def _legacy_save(obj, path, protocol=2):
# 1. input check
if not isinstance(obj, dict):
raise NotImplementedError(
"Now only supports save state_dict of Layer or Optimizer, "
"expect dict, but received %s." % type(obj))
if len(obj) == 0:
warnings.warn("The input state dict is empty, no need to save.")
if not isinstance(protocol, int):
raise ValueError(
"The 'protocol' MUST be `int`, but received {}".format(
type(protocol)))
if protocol < 2 or protocol > 4:
raise ValueError(
"Expected 1<'protocol'<5, but received protocol={}".format(
protocol))
if _is_file_path(path):
filename = os.path.basename(path)
if filename == "":
raise ValueError(
"The input path MUST be format of dirname/filename "
"[dirname\\filename in Windows system], but received "
"filename is empty string.")
# 2. save object
dirname = os.path.dirname(path)
if dirname and not os.path.exists(dirname):
os.makedirs(dirname)
if isinstance(obj, dict):
saved_obj = _build_saved_state_dict(obj)
saved_obj = _unpack_saved_dict(saved_obj, protocol)
# When value of dict is lager than 4GB ,there is a Bug on 'MAC python3'
if _is_file_path(
path) and sys.platform == 'darwin' and sys.version_info.major == 3:
pickle_bytes = pickle.dumps(saved_obj, protocol=protocol)
with open(path, 'wb') as f:
max_bytes = 2**30
for i in range(0, len(pickle_bytes), max_bytes):
f.write(pickle_bytes[i:i + max_bytes])
else:
with _open_file_buffer(path, 'wb') as f:
pickle.dump(saved_obj, f, protocol=protocol)
def _save_selected_rows(selected_rows, file_name):
if not selected_rows.get_tensor()._is_initialized():
raise ValueError("The saved tensor is not initialized.")
if _is_file_path(file_name):
# '_seek' is the end position of this SelectedRows in the file.
_seek = core.save_selected_rows(selected_rows, file_name)
elif _is_memory_buffer(file_name):
selected_rows_bytes = core.save_selected_rows_to_memory(selected_rows)
with _open_file_buffer(file_name, 'wb') as f:
f.write(selected_rows_bytes)
_seek = f.tell()
else:
raise NotImplementedError(
'Only supports saving objects to file or BytesIO, but received {}'.
format(type(file_name)))
return _seek
def _load_lod_tensor(file_name):
temp_t = paddle.fluid.core.LoDTensor()
if _is_file_path(file_name):
# '_seek' is the end position of this tensor in the file.
_seek = paddle.fluid.core.load_lod_tensor(temp_t, file_name)
elif _is_memory_buffer(file_name):
with _open_file_buffer(file_name, 'rb') as f:
tensor_bytes = f.read()
paddle.fluid.core.load_lod_tensor_from_memory(temp_t, tensor_bytes)
_seek = f.tell()
else:
raise NotImplementedError(
'Only supports load objects from file or BytesIO, but received {}'.
format(type(file_name)))
return temp_t, _seek
def _load_selected_rows(file_name):
temp_sr = core.SelectedRows()
if _is_file_path(file_name):
# '_seek' is the end position of this SelectedRows in the file.
_seek = core.load_selected_rows(temp_sr, file_name)
elif _is_memory_buffer(file_name):
with _open_file_buffer(file_name, 'rb') as f:
selected_rows_bytes = f.read()
paddle.fluid.core.load_selected_rows_from_memory(
temp_sr, selected_rows_bytes)
_seek = f.tell()
else:
raise NotImplementedError(
'Only supports load objects from file or BytesIO, but received {}'.
format(type(file_name)))
return temp_sr, _seek
def _save_lod_tensor(tensor, file_name):
if not tensor._is_initialized():
raise ValueError("The saved tensor is not initialized.")
if _is_file_path(file_name):
_seek = core.save_lod_tensor(tensor, file_name)
# '_seek' is the end position of this tensor in the file.
elif _is_memory_buffer(file_name):
tensor_bytes = core.save_lod_tensor_to_memory(tensor)
with _open_file_buffer(file_name, 'wb') as f:
f.write(tensor_bytes)
_seek = f.tell()
else:
raise NotImplementedError(
'Only supports saving objects to file or BytesIO, but received {}'.
format(type(file_name)))
return _seek
def load(path, **configs):
'''
Load an object can be used in paddle from specified path.
