@@ -6,7 +6,7 @@ SPHINXOPTS = | |||
SPHINXAPIDOC = sphinx-apidoc | |||
SPHINXBUILD = sphinx-build | |||
SPHINXPROJ = fastNLP | |||
SPHINXEXCLUDE = ../fastNLP/transformers/* ../fastNLP/modules/* ../fastNLP/core/drivers/torch_paddle_driver/* ../fastNLP/core/utils/torch_paddle_utils.py | |||
SPHINXEXCLUDE = ../fastNLP/transformers/* | |||
SOURCEDIR = source | |||
BUILDDIR = build | |||
PORT = 9000 | |||
@@ -16,7 +16,7 @@ help: | |||
@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) | |||
apidoc: | |||
$(SPHINXAPIDOC) -efM -d 6 -o source ../$(SPHINXPROJ) $(SPHINXEXCLUDE) | |||
$(SPHINXAPIDOC) -efM -o source ../$(SPHINXPROJ) $(SPHINXEXCLUDE) | |||
server: | |||
cd build/html && python -m http.server $(PORT) | |||
@@ -24,6 +24,9 @@ server: | |||
delete: | |||
rm -f source/$(SPHINXPROJ).* source/modules.rst && rm -rf build | |||
web: | |||
make html && make server | |||
dev: | |||
make delete && make apidoc && make html && make server | |||
@@ -42,7 +42,8 @@ extensions = [ | |||
'sphinx.ext.viewcode', | |||
'sphinx.ext.autosummary', | |||
'sphinx.ext.mathjax', | |||
'sphinx.ext.todo' | |||
'sphinx.ext.todo', | |||
'sphinx_autodoc_typehints' | |||
] | |||
autodoc_default_options = { | |||
@@ -53,8 +54,10 @@ autodoc_default_options = { | |||
add_module_names = False | |||
autosummary_ignore_module_all = False | |||
autodoc_typehints = "description" | |||
# autodoc_typehints = "description" | |||
autoclass_content = "class" | |||
typehints_fully_qualified = False | |||
typehints_defaults = "comma" | |||
# Add any paths that contain templates here, relative to this directory. | |||
templates_path = ['_templates'] | |||
@@ -168,8 +171,8 @@ texinfo_documents = [ | |||
# -- Extension configuration ------------------------------------------------- | |||
def maybe_skip_member(app, what, name, obj, skip, options): | |||
# if obj.__doc__ is None: | |||
# return True | |||
if obj.__doc__ is None: | |||
return True | |||
if name == "__init__": | |||
return False | |||
if name.startswith("_"): | |||
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.callbacks.torch_callbacks | |||
@@ -18,7 +18,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.callbacks.callback | |||
fastNLP.core.callbacks.callback_event | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.callbacks.torch_callbacks.torch_grad_clip_callback | |||
fastNLP.core.callbacks.torch_callbacks.torch_lr_sched_callback |
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.collators.padders.exceptions | |||
fastNLP.core.collators.padders.get_padder | |||
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.collators.padders | |||
@@ -18,7 +18,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.collators.collator | |||
fastNLP.core.collators.packer_unpacker |
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.controllers.loops.evaluate_batch_loop | |||
fastNLP.core.controllers.loops.loop | |||
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.controllers.loops | |||
fastNLP.core.controllers.utils | |||
@@ -19,7 +19,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.controllers.evaluator | |||
fastNLP.core.controllers.trainer |
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.controllers.utils.state | |||
fastNLP.core.controllers.utils.utils |
@@ -10,6 +10,6 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.dataloaders.jittor_dataloader.fdl |
@@ -10,6 +10,6 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.dataloaders.paddle_dataloader.fdl |
@@ -0,0 +1,7 @@ | |||
fastNLP.core.dataloaders.prepare\_dataloader module | |||
=================================================== | |||
.. automodule:: fastNLP.core.dataloaders.prepare_dataloader | |||
:members: | |||
:undoc-members: | |||
:show-inheritance: |
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.dataloaders.jittor_dataloader | |||
fastNLP.core.dataloaders.paddle_dataloader | |||
@@ -20,7 +20,8 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.dataloaders.mix_dataloader | |||
fastNLP.core.dataloaders.prepare_dataloader | |||
fastNLP.core.dataloaders.utils |
@@ -10,6 +10,6 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.dataloaders.torch_dataloader.fdl |
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.dataset.dataset | |||
fastNLP.core.dataset.field | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.drivers.jittor_driver.initialize_jittor_driver | |||
fastNLP.core.drivers.jittor_driver.jittor_driver | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.drivers.paddle_driver.dist_utils | |||
fastNLP.core.drivers.paddle_driver.fleet | |||
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.drivers.jittor_driver | |||
fastNLP.core.drivers.paddle_driver | |||
@@ -20,7 +20,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.drivers.choose_driver | |||
fastNLP.core.drivers.driver | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.drivers.torch_driver.ddp | |||
fastNLP.core.drivers.torch_driver.dist_utils | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.log.handler | |||
fastNLP.core.log.highlighter | |||
@@ -10,6 +10,6 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.metrics.backend.jittor_backend.backend |
@@ -10,6 +10,6 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.metrics.backend.paddle_backend.backend |
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.metrics.backend.jittor_backend | |||
