def _process_parameter(self):
global_block = self.inference_program.global_block()
param_attrs = self.config.desc.extra_info.map.data['param_attrs']
for key, param_attr in param_attrs.map.data.items():
param = from_flexible_data_to_param(param_attr)
param['name'] = HUB_VAR_PREFIX + key
if (param['name'] not in global_block.vars):
continue
var = global_block.var(param['name'])
global_block.create_parameter(**param,
shape=var.shape,
dtype=var.dtype,
type=var.type,
lod_level=var.lod_level,
error_clip=var.error_clip,
stop_gradient=var.stop_gradient,
is_data=var.is_data)
def test_convert_do_model_average(self):
program = fluid.Program()
with fluid.program_guard(program):
input = fluid.layers.data(name="test", shape=[1], dtype="float32")
fluid.layers.fc(input=input,
size=10,
param_attr=fluid.ParamAttr(name="fc_w",
do_model_average=True))
fc_w = [
param for param in
fluid.default_main_program().global_block().iter_parameters()
][0]
flexible_data = module_desc_pb2.FlexibleData()
from_param_to_flexible_data(fc_w, flexible_data)
param_dict = from_flexible_data_to_param(flexible_data)
assert fc_w.do_model_average.__class__ == param_dict[
'do_model_average'].__class__, "do_model_average type convert error!"
assert fc_w.do_model_average == param_dict[
'do_model_average'], "do_model_average value convert error!"
def test_convert_optimize_attr(self):
program = fluid.Program()
with fluid.program_guard(program):
input = fluid.layers.data(name="test", shape=[1], dtype="float32")
fluid.layers.fc(input=input,
size=10,
param_attr=fluid.ParamAttr(name="fc_w",
learning_rate=5))
fc_w = [
param for param in
fluid.default_main_program().global_block().iter_parameters()
][0]
flexible_data = module_desc_pb2.FlexibleData()
from_param_to_flexible_data(fc_w, flexible_data)
param_dict = from_flexible_data_to_param(flexible_data)
assert fc_w.optimize_attr.__class__ == param_dict[
'optimize_attr'].__class__, "optimize_attr type convert error!"
assert fc_w.optimize_attr == param_dict[
'optimize_attr'], "optimize_attr value convert error!"
def test_convert_l2_regularizer(self):
program = fluid.Program()
with fluid.program_guard(program):
input = fluid.layers.data(name="test", shape=[1], dtype="float32")
fluid.layers.fc(input=input,
size=10,
param_attr=fluid.ParamAttr(
name="fc_w",
regularizer=fluid.regularizer.L2Decay(
regularization_coeff=1.5)))
fc_w = [
param for param in
fluid.default_main_program().global_block().iter_parameters()
][0]
flexible_data = module_desc_pb2.FlexibleData()
from_param_to_flexible_data(fc_w, flexible_data)
param_dict = from_flexible_data_to_param(flexible_data)
assert fc_w.regularizer.__class__ == param_dict[
'regularizer'].__class__, "regularzier type convert error!"
assert fc_w.regularizer._regularization_coeff == param_dict[
'regularizer']._regularization_coeff, "regularzier value convert error!"
def test_convert_gradient_clip_by_normal(self):
program = fluid.Program()
with fluid.program_guard(program):
input = fluid.layers.data(name="test", shape=[1], dtype="float32")
fluid.layers.fc(
input=input,
size=10,
param_attr=fluid.ParamAttr(
name="fc_w",
gradient_clip=fluid.clip.GradientClipByNorm(clip_norm=1)))
fc_w = [
param for param in
fluid.default_main_program().global_block().iter_parameters()
][0]
flexible_data = module_desc_pb2.FlexibleData()
from_param_to_flexible_data(fc_w, flexible_data)
param_dict = from_flexible_data_to_param(flexible_data)
assert fc_w.gradient_clip_attr.__class__ == param_dict[
'gradient_clip_attr'].__class__, "clip type convert error!"
assert fc_w.gradient_clip_attr.clip_norm == param_dict[
'gradient_clip_attr'].clip_norm, "clip value convert error!"