.. note::
Now supports loading ``state_dict`` of Layer/Optimizer, Tensor and nested structure containing Tensor, Program.
.. note::
In order to use the model parameters saved by paddle more efficiently,
``paddle.load`` supports loading ``state_dict`` of Layer from the result of
other save APIs except ``paddle.save`` , but the argument ``path`` format is
different:
1. loading from ``paddle.static.save`` or ``paddle.Model().save(training=True)`` ,
``path`` needs to be a complete file name, such as ``model.pdparams`` or
``model.pdopt`` ;
2. loading from ``paddle.jit.save`` or ``paddle.static.save_inference_model``
or ``paddle.Model().save(training=False)`` , ``path`` need to be a file prefix,
such as ``model/mnist``, and ``paddle.load`` will get information from
``mnist.pdmodel`` and ``mnist.pdiparams`` ;
3. loading from paddle 1.x APIs ``paddle.fluid.io.save_inference_model`` or
``paddle.fluid.io.save_params/save_persistables`` , ``path`` need to be a
directory, such as ``model`` and model is a directory.
.. note::
If you load ``state_dict`` from the saved result of static mode API such as
``paddle.static.save`` or ``paddle.static.save_inference_model`` ,
the structured variable name in dynamic mode will cannot be restored.
You need to set the argument ``use_structured_name=False`` when using
``Layer.set_state_dict`` later.
Args:
path(str|BytesIO) : The path/buffer to load the target object. Generally, the path is the target
file path. When loading state_dict from the saved result of the API used to save
the inference model, the path may be a file prefix or directory.
**configs (dict, optional): other load configuration options for compatibility. We do not
recommend using these configurations, they may be removed in the future. If not necessary,
DO NOT use them. Default None.
The following options are currently supported:
(1) model_filename (str): The inference model file name of the paddle 1.x
``save_inference_model`` save format. Default file name is :code:`__model__` .
(2) params_filename (str): The persistable variables file name of the paddle 1.x
``save_inference_model`` save format. No default file name, save variables separately
by default.
(3) return_numpy(bool): If specified as True, return tensor as numpy.ndarray, otherwise return tensor as paddle.Tensor.
Default False.
Returns:
Object(Object): a target object can be used in paddle
Examples:
.. code-block:: python
# example 1: dynamic graph
import paddle
emb = paddle.nn.Embedding(10, 10)
layer_state_dict = emb.state_dict()
# save state_dict of emb
paddle.save(layer_state_dict, "emb.pdparams")
scheduler = paddle.optimizer.lr.NoamDecay(
d_model=0.01, warmup_steps=100, verbose=True)
adam = paddle.optimizer.Adam(
learning_rate=scheduler,
parameters=emb.parameters())
opt_state_dict = adam.state_dict()
# save state_dict of optimizer
paddle.save(opt_state_dict, "adam.pdopt")
# save weight of emb
paddle.save(emb.weight, "emb.weight.pdtensor")
# load state_dict of emb
load_layer_state_dict = paddle.load("emb.pdparams")
# load state_dict of optimizer
load_opt_state_dict = paddle.load("adam.pdopt")
# load weight of emb
load_weight = paddle.load("emb.weight.pdtensor")
# example 2: Load multiple state_dict at the same time
from paddle import nn
from paddle.optimizer import Adam
layer = paddle.nn.Linear(3, 4)
adam = Adam(learning_rate=0.001, parameters=layer.parameters())
obj = {'model': layer.state_dict(), 'opt': adam.state_dict(), 'epoch': 100}
path = 'example/model.pdparams'
paddle.save(obj, path)
obj_load = paddle.load(path)
# example 3: static graph
import paddle
import paddle.static as static
paddle.enable_static()
# create network
x = paddle.static.data(name="x", shape=[None, 224], dtype='float32')
z = paddle.static.nn.fc(x, 10)
place = paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
prog = paddle.static.default_main_program()
for var in prog.list_vars():