fastNLP.core.metrics.backend.paddle_backend | |||
@@ -20,7 +20,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.metrics.backend.auto_backend | |||
fastNLP.core.metrics.backend.backend |
@@ -10,6 +10,6 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.metrics.backend.torch_backend.backend |
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.metrics.backend | |||
@@ -18,7 +18,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.metrics.accuracy | |||
fastNLP.core.metrics.classify_f1_pre_rec_metric | |||
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.callbacks | |||
fastNLP.core.collators | |||
@@ -27,6 +27,6 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.vocabulary |
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.samplers.conversion_utils | |||
fastNLP.core.samplers.mix_sampler | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core.utils.cache_results | |||
fastNLP.core.utils.dummy_class | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.envs.distributed | |||
fastNLP.envs.env | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.io.loader.classification | |||
fastNLP.io.loader.conll | |||
@@ -10,7 +10,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.io.pipe.classification | |||
fastNLP.io.pipe.conll | |||
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.io.loader | |||
fastNLP.io.pipe | |||
@@ -19,7 +19,7 @@ Submodules | |||
---------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.io.data_bundle | |||
fastNLP.io.embed_loader | |||
@@ -10,7 +10,7 @@ Subpackages | |||
----------- | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP.core | |||
fastNLP.envs | |||
@@ -2,6 +2,6 @@ fastNLP | |||
======= | |||
.. toctree:: | |||
:maxdepth: 6 | |||
:maxdepth: 4 | |||
fastNLP |
@@ -63,7 +63,6 @@ __all__ = [ | |||
"PaddleFleetDriver", | |||
"JittorSingleDriver", | |||
"JittorMPIDriver", | |||
"TorchPaddleDriver", | |||
# log | |||
"logger", | |||
@@ -10,8 +10,8 @@ from .callback_event import Event | |||
from .callback import Callback | |||
from fastNLP.core.log import logger | |||
from .progress_callback import ProgressCallback, choose_progress_callback | |||
from fastNLP.envs import rank_zero_call | |||
from fastNLP.core.utils.utils import _get_fun_msg | |||
from ..utils.exceptions import EarlyStopException | |||
from ..utils.utils import _get_fun_msg | |||
def _transfer(func): | |||
@@ -25,6 +25,8 @@ def _transfer(func): | |||
for callback_fn in manager.callback_fns[func.__name__]: | |||
try: | |||
callback_fn(*arg, **kwargs) | |||
except EarlyStopException as e: | |||
raise e | |||
except BaseException as e: | |||
logger.error(f"The following callback_fn raise exception:{_get_fun_msg(callback_fn)}.") | |||
raise e | |||
@@ -186,6 +188,8 @@ class CallbackManager: | |||
for each_callback_filters in self._callback_filters: | |||
if each_callback_filters[0] not in _record_duplicated_callback_names: | |||
_record_duplicated_callback_names.add(each_callback_filters[0]) | |||
if 'filter_states' not in states[each_callback_filters[0]]: | |||
states[each_callback_filters[0]]["filter_states"] = {} | |||
states[each_callback_filters[0]]["filter_states"][each_callback_filters[1]] = each_callback_filters[2].state_dict() | |||
# 3. 保存 callback_counter; | |||
@@ -212,7 +216,9 @@ class CallbackManager: | |||
if each_callback_filters[0] in states: | |||
if each_callback_filters[0] not in _already_loaded_callback_names: | |||
_already_loaded_callback_names.add(each_callback_filters[0]) | |||
each_callback_filters[2].load_state_dict(states[each_callback_filters[0]]["filter_states"][each_callback_filters[1]]) | |||
if 'filter_states' in states[each_callback_filters[0]] and \ | |||
each_callback_filters[1] in states[each_callback_filters[0]]['filter_states']: | |||
each_callback_filters[2].load_state_dict(states[each_callback_filters[0]]['filter_states'][each_callback_filters[1]]) | |||
else: | |||
_duplicated_callback_names.add(each_callback_filters[0]) | |||
@@ -19,7 +19,7 @@ class CheckpointCallback(Callback): | |||
only_state_dict: bool = True, model_save_fn: Optional[Callable] = None, save_object: str = 'model', | |||
save_evaluate_results=True, **kwargs): | |||
""" | |||
保存模型 checkpoint 的 callback ,其保存的文件目录以及文件名命名规则如下:: | |||
保存 checkpoint 的 callback ,其保存的文件目录以及文件名命名规则如下:: | |||
- folder/ | |||
- YYYY-mm-dd-HH_MM_SS_fffff/ # 自动根据当前脚本的启动时间创建的 | |||
@@ -29,8 +29,9 @@ class CheckpointCallback(Callback): | |||
- {save_object}-epoch_{epoch_idx}-batch_{global_batch_idx}-exception_{exception_type}/ # exception时保存。 | |||
- {save_object}-epoch_{epoch_idx}-batch_{global_batch_idx}-{monitor}_{monitor_value}/ # 满足topk条件存储文件名 | |||
model_save_fn 为 None ,则以上每个 folder 中,将生成 fastnlp_model.pkl.tar 文件。 | |||
若 model_save_fn 不为 None,则 fastNLP 将 folder 绝对路径传递给该函数,fastNLP 在该 folder 下不进行模型保存。 | |||
model_save_fn 为 None ,则以上每个 folder 中,将生成 fastnlp_model.pkl.tar 文件。若 model_save_fn 不为 None, | |||
则 fastNLP 将 folder 绝对路径传递给该函数,fastNLP 在该 folder 下不进行模型保存。默认情况下,本 checkpoint 只保存了 model | |||
的状态;如还需保存 Trainer 的状态以断点重训的话,请使用 ``save_object='trainer'`` 。 | |||
:param monitor: 监控的 metric 值。如果在 evaluation 结果中没有找到完全一致的名称,将使用 最长公共字符串算法 找到最匹配 | |||
的那个作为 monitor 。如果为 None,将尝试使用 Trainer 设置的 monitor 。也可以传入一个函数,接受参数为 evaluation 的结 | |||
@@ -46,22 +47,14 @@ class CheckpointCallback(Callback): | |||
:param only_state_dict: 保存模型时是否只保存 state_dict 。当 model_save_fn 不为 None 时,该参数无效。 | |||
