def test_while_no_params_filename(self):
# Phase 1. run and save static model
self.train_and_save_model()
# # Phase 2. load model & train dygraph
with unique_name.guard():
dy_out, dy_param_init_value, dy_param_value = \
self.load_and_train_dygraph()
with unique_name.guard():
static_out, static_param_init_value, static_param_value = \
self.load_and_train_static()
# Phase 3. compare
with unique_name.guard():
dict_old_new_init = rename_var_with_generator(
static_param_init_value.keys())
for key, value in six.iteritems(static_param_init_value):
key = dict_old_new_init[key]
self.assertTrue(np.array_equal(value, dy_param_init_value[key]))
self.assertTrue(np.allclose(static_out, dy_out))
for key, value in six.iteritems(static_param_value):
key += LOADED_VAR_SUFFIX
self.assertTrue(np.allclose(value, dy_param_value[key], atol=1e-5))
def test_save_load_finetune_load(self):
model_path = "test_jit_save_load_save_without_running/model"
IMAGE_SIZE = 224
inps0 = paddle.randn([1, IMAGE_SIZE])
inps1 = paddle.randn([2, IMAGE_SIZE])
# Use new namespace
with unique_name.guard():
layer_save = LayerSaved(IMAGE_SIZE, IMAGE_SIZE)
#save
paddle.jit.save(layer_save,
model_path,
input_spec=[
paddle.static.InputSpec(shape=[None, IMAGE_SIZE],
dtype='float32')
])
result_00 = layer_save(inps0)
result_01 = layer_save(inps1)
#load and save without running
with unique_name.guard():
layer_load = paddle.jit.load(model_path)
paddle.jit.save(layer_load,
model_path,
input_spec=[
paddle.static.InputSpec(
shape=[None, IMAGE_SIZE], dtype='float32')
])
#reload
layer_reload = paddle.jit.load(model_path)
result_10 = layer_reload(inps0)
result_11 = layer_reload(inps1)
self.assertTrue(float((result_00 - result_10).abs().max()) < 1e-5)
self.assertTrue(float((result_01 - result_11).abs().max()) < 1e-5)
def test_save_load_finetune_load(self):
model_path = "test_jit_save_load_finetune_load/model"
IMAGE_SIZE = 224
inps0 = paddle.randn([1, IMAGE_SIZE])
inps1 = paddle.randn([2, IMAGE_SIZE])
# Use new namespace
with unique_name.guard():
layer_save = LayerSaved(IMAGE_SIZE, IMAGE_SIZE)
layer_save(inps0)
#save
paddle.jit.save(layer_save, model_path)
#load
with unique_name.guard():
layer_load = LayerLoadFinetune(IMAGE_SIZE, IMAGE_SIZE, model_path)
#train
train(layer_load, input_size=IMAGE_SIZE)
result_00 = layer_load(inps0)
result_01 = layer_load(inps1)
#save
paddle.jit.save(layer_load, model_path)
#load
layer_finetune = paddle.jit.load(model_path)
result_10 = layer_finetune(inps0)
result_11 = layer_finetune(inps1)
self.assertTrue(float((result_00 - result_10).abs().max()) < 1e-5)
self.assertTrue(float(((result_01 - result_11)).abs().max()) < 1e-5)
def __init__(self, programs, persistable_vars):
super(TranslatedLayer, self).__init__()
if not isinstance(programs, dict):
raise TypeError(
"TranslatedLayer need to use _ProgramHolder's dict for initialization."
)
if not isinstance(persistable_vars, dict):
raise TypeError(
"TranslatedLayer need to use persistable variable dict for initialization."
)
self._program_holder_dict = programs
# NOTE(chenweihang): [ why not use var name directly? ]
# When add parameter or buffer to Layer by follow apis,
# the variable name can't contain `.`, beccause which may cause
# AttributeError when access the newly added parameter or buffer
# in the form of `self.**.**``, but the ParamBase or BarBase
# name contains `.` originally, such as `linear_0.w_0`, so here
# need to generate new var name for each var
self._persistable_var_name_dict = dict()
# the TranslatedLayer object holded var names count started from 0
with unique_name.guard():
for name, var in persistable_vars.items():
if isinstance(var, framework.ParamBase):
dy_name = _generate_unique_var_name(PARAMETER_NAME_PREFIX)
self._persistable_var_name_dict[name] = dy_name
self.add_parameter(dy_name, var)
elif isinstance(var, core.VarBase):
dy_name = _generate_unique_var_name(BUFFER_NAME_PREFIX)
self._persistable_var_name_dict[name] = dy_name
self.register_buffer(dy_name, var)
else:
raise TypeError(
"Adding persistent variable which to layer is not supported now"
)
self._is_test = True