if list(var.shape) == [224, 10]:
tensor = var.get_value()
break
# save/load tensor
path_tensor = 'temp/tensor.pdtensor'
paddle.save(tensor, path_tensor)
load_tensor = paddle.load(path_tensor)
# save/load state_dict
path_state_dict = 'temp/model.pdparams'
paddle.save(prog.state_dict("param"), path_tensor)
load_state_dict = paddle.load(path_tensor)
# example 4: load program
import paddle
paddle.enable_static()
data = paddle.static.data(
name='x_static_save', shape=(None, 224), dtype='float32')
y_static = z = paddle.static.nn.fc(data, 10)
main_program = paddle.static.default_main_program()
path = "example/main_program.pdmodel"
paddle.save(main_program, path)
load_main = paddle.load(path)
print(load_main)
# example 5: save object to memory
from io import BytesIO
import paddle
from paddle.nn import Linear
paddle.disable_static()
linear = Linear(5, 10)
state_dict = linear.state_dict()
byio = BytesIO()
paddle.save(state_dict, byio)
tensor = paddle.randn([2, 3], dtype='float32')
paddle.save(tensor, byio)
byio.seek(0)
# load state_dict
dict_load = paddle.load(byio)
'''
if _is_memory_buffer(path) or os.path.isfile(path):
config = _parse_load_config(configs)
exception_type = pickle.UnpicklingError
try:
with _open_file_buffer(path, 'rb') as f:
# When value of dict is lager than 4GB ,there is a Bug on 'MAC python3'
if _is_file_path(
path
) and sys.platform == 'darwin' and sys.version_info.major == 3:
load_result = _pickle_loads_mac(path, f)
else:
load_result = pickle.load(f, encoding='latin1')
# TODO(weixin):If `obj` is any object, the judgment condition should be more precise.
if isinstance(load_result, dict):
load_result = _pack_loaded_dict(load_result)
# paddle2.0: paddle.save/load
if "StructuredToParameterName@@" in load_result:
for key in load_result["StructuredToParameterName@@"]:
if isinstance(load_result[key], np.ndarray):
load_result[key] = _ndarray_to_tensor(
load_result[key], config.return_numpy)
if not config.keep_name_table and "StructuredToParameterName@@" in load_result:
del load_result["StructuredToParameterName@@"]
else:
# paddle2.1 static.save/load
load_result = _parse_load_result(load_result,
config.return_numpy)
else:
load_result = _parse_load_result(load_result,
config.return_numpy)
except exception_type as msg_pickle:
try:
tensor, _ = _load_selected_rows(path)
return tensor
except:
try:
tensor, _ = _load_lod_tensor(path)
if config.return_numpy:
return np.array(tensor)
else:
if _non_static_mode():
return _lod_tensor2varbase(tensor)
return tensor
except:
try:
with _open_file_buffer(path, "rb") as f:
program_desc_str = f.read()
program = Program.parse_from_string(
program_desc_str)
return program
except:
raise ValueError(
"`paddle.load` can not parse the file:{}.".format(
path))
else:
load_result = _legacy_load(path, **configs)
return load_result
def save(obj, path, protocol=4, **configs):
'''
Save an object to the specified path.
.. note::
Now supports saving ``state_dict`` of Layer/Optimizer, Tensor and nested structure containing Tensor, Program.
.. note::
Different from ``paddle.jit.save``, since the save result of ``paddle.save`` is a single file,
there is no need to distinguish multiple saved files by adding a suffix. The argument ``path``
of ``paddle.save`` will be directly used as the saved file name instead of a prefix.
In order to unify the saved file name format, we recommend using the paddle standard suffix:
1. for ``Layer.state_dict`` , recommend to use ``.pdparams`` ;
2. for ``Optimizer.state_dict`` , recommend to use ``.pdopt`` .
For specific examples, please refer to API code examples.
Args:
obj(Object) : The object to be saved.
path(str|BytesIO) : The path/buffer of the object to be saved.
If saved in the current directory, the input path string will be used as the file name.
protocol(int, optional): The protocol version of pickle module must be greater than 1 and less than 5.