:param model_save_fn: 个性化的保存函数,当触发保存操作时,就调用这个函数,这个函数应当接受一个文件夹作为参数,不返回任何东西。 | |||
如果传入了 model_save_fn 函数,fastNLP 将不再进行模型相关的保存。在多卡场景下,我们只在 rank 0 上会运行该函数。 | |||
:param save_object: 可选 ['trainer', 'model'],表示在保存时的保存对象为 trainer+model 还是 只是model 。 | |||
:param save_object: 可选 ['trainer', 'model'],表示在保存时的保存对象为 ``trainer+model`` 还是 只是 ``model`` 。如果 | |||
保存 ``trainer`` 对象的话,将会保存 :class:~fastNLP.Trainer 的相关状态,可以通过 :meth:`Trainer.load` 加载该断 | |||
点继续训练。如果保存的是 ``Model`` 对象,则可以通过 :meth:`Trainer.load_model` 加载该模型权重。 | |||
:param save_evaluate_results: 是否保存 evaluate 的结果。如果为 True ,在保存 topk 模型的 folder 中还将额外保存一个 | |||
fastnlp_evaluate_results.json 文件,记录当前的 results。仅在设置了 topk 的场景下有用,默认为 True 。 | |||
:param kwargs: | |||
""" | |||
super().__init__() | |||
if folder is None: | |||
logger.warning( | |||
"Parameter `folder` is None, and we will use the current work directory to find and load your model.") | |||
folder = Path.cwd() | |||
folder = Path(folder) | |||
if not folder.exists(): | |||
raise NotADirectoryError(f"Path '{folder.absolute()}' is not existed!") | |||
elif folder.is_file(): | |||
raise ValueError("Parameter `folder` should be a directory instead of a file.") | |||
if every_n_epochs is not None: | |||
if not isinstance(every_n_epochs, int) or every_n_epochs < 1: | |||
raise ValueError("Parameter `every_n_epochs` should be an int and greater than or equal to 1.") | |||
@@ -74,12 +67,6 @@ class CheckpointCallback(Callback): | |||
else: | |||
every_n_batches = sys.maxsize # 使得没有数字可以整除 | |||
if topk is not None: | |||
if not isinstance(topk, int): | |||
raise ValueError("Parameter `topk` should be an int.") | |||
else: | |||
topk = 0 | |||
if on_exceptions is not None: | |||
if not isinstance(on_exceptions, Sequence): | |||
on_exceptions = [on_exceptions] | |||
@@ -19,7 +19,8 @@ class LoadBestModelCallback(HasMonitorCallback): | |||
model_load_fn:Optional[Callable] = None, | |||
delete_after_train:bool = True): | |||
""" | |||
保存最佳的 monitor 值最佳的模型,并在训练结束的时候重新加载模型。仅在训练正常结束的时候才能加载最好的模型。 | |||
保存最佳的 monitor 值最佳的模型,并在训练结束的时候重新加载模型,默认会在加载之后删除权重文件。仅在训练正常结束的时候才能加载 | |||
最好的模型。 | |||
:param str monitor: 监控的 metric 值。如果在 evaluation 结果中没有找到完全一致的名称,将使用 最长公共字符串算法 找到最匹配 | |||
的那个作为 monitor 。如果为 None,将尝试使用 Trainer 设置的 monitor 。也可以传入一个函数,接受参数为 evaluation 的结 | |||
@@ -33,9 +33,8 @@ class Saver: | |||
:param kwargs: 更多需要传递给 Trainer.save() 或者 Trainer.save_model() 接口的参数。 | |||
""" | |||
if folder is None: | |||
logger.rank_zero_warning( | |||
"Parameter `folder` is None, and we will use the current work directory to find and load your model.") | |||
folder = Path.cwd() | |||
folder = Path.cwd().absolute() | |||
logger.info(f"Parameter `folder` is None, and we will use {folder} to save and load your model.") | |||
folder = Path(folder) | |||
if not folder.exists(): | |||
folder.mkdir(parents=True, exist_ok=True) | |||
@@ -8,8 +8,8 @@ from fastNLP.core.utils.utils import _get_fun_msg | |||
def _get_monitor_value(monitor: Union[callable, str], real_monitor: Optional[str], res: dict) ->Tuple[str, float]: | |||
""" | |||
从res中寻找 monitor 并返回。如果 monitor 没找到则尝试用 _real_monitor ,若 _real_monitor 为 None 则尝试使用 monitor 的值进行 | |||
匹配。 | |||
从 ``res`` 中寻找 ``monitor`` 并返回。如果 ``monitor`` 没找到则尝试用 ``_real_monitor`` ,若 ``_real_monitor`` 为 ``None`` | |||
则尝试使用 ``monitor`` 的值进行匹配。 | |||
:param monitor: | |||
:param real_monitor: | |||
@@ -121,7 +121,7 @@ def get_padder(batch_field:Sequence[Any], pad_val, dtype, backend, field_name)-> | |||
# 这里 ele_dtype 传入为 None 的原因是防止出现 paddle tensor 转换为 torch tensor | |||
return TorchTensorPadder(pad_val=pad_val, ele_dtype=None, dtype=dtype) | |||
elif backend == 'paddle': | |||
return PaddleTensorPadder(pad_val=pad_val, ele_dtype=None, dtype=dtype) | |||
return PaddleTensorPadder(pad_val=pad_val, ele_dtype=ele_dtype, dtype=dtype) | |||
elif backend == 'jittor': | |||
return JittorTensorPadder(pad_val=pad_val, ele_dtype=ele_dtype, dtype=dtype) | |||
else: | |||
@@ -141,7 +141,10 @@ class PaddleTensorPadder(Padder): | |||
shapes = [field.shape for field in batch_field] | |||
max_shape = [len(batch_field)] + [max(*_) for _ in zip(*shapes)] | |||
array = np.full(max_shape, fill_value=pad_val) | |||
if isinstance(batch_field[0], paddle.Tensor): | |||
array = paddle.full(max_shape, fill_value=pad_val, dtype=dtype) | |||
else: | |||
array = np.full(max_shape, fill_value=pad_val, dtype=batch_field[0].dtype) | |||
for i, field in enumerate(batch_field): | |||
slices = (i, ) + tuple(slice(0, s) for s in shapes[i]) | |||
array[slices] = field | |||
@@ -118,8 +118,8 @@ class TorchTensorPadder(Padder): | |||
batch_field = [torch.tensor(field.tolist(), dtype=dtype) for field in batch_field] | |||
else: | |||
device = batch_field[0].device | |||
if dtype is None: | |||
dtype = batch_field[0].dtype | |||
if dtype is None: | |||
dtype = batch_field[0].dtype | |||
except AttributeError: | |||
raise RuntimeError(f"If the field is not a torch.Tensor (it is {type(batch_field[0])}), " | |||
f"it must have tolist() method.") | |||
@@ -56,6 +56,8 @@ class Evaluator: | |||
* ddp_kwargs -- 用于在使用 ``TorchDDPDriver`` 时指定 ``DistributedDataParallel`` 初始化时的参数;例如传入 | |||
{'find_unused_parameters': True} 来解决有参数不参与前向运算导致的报错等; | |||
* torch_non_blocking -- 表示用于 pytorch 的 tensor 的 to 方法的参数 non_blocking; | |||
* *data_device* -- 表示如果用户的模型 device (在 Driver 中对应为参数 model_device)为 None 时,我们会将数据迁移到 data_device 上; | |||
