_size")
channel_mutiplier = int(out_channels / in_channels)
param_data = param_data.reshape(
(channel_mutiplier, in_channels, *kernel_size))
tensor.from_ndarray(param_data)
tensor_util.convert_parameter_tensor_format(
tensor, key_names.FrameworkType.TORCH,
key_names.FrameworkType.NNDCT)
def update_param_to_nndct(self, node, param_name, param_data):
for param_type, tensor in node.op.params.items():
if tensor.name == param_name:
if node.op.type in [
NNDCT_OP.CONVTRANSPOSE2D, NNDCT_OP.CONVTRANSPOSE3D
] and param_type == node.op.ParamName.WEIGHTS:
param_data = np.copy(param_data).swapaxes(1, 0)
param_data = np.ascontiguousarray(param_data)
if node.op.type in [
NNDCT_OP.DEPTHWISE_CONV2D, NNDCT_OP.DEPTHWISE_CONV3D
] and param_type == node.op.ParamName.WEIGHTS:
in_channels = node.node_config("in_channels")
out_channels = node.node_config("out_channels")
kernel_size = node.node_config("kernel_size")
channel_mutiplier = int(out_channels / in_channels)
param_data = param_data.reshape(
(channel_mutiplier, in_channels, *kernel_size))
if node.op.type in [
NNDCT_OP.DEPTHWISE_CONVTRANSPOSE2D,
NNDCT_OP.DEPTHWISE_CONVTRANSPOSE3D
] and param_type == node.op.ParamName.WEIGHTS:
in_channels = node.node_config("in_channels")
out_channels = node.node_config("out_channels")
kernel_size = node.node_config("kernel_size")
channel_mutiplier = int(out_channels / in_channels)
param_data = param_data.reshape(
(in_channels, channel_mutiplier, *kernel_size))
param_data = np.copy(param_data).swapaxes(0, 1)
param_data = np.ascontiguousarray(param_data)
origin_shape = tensor.shape
tensor.from_ndarray(param_data)
tensor_util.convert_parameter_tensor_format(
tensor, key_names.FrameworkType.TORCH,
key_names.FrameworkType.NNDCT)
NndctScreenLogger().check(
f"The shape of data '{tensor.shape}' must be consistent with that of original data '{origin_shape}' for {tensor.name}",
origin_shape == tensor.shape)
def param_to_torch_format(tensor):
tu.convert_parameter_tensor_format(tensor, FrameworkType.NNDCT,
FrameworkType.TORCH)
def param_to_nndct_format(tensor):
tu.convert_parameter_tensor_format(tensor, FrameworkType.TORCH,
FrameworkType.NNDCT)
def nndct_param_to_tf(tensor):
return convert_parameter_tensor_format(tensor, 'nndct', 'tensorflow')
def tf_param_to_nndct(tensor):
return convert_parameter_tensor_format(tensor, 'tensorflow', 'nndct')