Default: 4
**configs(dict, optional): optional keyword arguments. The following options are currently supported:
use_binary_format(bool): When the saved object is static graph variable, you can specify ``use_binary_for_var``.
If True, save the file in the c++ binary format when saving a single static graph variable; otherwise, save it in pickle format.
Default: False
Returns:
None
Examples:
.. code-block:: python
# example 1: dynamic graph
import paddle
emb = paddle.nn.Embedding(10, 10)
layer_state_dict = emb.state_dict()
# save state_dict of emb
paddle.save(layer_state_dict, "emb.pdparams")
scheduler = paddle.optimizer.lr.NoamDecay(
d_model=0.01, warmup_steps=100, verbose=True)
adam = paddle.optimizer.Adam(
learning_rate=scheduler,
parameters=emb.parameters())
opt_state_dict = adam.state_dict()
# save state_dict of optimizer
paddle.save(opt_state_dict, "adam.pdopt")
# save weight of emb
paddle.save(emb.weight, "emb.weight.pdtensor")
# example 2: Save multiple state_dict at the same time
from paddle import nn
from paddle.optimizer import Adam
layer = paddle.nn.Linear(3, 4)
adam = Adam(learning_rate=0.001, parameters=layer.parameters())
obj = {'model': layer.state_dict(), 'opt': adam.state_dict(), 'epoch': 100}
path = 'example/model.pdparams'
paddle.save(obj, path)
# example 3: static graph
import paddle
import paddle.static as static
paddle.enable_static()
# create network
x = paddle.static.data(name="x", shape=[None, 224], dtype='float32')
z = paddle.static.nn.fc(x, 10)
place = paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
prog = paddle.static.default_main_program()
for var in prog.list_vars():
if list(var.shape) == [224, 10]:
tensor = var.get_value()
break
# save/load tensor
path_tensor = 'temp/tensor.pdtensor'
paddle.save(tensor, path_tensor)
# save/load state_dict
path_state_dict = 'temp/model.pdparams'
paddle.save(prog.state_dict("param"), path_tensor)
# example 4: save program
import paddle
paddle.enable_static()
data = paddle.static.data(
name='x_static_save', shape=(None, 224), dtype='float32')
y_static = z = paddle.static.nn.fc(data, 10)
main_program = paddle.static.default_main_program()
path = "example/main_program.pdmodel"
paddle.save(main_program, path)
# example 5: save object to memory
from io import BytesIO
import paddle
from paddle.nn import Linear
paddle.disable_static()
linear = Linear(5, 10)
state_dict = linear.state_dict()
byio = BytesIO()
paddle.save(state_dict, byio)
tensor = paddle.randn([2, 3], dtype='float32')
paddle.save(tensor, byio)
'''
if _is_file_path(path):
# 1. input check
filename = os.path.basename(path)
if filename == "":
raise ValueError(
"The input path MUST be format of dirname/filename "
"[dirname\\filename in Windows system], but received "
"filename is empty string.")
# 2. save object
dirname = os.path.dirname(path)
if dirname and not os.path.exists(dirname):
os.makedirs(dirname)
elif not _is_memory_buffer(path):
raise ValueError(
"only supports saving objects to file and `BytesIO`, but got {}".
format(type(path)))
config = _parse_save_config(configs)
if not isinstance(config.use_binary_format, bool):
raise TypeError(
"Type of `use_binary_format` should be bool, but received {}.".
format(type(config.use_binary_format)))
if config.use_binary_format:
_save_binary_var(obj, path)
else:
# `protocol` need to be used, `pickle_protocol` is a deprecated arg.
if config.pickle_protocol is not None:
protocol = config.pickle_protocol
warnings.warn(
"'pickle_protocol' is a deprecated argument. Please use 'protocol' instead."
)
if isinstance(obj, Program):
obj.desc.flush()
with _open_file_buffer(path, "wb") as f:
f.write(obj.desc.serialize_to_string())
elif _is_state_dict(obj):
if _non_static_mode():
_legacy_save(obj, path, protocol)
else:
_legacy_static_save(obj, path, protocol)
else:
with _open_file_buffer(path, 'wb') as f:
_pickle_save(obj, f, protocol)