注意如果 model_device 为 None,那么 data_device 不会起作用; | |||
* *model_use_eval_mode* (``bool``) -- | |||
是否在 evaluate 的时候将 model 的状态设置成 eval 状态。在 eval 状态下,model 的 | |||
dropout 与 batch normalization 将会关闭。默认为True。如果为 False,fastNLP 不会对 model 的 evaluate 状态做任何设置。无论 | |||
@@ -234,8 +236,7 @@ class Evaluator: | |||
""" | |||
调用所有 metric 的 reset() 方法,清除累积的状态。 | |||
Returns: | |||
:return: | |||
""" | |||
self.metrics_wrapper.reset() | |||
@@ -357,6 +358,11 @@ class _MetricsWrapper: | |||
metric.update(res) | |||
def reset(self): | |||
""" | |||
将 Metric 中的状态重新设置。 | |||
:return: | |||
""" | |||
for metric in self._metrics: | |||
if _is_allennlp_metric(metric): | |||
metric.get_metric(reset=True) | |||
@@ -1,4 +1,10 @@ | |||
from typing import Union, Optional, List, Callable, Dict, Sequence, BinaryIO, IO | |||
""" | |||
``Trainer`` 是 fastNLP 用于训练模型的专门的训练器,其支持多种不同的驱动模式 ``Driver``,不仅包括最为经常使用的 DDP,而且还支持 jittor 等国产 | |||
的训练框架;新版的 fastNLP 新加入了方便的 callback 函数修饰器,并且支持定制用户自己特定的训练循环过程;通过使用该训练器,用户只需要自己实现 | |||
模型部分,而将训练层面的逻辑完全地交给 fastNLP; | |||
""" | |||
from typing import Union, Optional, List, Callable, Dict, BinaryIO | |||
from functools import partial | |||
from collections import defaultdict | |||
import copy | |||
@@ -7,7 +13,6 @@ from dataclasses import is_dataclass | |||
import os | |||
from pathlib import Path | |||
import io | |||
import inspect | |||
__all__ = [ | |||
'Trainer', | |||
@@ -62,12 +67,20 @@ class Trainer(TrainerEventTrigger): | |||
**kwargs | |||
): | |||
r""" | |||
`Trainer` 是 fastNLP 用于训练模型的专门的训练器,其支持多种不同的驱动模式,不仅包括最为经常使用的 DDP,而且还支持 jittor 等国产 | |||
的训练框架;新版的 fastNLP 新加入了方便的 callback 函数修饰器,并且支持定制用户自己特定的训练循环过程;通过使用该训练器,用户只需 | |||
要自己实现模型部分,而将训练层面的逻辑完全地交给 fastNLP; | |||
:param model: 训练所需要的模型,例如 ``torch.nn.Module``; | |||
.. note:: | |||
当使用 pytorch 时,注意参数 ``model`` 在大多数情况下为 ``nn.Module``。但是您仍能够通过使用一些特定的组合来使用情况,如下所示: | |||
:param model: 训练所需要的模型,目前支持 pytorch; | |||
:param driver: 训练模型所使用的具体的驱动模式,应当为以下选择中的一个:["torch",],之后我们会加入 jittor、paddle 等 | |||
1. 当希望使用 ``DataParallel`` 时,您应当使用 ``TorchSingleDriver``,意味着您在初始化 ``Trainer`` 时参数 ``device`` 不应当为 | |||
一个 ``List``; | |||
2. 当您选择自己初始化 ``init_process_group`` 时(这种情况要求您传入的 ``model`` 参数一定为 ``DistributedDataParallel``), | |||
您应当使用 ``TorchDDPDriver``,意味着您需要通过 ``python -m torch.distributed.launch`` 的方式来启动训练,此时参数 ``device`` | |||
应当设置为 None(此时我们会忽略该参数),具体见下面对于参数 ``device`` 的更详细的解释。 | |||
:param driver: 训练模型所使用的具体的驱动模式,应当为以下选择中的一个:["torch"],之后我们会加入 jittor、paddle 等 | |||
国产框架的训练模式;其中 "torch" 表示使用 ``TorchSingleDriver`` 或者 ``TorchDDPDriver``,具体使用哪一种取决于参数 ``device`` | |||
的设置; | |||
:param train_dataloader: 训练数据集,注意其必须是单独的一个数据集,不能是 List 或者 Dict; | |||
@@ -80,79 +93,248 @@ class Trainer(TrainerEventTrigger): | |||
device 的可选输入如下所示: | |||
* *str*: 例如 'cpu', 'cuda', 'cuda:0', 'cuda:1' 等; | |||
* *torch.device*: 将模型装载到 ``torch.device`` 上; | |||
* *torch.device*: 例如 'torch.device("cuda:0")'; | |||
* *int*: 将使用 ``device_id`` 为该值的 ``gpu`` 进行训练;如果值为 -1,那么默认使用全部的显卡,此时使用的 driver 实例是 `TorchDDPDriver`; | |||
* *list(int)*: 如果多于 1 个device,应当通过该种方式进行设定;注意此时我们一定会使用 ``TorchDDPDriver``,不管您传入的列表的长度是 1 还是其它值; | |||
* *None*: 为None则不对模型进行任何处理; | |||
* *None*: 仅当用户自己通过训练框架提供的并行训练启动脚本开启 ddp 进程时为 None; | |||
.. note:: | |||
如果希望使用 ``TorchDDPDriver``,在初始化 ``Trainer`` 时您应当使用:: | |||
Trainer(driver="torch", device=[0, 1]) | |||
注意如果这时 ``device=[0]``,我们仍旧会使用 ``TorchDDPDriver``。 | |||
如果希望使用 ``TorchSingleDriver``,则在初始化 ``Trainer`` 时您应当使用:: | |||
Trainer(driver="torch", device=0) | |||
.. node:: | |||
.. warning:: | |||
如果希望使用 ``TorchDDPDriver`` | |||
注意参数 ``device`` 仅当您通过 pytorch 或者其它训练框架自身的并行训练启动脚本启动 ddp 训练时才允许为 ``None``! | |||
例如,当您使用:: | |||
python -m torch.distributed.launch --nproc_per_node 2 train.py | |||
来使用 ``TorchDDPDriver`` 时,此时参数 ``device`` 不再有效(不管您是否自己初始化 ``init_process_group``),我们将直接 | |||
通过 ``torch.device(f"cuda:{local_rank}")`` 来获取当前进程所使用的的具体的 gpu 设备。因此此时您需要使用 ``os.environ["CUDA_VISIBLE_DEVICES"]`` | |||
来指定要使用的具体的 gpu 设备。 | |||
另一点需要注意的是,当您没有选择自己初始化 ``init_process_group`` 时,我们仍旧会帮助您把模型和数据迁移到当前进程所使用的 | |||
具体的 gpu 设备上。但是如果您选择自己在 ``Trainer`` 初始化前(意味着在 ``driver`` 的 ``setup`` 前)初始化 ``init_process_group``, | |||
那么对于模型的迁移应当完全由您自己来完成。此时对于数据的迁移,如果您在 ``Trainer`` 初始化时指定了参数 ``data_device``,那么 | |||
我们会将数据迁移到 ``data_device`` 上;如果其为 None,那么将数据迁移到正确的设备上应当由您自己来完成。 | |||
对于使用 ``TorchDDPDriver`` 的更多细节,请见 :class:`fastNLP.core.drivers.torch_driver.TorchDDPDriver`。 | |||
:param n_epochs: 训练总共的 epoch 的数量,默认为 20; | |||
:param evaluate_dataloaders: 验证数据集,其可以是单独的一个数据集,也可以是多个数据集;当为多个数据集时,注意其必须是 Dict;默认 | |||
为 None; | |||
:param batch_step_fn: 定制每次 train batch 执行的函数。该函数应接受两个参数为 `trainer` 和`batch`,不需要要返回值;可以 | |||
参考 fastNLP.core.controllers.loops.train_batch_loop.TrainBatchLoop中的batch_step_fn函数。 | |||
:param evaluate_batch_step_fn: 定制每次 evaluate batch 执行的函数。该函数应接受的两个参数为 `evaluator` 和 `batch`, | |||
不需要有返回值;可以参考 fastNLP.core.controllers.loops.evaluate_batch_loop.EvaluateBatchLoop中的batch_step_fn函数。 | |||
:param train_fn: 用来控制 `Trainer` 在训练的前向传播过程中是调用模型的哪一个函数,例如是 `train_step` 还是 `forward`; | |||
默认为 None,如果该值是 None,那么我们会默认使用 `train_step` 当做前向传播的函数,如果在模型中没有找到该方法, | |||
则使用模型默认的前向传播函数。 | |||
:param evaluate_fn: 用来控制 `Trainer` 中内置的 `Evaluator` 的模式,应当为 None 或者一个字符串;其使用方式和 train_fn 类似; | |||
注意该参数我们会直接传给 Trainer 中内置的 Evaluator(如果不为 None);如果该值为 None ,将首先尝试寻找模型中是否有 | |||
evaluate_step 这个函数,如果没有则使用 forward 函数。 | |||
:param callbacks: 训练当中触发的 callback 类,该参数应当为一个列表,其中的每一个元素都应当继承 `Callback` 类; | |||
:param metrics: 应当为一个字典,其中 key 表示 monitor,例如 {"acc1": AccMetric(), "acc2": AccMetric()}; | |||
:param evaluate_every: 可以为负数、正数或者函数;为负数时表示每隔几个 epoch evaluate 一次;为正数则表示每隔几个 batch evaluate 一次; | |||
为函数时表示用户自己传入的用于控制 Trainer 中的 evaluate 的频率的函数,该函数的应该接受当前 trainer 对象作为参数,并 | |||
返回一个 bool 值,返回为 True 说明需要进行 evaluate ;将在每个 batch 结束后调用该函数判断是否需要 evaluate 。 | |||
:param input_mapping: 应当为一个字典或者一个函数,表示在当前 step 拿到一个 batch 的训练数据后,应当做怎样的映射处理;如果其是 | |||
一个字典,并且 batch 也是一个 `Dict`,那么我们会把 batch 中同样在 input_mapping 中的 key 修改为 input_mapping 的对应 key 的 | |||
value;如果 batch 是一个 `dataclass`,那么我们会先将该 dataclass 转换为一个 Dict,然后再进行上述转换;如果 batch 此时是其它 | |||
类型,那么我们将会直接报错;如果 input_mapping 是一个函数,那么对于取出的 batch,我们将不会做任何处理,而是直接将其传入该函数里; | |||
注意该参数会被传进 `Evaluator` 中;因此你可以通过该参数来实现将训练数据 batch 移到对应机器上的工作(例如当参数 `device` 为 None 时); | |||
如果 train 和 evaluate 需要使用不同的 input_mapping, 请使用 train_input_mapping 与 evaluate_input_mapping 设置。 | |||
:param output_mapping: 应当为一个字典或者函数。作用和 input_mapping 类似,区别在于其用于转换输出;如果 output_mapping 是一个 | |||
函数,那么我们将会直接将模型的输出传给该函数;如果其是一个 `Dict`,那么我们需要 batch 必须是 `Dict` 或者 `dataclass` 类型, | |||
如果 batch 是一个 `Dict`,那么我们会把 batch 中同样在 output_mapping 中的 key 修改为 output_mapping 的对应 key 的 value; | |||
如果 batch 是一个 `dataclass`,那么我们会先将该 dataclass 转换为一个 Dict,然后再进行上述转换; | |||
如果 train 和 evaluate 需要使用不同的 output_mapping, 请使用 train_output_mapping 与 evaluate_output_mapping 设置。 | |||
:param model_wo_auto_param_call: 是否关闭在训练时调用我们的 auto_param_call 来自动匹配 batch 和 forward 函数的参数的行为; | |||
如果该值为 False,并且当 batch 为字典时,我们会根据 forward 所需要的参数从 batch 中提取对应的对象,传入到 forward 函数中;如果该值 | |||
为 True,那么我们会将 batch 直接透传给模型。注意该参数应用于 `train_step`, `evaluate_step` 和 `test_step`; | |||
:param accumulation_steps: 梯度累积的步数,表示每隔几个 batch 优化器迭代一次;默认为 1; | |||
:param fp16: 是否开启混合精度训练;默认为 False; | |||
:param monitor: 当存在 evaluate_dataloaders 时,默认的 monitor metric 的名字。传入的 callback 如果有 monitor 参数且没有 | |||
在 callback 初始化设定的,将采取这个值。如果在 evaluation 结果中没有找到完全一致的名称,将使用 最长公共字符串算法 找到最匹配 | |||
的那个作为 monitor 。也可以传入一个函数,接受参数为 evaluation 的结果(字典类型),返回一个 float 值作为 monitor 的结果。 | |||
如果 evaluate_dataloaders 与 metrics 没有提供,该参数无意义。 | |||
:param larger_better: monitor 的值是否是越大越好。 | |||
:param marker: 用于标记一个 Trainer 实例,从而在用户调用 `Trainer.on` 函数时,标记该 callback 函数属于哪一个具体的 'trainer' 实例;默认为 None; | |||
:param kwargs: 一些其它的可能需要的参数,见下方的说明 | |||
为 None; | |||
:param batch_step_fn: 定制每次训练时前向运行一个 batch 的数据所执行的函数。该函数应接受两个参数为 ``trainer`` 和 ``batch``, | |||
不需要要返回值;更详细的使用位置和说明请见 :meth:`fastNLP.core.controllers.TrainBatchLoop.batch_step_fn`; | |||
:param evaluate_batch_step_fn: 定制每次验证时前向运行一个 batch 的数据所执行的函数。该函数应接受的两个参数为 ``evaluator`` 和 ``batch``, | |||
不需要有返回值;可以参考 :meth:`fastNLP.core.controllers.EvaluateBatchLoop.batch_step_fn`; | |||
:param train_fn: 用来控制 ``Trainer`` 在训练的前向传播过程中是调用模型的哪一个函数,例如是 ``train_step`` 还是 ``forward``; | |||
默认为 ``None``,如果该值是 ``None``,那么我们会默认使用 ``train_step`` 当做前向传播的函数,如果在模型的定义类中没有找到该方法, | |||
则使用模型默认的前向传播函数,例如对于 pytorch 来说就是 ``forward``。 | |||
.. note:: | |||
在 fastNLP 中,对于训练时使用的前向传播函数的查找逻辑如下所示: | |||
1. 如果 ``train_fn`` 为 None,那么在 model 的类 Model 中寻找方法 ``Model.train_step``;如果没有找到,那么默认使用 ``Model.forward``; | |||
2. 如果 ``train_fn`` 为一个字符串,例如 'my_step_fn',那么我们首先会在 model 的类 Model 中寻找方法 ``Model.my_step_fn``, | |||
如果没有找到,那么会直接报错; | |||
:param evaluate_fn: 用来控制 ``Trainer`` 中内置的 ``Evaluator`` 在验证的前向传播过程中是调用模型的哪一个函数,应当为 ``None`` | |||
或者一个字符串;其使用方式和 train_fn 类似;具体可见 :class:`fastNLP.core.controllers.Evaluator`; | |||
:param callbacks: 训练当中触发的 callback 类,该参数应当为一个列表,其中的每一个元素都应当继承 ``Callback`` 类;具体可见 | |||
:class:`fastNLP.core.callbacks.Callback`; | |||
:param metrics: 用于传给 ``Trainer`` 内部的 ``Evaluator`` 实例来进行训练过程中的验证。其应当为一个字典,其中 key 表示 monitor, | |||
例如 {"acc1": AccMetric(), "acc2": AccMetric()}; | |||
目前我们支持的 ``metric`` 的种类有以下几种: | |||
1. fastNLP 自己的 ``metric``:详见 :class:`fastNLP.core.metrics.Metric`; | |||
2. torchmetrics; | |||
3. allennlp.training.metrics; | |||
4. paddle.metric; | |||
:param evaluate_every: 用来控制 ``Trainer`` 内部的 ``Evaluator`` 验证的频率,其可以为负数、正数或者函数: | |||
1. 为负数时表示每隔几个 ``epoch`` evaluate 一次; | |||
2. 为正数则表示每隔几个 ``batch`` evaluate 一次; | |||
3. 为函数时表示用户自己传入的用于控制 evaluate 的频率的函数,该函数的应该接受当前 trainer 对象作为参数,并 | |||
返回一个 bool 值,返回为 True 说明需要进行 evaluate ;将在每个 ``batch`` 结束后调用该函数判断是否需要 evaluate; | |||
.. note:: | |||
如果参数 ``evaluate_every`` 为函数,其应当类似: | |||
>>> def my_evaluate_every(trainer) -> bool: | |||
... if (trainer.global_forward_batches+1) % 1000 == 0: | |||
... return True | |||
... else: | |||
... return False | |||
该函数表示当每经过 1000 个 batch,``Trainer`` 中内置的 ``Evaluator`` 就会验证一次; | |||
另一个需要注意的事情在于该函数会在每一次 batch 的结尾进行调用,当该函数返回 ``True`` 时,``Evaluator`` 才会进行验证; | |||
:param input_mapping: 应当为一个字典或者一个函数,表示在当前 step 拿到一个 batch 的训练数据后,应当做怎样的映射处理: | |||
1. 如果 ``input_mapping`` 是一个字典: | |||
1. 如果此时 batch 也是一个 ``Dict``,那么我们会把 batch 中同样在 ``input_mapping`` 中的 key 修改为 ``input_mapping`` 的对应 ``key`` 的 ``value``; | |||
2. 如果此时 batch 是一个 ``dataclass``,那么我们会先将其转换为一个 ``Dict``,然后再进行上述转换; | |||
3. 如果此时 batch 此时是其它类型,那么我们将会直接报错; | |||
2. 如果 ``input_mapping`` 是一个函数,那么对于取出的 batch,我们将不会做任何处理,而是直接将其传入该函数里; | |||
注意该参数会被传进 ``Evaluator`` 中;因此你可以通过该参数来实现将训练数据 batch 移到对应机器上的工作(例如当参数 ``device`` 为 ``None`` 时); | |||
如果 ``Trainer`` 和 ``Evaluator`` 需要使用不同的 ``input_mapping``, 请使用 ``train_input_mapping`` 与 ``evaluate_input_mapping`` 分别进行设置。 | |||
:param output_mapping: 应当为一个字典或者函数。作用和 ``input_mapping`` 类似,区别在于其用于转换输出: | |||
1. 如果 ``output_mapping`` 是一个 ``Dict``,那么我们需要模型的输出必须是 ``Dict`` 或者 ``dataclass`` 类型: | |||
1. 如果此时模型的输出是一个 ``Dict``,那么我们会把输出中同样在 ``output_mapping`` 中的 key 修改为 ``output_mapping`` 的对应 key 的 value; | |||
2. 如果此时模型的输出是一个 ``dataclass``,那么我们会先将其转换为一个 Dict,然后再进行上述转换; | |||
2. 如果 ``output_mapping`` 是一个函数,那么我们将会直接将模型的输出传给该函数; | |||
如果 ``Trainer`` 和 ``Evaluator`` 需要使用不同的 ``output_mapping``, 请使用 ``train_output_mapping`` 与 ``evaluate_output_mapping`` 分别进行设置; | |||
.. note:: | |||
``input_mapping`` 和 ``output_mapping`` 与 fastNLP 的一个特殊的概念 **'参数绑定'** 高度相关,它们的存在也是为了 fastNLP | |||
中的参数匹配能够正确地运行; | |||
.. todo:: | |||
之后链接上 参数匹配 的文档; | |||
.. warning:: | |||
如果 ``Trainer`` 的参数 ``output_mapping`` 不为 ``None``,请保证其返回的一定是一个字典,并且其中含有关键字 **'loss'**; | |||
:param model_wo_auto_param_call: 是否关闭在训练时调用我们的 ``auto_param_call`` 函数来自动匹配 batch 和前向函数的参数的行为; | |||
1. 如果该值为 ``False``,并且当 batch 为字典时,我们会根据**前向函数**所需要的参数从 batch 中提取对应的对象,然后传入到**前向函数**中; | |||
2. 如果该值为 ``True``,那么我们会将 batch 直接透传给模型; | |||
.. todo:: | |||
之后链接上 参数匹配 的文档; | |||
函数 ``auto_param_call`` 详见 :func:`fastNLP.core.utils.auto_param_call`; | |||
:param accumulation_steps: 梯度累积的步数,表示每隔几个 batch 才让优化器迭代一次,默认为 1; | |||
:param fp16: 是否开启混合精度训练,默认为 False; | |||
:param monitor: 对于一些特殊的 ``Callback``,例如 :class:`fastNLP.core.callbacks.CheckpointCallback`,它们需要参数 ``monitor`` | |||
来从 ``Evaluator`` 的验证结果中获取当前评测的值,从而来判断是否执行一些特殊的操作。例如,对于 ``CheckpointCallback`` 而言,如果我们 | |||
想要每隔一个 epoch 让 ``Evaluator`` 进行一次验证,然后保存训练以来的最好的结果;那么我们需要这样设置: | |||
.. code-block:: | |||
trainer = Trainer( | |||
..., | |||
metrics={'acc': accMetric()}, | |||
callbacks=[CheckpointCallback( | |||
..., | |||
monitor='acc', | |||
topk=1 | |||
)] | |||
) | |||
这意味着对于 ``CheckpointCallback`` 来说,*'acc'* 就是一个监测的指标,用于在 ``Evaluator`` 验证后取出其需要监测的那个指标的值。 | |||
``Trainer`` 中的参数 ``monitor`` 的作用在于为没有设置 ``monitor`` 参数但是需要该参数的 *callback* 实例设置该值。关于 ``monitor`` | |||
参数更详细的说明,请见 :class:`fastNLP.core.callbacks.CheckpointCallback`; | |||
注意该参数仅当 ``Trainer`` 内置的 ``Evaluator`` 不为 None 时且有需要该参数但是没有设置该参数的 *callback* 实例才有效; | |||
:param larger_better: 对于需要参数 ``monitor`` 的 *callback* 来说,``monitor`` 的值是否是越大越好;类似于 ``monitor``,其作用 | |||
在于为没有设置 ``larger_better`` 参数但是需要该参数的 *callback* 实例设置该值; | |||
注意该参数仅当 ``Trainer`` 内置的 ``Evaluator`` 不为 None 时且有需要该参数但是没有设置该参数的 *callback* 实例才有效; | |||
:param marker: 用于标记一个 ``Trainer`` 实例,从而在用户调用 ``Trainer.on`` 函数时,标记该函数属于哪一个具体的 ``Trainer`` 实例;默认为 None; | |||
.. note:: | |||
marker 的使用场景主要在于如果一个脚本中含有多个 ``Trainer`` 实例,并且含有多个使用 ``Trainer.on`` 修饰的函数时,不同的函数属于 | |||
不同的 ``Trainer`` 实例; | |||
此时,通过将修饰器 ``Trainer.on`` 的参数 ``marker`` 和 ``Trainer`` 的参数 ``marker`` 置为相同,就可以使得该函数只会在这一 | |||
``Trainer`` 实例中被调用;例如, | |||
.. code-block:: | |||
@Trainer.on(Event.on_train_begin(), marker='trainer1') | |||
def fn(trainer): | |||
... | |||
trainer = Trainer( | |||
..., | |||
marker='trainer1' | |||
) | |||
另一点需要说明的是,如果一个被 ``Trainer.on`` 修饰的函数,其修饰时没有指明 ``marker``,那么会将该函数传给代码位于其之后的 | |||
第一个 ``Trainer`` 实例,即使该 ``Trainer`` 实例的 marker 不为 None;这一点详见 :meth:`~fastNLP.core.controllers.Trainer.on` | |||
:kwargs: | |||
* *torch_kwargs* -- 用于在指定 ``driver`` 为 'torch' 时设定具体 driver 实例的一些参数: | |||
* ddp_kwargs -- 用于在使用 ``TorchDDPDriver`` 时指定 ``DistributedDataParallel`` 初始化时的参数;例如传入 | |||
{'find_unused_parameters': True} 来解决有参数不参与前向运算导致的报错等; | |||
* set_grad_to_none -- 是否在训练过程中在每一次 optimizer 更新后将 grad 置为 None; | |||
* torch_non_blocking -- 表示用于 pytorch 的 tensor 的 to 方法的参数 non_blocking; | |||
* *data_device* -- 表示如果用户的模型 device (在 Driver 中对应为参数 model_device)为 None 时,我们会将数据迁移到 data_device 上; | |||
注意如果 model_device 为 None,那么 data_device 不会起作用; | |||
* *use_dist_sampler* -- 表示是否使用分布式的 sampler 。在多卡时,分布式 sampler 将自动决定每张卡上读取的 sample ,使得一个epoch | |||
内所有卡的 sample 加起来为一整个数据集的 sample。默认会根据 driver 是否为分布式进行设置。 | |||
* *evaluate_use_dist_sampler* -- 表示在 Evaluator 中在使用 分布式 的时候是否将 dataloader 的 sampler 替换为分布式的 sampler;默认为 True; | |||
* *data_device* -- 一个具体的 driver 实例中,有 ``model_device`` 和 ``data_device``,前者表示模型所在的设备,后者表示 | |||
当 ``model_device`` 为 None 时应当将数据迁移到哪个设备; | |||
.. note:: | |||
注意您在绝大部分情况下不会用到该参数! | |||
1. 当 driver 实例的 ``model_device`` 不为 None 时,该参数无效; | |||
2. 对于 pytorch,仅当用户自己通过 ``python -m torch.distributed.launch`` 并且自己初始化 ``init_process_group`` 时, | |||
driver 实例的 ``model_device`` 才会为 None; | |||
* *use_dist_sampler* -- 表示是否使用分布式的 ``sampler``。在多卡时,分布式 ``sampler`` 将自动决定每张卡上读取的 sample ,使得一个 epoch | |||
内所有卡的 sample 加起来为一整个数据集的 sample。默认会根据 driver 是否为分布式进行设置。 | |||
* *evaluate_use_dist_sampler* -- 表示在 ``Evaluator`` 中在使用分布式的时候是否将 dataloader 的 ``sampler`` 替换为分布式的 ``sampler``;默认为 ``True``; | |||
* *output_from_new_proc* -- 应当为一个字符串,表示在多进程的 driver 中其它进程的输出流应当被做如何处理;其值应当为以下之一: | |||
["all", "ignore", "only_error"];当该参数的值不是以上值时,该值应当表示一个文件夹的名字,我们会将其他 rank 的输出流重定向到 | |||
log 文件中,然后将 log 文件保存在通过该参数值设定的文件夹中;默认为 "only_error"; | |||
注意该参数仅当使用分布式的 ``driver`` 时才有效,例如 ``TorchDDPDriver``; | |||
* *progress_bar* -- 以哪种方式显示 progress ,目前支持[None, 'raw', 'rich', 'auto'] 或者 RichCallback, RawTextCallback对象, | |||
默认为 auto , auto 表示如果检测到当前 terminal 为交互型则使用 RichCallback,否则使用 RawTextCallback对象。如果 | |||
需要定制 progress bar 的参数,例如打印频率等,可以传入 RichCallback, RawTextCallback 对象。 | |||
* *train_input_mapping* -- 与 input_mapping 一致,但是只用于 train 中。与 input_mapping 互斥。 | |||
* *train_output_mapping* -- 与 output_mapping 一致,但是只用于 train 中。与 output_mapping 互斥。 | |||
* *evaluate_input_mapping* -- 与 input_mapping 一致,但是只用于 evaluate 中。与 input_mapping 互斥。 | |||
* *evaluate_output_mapping* -- 与 output_mapping 一致,但是只用于 evaluate 中。与 output_mapping 互斥。 | |||
默认为 auto , auto 表示如果检测到当前 terminal 为交互型则使用 RichCallback,否则使用 RawTextCallback对象。如果 | |||
需要定制 progress bar 的参数,例如打印频率等,可以传入 RichCallback, RawTextCallback 对象。 | |||
* *train_input_mapping* -- 与 input_mapping 一致,但是只用于 ``Trainer`` 中。与 input_mapping 互斥。 | |||
* *train_output_mapping* -- 与 output_mapping 一致,但是只用于 ``Trainer`` 中。与 output_mapping 互斥。 | |||
* *evaluate_input_mapping* -- 与 input_mapping 一致,但是只用于 ``Evaluator`` 中。与 input_mapping 互斥。 | |||
* *evaluate_output_mapping* -- 与 output_mapping 一致,但是只用于 ``Evaluator`` 中。与 output_mapping 互斥。 | |||
.. note:: | |||
``Trainer`` 是通过在内部直接初始化一个 ``Evaluator`` 来进行验证; | |||
``Trainer`` 内部的 ``Evaluator`` 默认是 None,如果您需要在训练过程中进行验证,你需要保证这几个参数得到正确的传入: | |||
必须的参数:1. ``metrics``;2. ``evaluate_dataloaders``; | |||
可选的其它参数:1. ``evaluate_batch_step_fn;2. ``evaluate_fn``;3. ``evaluate_every``;4. ``input_mapping``; | |||
5. ``output_mapping``; 6. ``model_wo_auto_param_call``;7. ``fp16``;8. ``monitor``;9. ``larger_better``; | |||
.. warning:: | |||
如果 ``Trainer`` 中内置的 ``Evaluator`` 实例不为 ``None``,那么需要注意 ``Trainer`` 中的一些参数是与 ``Evaluator`` 一致的,它们分别为: | |||
1. ``Evaluator`` 在初始化时的 ``driver`` 参数是 ``Trainer`` 中已经实例化过的 driver;这一点使得一些参数对于 ``Trainer`` 内部的 | |||
``Evaluator`` 没有用处,例如 ``device``,``torch_kwargs``,``data_device`` 和 ``output_from_new_proc`` 等; | |||
2. ``input_mapping``,``output_mapping``,``model_wo_auto_param_call`` 和 ``fp16`` 是 ``Trainer`` 和其内部默认的 | |||
``Evaluator`` 是一致的; | |||
当然,对于 ``input_mapping`` 和 ``output_mapping``,您可以通过添加 ``kwargs`` 中的参数 ``evaluate_input_mapping`` 和 | |||
``evaluate_output_mapping`` 来单独为 ``Evaluator`` 进行更细致的订制。 | |||
另一方面,注意一些专门独属于 ``Evaluator`` 的参数仅当 ``Evaluator`` 不为 None 时才会生效。 | |||
""" | |||
self.model = model | |||
@@ -174,7 +356,7 @@ class Trainer(TrainerEventTrigger): | |||
evaluate_input_mapping = kwargs.get('evaluate_input_mapping', None) | |||
evaluate_output_mapping = kwargs.get('evaluate_output_mapping', None) | |||
train_input_mapping, train_output_mapping, evaluate_input_mapping, evaluate_output_mapping = \ | |||
train_input_mapping, train_output_mapping, evaluate_input_mapping, evaluate_output_mapping = \ | |||
_get_input_output_mapping(input_mapping, output_mapping, train_input_mapping, train_output_mapping, | |||
evaluate_input_mapping, evaluate_output_mapping) | |||
@@ -273,7 +455,7 @@ class Trainer(TrainerEventTrigger): | |||
if not (isinstance(progress_bar, str) or progress_bar is None): # 应该是ProgressCallback,获取其名称。 | |||
progress_bar = progress_bar.name | |||
self.evaluator = Evaluator(model=model, dataloaders=evaluate_dataloaders, metrics=metrics, | |||
driver=self.driver, device=device, evaluate_batch_step_fn=evaluate_batch_step_fn, | |||
driver=self.driver, evaluate_batch_step_fn=evaluate_batch_step_fn, | |||
evaluate_fn=evaluate_fn, input_mapping=evaluate_input_mapping, | |||
output_mapping=evaluate_output_mapping, fp16=fp16, verbose=0, | |||
use_dist_sampler=kwargs.get("evaluate_use_dist_sampler", None), | |||
@@ -302,7 +484,7 @@ class Trainer(TrainerEventTrigger): | |||
def run(self, num_train_batch_per_epoch: int = -1, num_eval_batch_per_dl: int = -1, | |||
num_eval_sanity_batch: int = 2, resume_from: str = None, resume_training: bool = True, | |||
catch_KeyboardInterrupt=None): | |||
""" | |||
r""" | |||
注意如果是断点重训的第一次训练,即还没有保存任何用于断点重训的文件,那么其应当置 resume_from 为 None,并且使用 ModelCheckpoint | |||
去保存断点重训的文件; | |||
:param num_train_batch_per_epoch: 每个 epoch 运行多少个 batch 即停止,-1 为根据 dataloader 有多少个 batch 决定。 | |||
@@ -491,6 +673,36 @@ class Trainer(TrainerEventTrigger): | |||
# do something | |||
# 以上函数会在 Trainer 每个新的 batch 开始的时候执行,但是是两个 batch 才执行一次。 | |||
.. note:: | |||
例如: | |||
.. code-block:: | |||
@Trainer.on(Event.on_train_begin()) | |||
def fn1(trainer): | |||
... | |||
@Trainer.on(Event.on_train_epoch_begin()) | |||
def fn2(trainer): | |||
... | |||
trainer1 = Trainer( | |||
..., | |||
marker='trainer1' | |||
) | |||
@Trainer.on(Event.on_fetch_data_begin()) | |||
def fn3(trainer): | |||
... | |||
trainer2 = Trainer( | |||
..., | |||
marker='trainer2' | |||
) | |||
注意如果你使用该函数修饰器来为你的训练添加 callback,请务必保证你加入 callback 函数的代码在实例化 `Trainer` 之前; | |||
:param event: 特定的 callback 时机,用户需要为该 callback 函数指定其属于哪一个 callback 时机。每个时机运行的函数应该包含 | |||
@@ -646,7 +858,7 @@ class Trainer(TrainerEventTrigger): | |||
self.driver.save_model(folder, only_state_dict, **kwargs) | |||
self.driver.barrier() | |||
def load_model(self, folder: Union[str, Path, BinaryIO, io.BytesIO], only_state_dict: bool = False, | |||
def load_model(self, folder: Union[str, Path, BinaryIO, io.BytesIO], only_state_dict: bool = True, | |||
model_load_fn: Optional[Callable] = None, **kwargs): | |||
""" | |||
加载模型 | |||
@@ -162,9 +162,9 @@ class PaddleDataLoader(DataLoader): | |||
def get_batch_indices(self) -> List[int]: | |||
""" | |||
获取当前 batch 的 idx | |||
获取当前 ``batch`` 中每条数据对应的索引。 | |||
:return: | |||
:return: 当前 ``batch`` 数据的索引 | |||
""" | |||
return self.cur_batch_indices | |||
@@ -10,7 +10,7 @@ from ..samplers import RandomBatchSampler, RandomSampler | |||
from .torch_dataloader import prepare_torch_dataloader | |||
from .paddle_dataloader import prepare_paddle_dataloader | |||
from .jittor_dataloader import prepare_jittor_dataloader | |||
from ...envs import FASTNLP_BACKEND, SUPPORT_BACKENDS, _module_available | |||
from ...envs import FASTNLP_BACKEND, SUPPORT_BACKENDS | |||
from ..log import logger | |||
@@ -170,9 +170,9 @@ class TorchDataLoader(DataLoader): | |||
def get_batch_indices(self) -> List[int]: | |||
""" | |||
获取当前 batch 的 idx | |||
获取当前 ``batch`` 中每条数据对应的索引。 | |||
:return: | |||
:return: 当前 ``batch`` 数据的索引 | |||
""" | |||
return self.cur_batch_indices | |||
@@ -400,15 +400,16 @@ class DataSet: | |||
new_field_name: str = None, num_proc: int = 0, | |||
progress_desc: str = None, show_progress_bar: bool = True): | |||
r""" | |||
将 DataSet 中的每个 instance 中的名为 `field_name` 的 field 传给 func,并获取它的返回值。 | |||
将 :class:`~DataSet` 每个 ``instance`` 中为 ``field_name`` 的 ``field`` 传给函数 ``func``,并获取函数的返回值。 | |||
:param field_name: 传入 func 的是哪个 field。 | |||
:param func: input是 instance 中名为 `field_name` 的 field 的内容。 | |||
:param new_field_name: 将 func 返回的内容放入到 `new_field_name` 这个 field 中,如果名称与已有的 field 相同,则覆 | |||
盖之前的 field。如果为 None 则不创建新的 field。 | |||
:param num_proc: 进程的数量。请注意,由于python语言的特性,多少进程就会导致多少倍内存的增长。 | |||
:param progress_desc: progress_desc 的值,默认为 Main | |||
:param show_progress_bar: 是否展示进度条,默认展示进度条 | |||
:param field_name: 传入 ``func`` 的 ``field`` 名称。 | |||
:param func: 一个函数,其输入是 ``instance`` 中名为 ``field_name`` 的 ``field`` 的内容。 | |||
:param new_field_name: 将 ``func`` 返回的内容放入到 ``new_field_name`` 对应的 ``field`` 中,如果名称与已有的 ``field`` 相同 | |||
则进行覆盖。如果为 ``None`` 则不会覆盖和创建 ``field`` 。 | |||
:param num_proc: 使用进程的数量。请注意,由于 ``python`` 语言的特性,使用了多少进程就会导致多少倍内存的增长。 | |||
:param progress_desc: 进度条的描述字符,默认为 ``Main``。 | |||
:param show_progress_bar: 是否展示进度条;默认为展示。 | |||
:return: 从函数 ``func`` 中得到的返回值。 | |||
""" | |||
assert len(self) != 0, "Null DataSet cannot use apply_field()." | |||
if not self.has_field(field_name=field_name): | |||
@@ -451,8 +452,8 @@ class DataSet: | |||
apply_out = self._apply_process(num_proc, func, progress_desc=progress_desc, | |||
show_progress_bar=show_progress_bar, _apply_field=field_name) | |||
# 只检测第一个数据是否为dict类型,若是则默认所有返回值为dict;否则报错。 | |||
if not isinstance(apply_out[0], dict): | |||
raise Exception("The result of func is not a dict") | |||
if not isinstance(apply_out[0], Mapping): | |||
raise Exception(f"The result of func is not a Mapping, but a {type(apply_out[0])}") | |||
for key, value in apply_out[0].items(): | |||
results[key] = [value] | |||
@@ -9,7 +9,6 @@ __all__ = [ | |||
"JittorDriver", | |||
"JittorSingleDriver", | |||
"JittorMPIDriver", | |||
"TorchPaddleDriver", | |||
'torch_seed_everything', | |||
'paddle_seed_everything', | |||
'optimizer_state_to_device' | |||
@@ -18,7 +17,6 @@ __all__ = [ | |||
from .torch_driver import TorchDriver, TorchSingleDriver, TorchDDPDriver, torch_seed_everything, optimizer_state_to_device | |||
from .jittor_driver import JittorDriver, JittorMPIDriver, JittorSingleDriver | |||
from .paddle_driver import PaddleDriver, PaddleFleetDriver, PaddleSingleDriver, paddle_seed_everything | |||
from .torch_paddle_driver import TorchPaddleDriver | |||
from .driver import Driver | |||
@@ -55,8 +55,8 @@ def initialize_torch_driver(driver: str, device: Optional[Union[str, "torch.devi | |||
elif each < 0: | |||
raise ValueError("When parameter `device` is 'Sequence' type, the value in it should be bigger than 0.") | |||
elif each >= _could_use_device_num: | |||
raise ValueError("When parameter `device` is 'Sequence' type, the value in it should not be bigger than" | |||
" the available gpu number.") | |||
raise ValueError(f"When parameter `device` is 'Sequence' type, the value in it should not be bigger than" | |||
f" the available gpu number:{_could_use_device_num}.") | |||
device = [torch.device(f"cuda:{w}") for w in device] | |||
elif device is not None and not isinstance(device, torch.device): | |||
raise ValueError("Parameter `device` is wrong type, please check our documentation for the right use.") | |||
@@ -167,6 +167,12 @@ class TorchDriver(Driver): | |||
""" | |||
model = self.unwrap_model() | |||
res = torch.load(filepath, map_location='cpu') | |||
if isinstance(res, dict) and only_state_dict is False: | |||
logger.rank_zero_warning(f"It seems like that {filepath} only contains state, you may need to use " | |||
f"`only_state_dict=True`") | |||
elif not isinstance(res, dict) and only_state_dict is True: | |||
logger.rank_zero_warning(f"It seems like that {filepath} is not state, you may need to use " | |||
f"`only_state_dict=False`") | |||
if only_state_dict: | |||
model.load_state_dict(res) | |||
else: | |||
@@ -1,5 +0,0 @@ | |||
__all__ = [ | |||
"TorchPaddleDriver", | |||
] | |||
from .torch_paddle_driver import TorchPaddleDriver |
@@ -1,193 +0,0 @@ | |||
from typing import Optional, Dict, Union, Callable, Tuple | |||
from fastNLP.envs.imports import _NEED_IMPORT_PADDLE, _NEED_IMPORT_TORCH | |||
from fastNLP.core.utils.utils import _get_fun_msg | |||
if _NEED_IMPORT_PADDLE: | |||
import paddle | |||
from paddle.io import DataLoader as PaddleDataLoader | |||
from paddle.optimizer import Optimizer as PaddleOptimizer | |||
if _NEED_IMPORT_TORCH: | |||
import torch | |||
from torch.utils.data import DataLoader as TorchDataLoader | |||
from torch.optim import Optimizer as TorchOptimizer | |||
from fastNLP.core.drivers.driver import Driver | |||
from fastNLP.envs.distributed import rank_zero_call | |||
from fastNLP.core.utils.utils import auto_param_call, apply_to_collection | |||
from fastNLP.core.log.logger import logger | |||
from fastNLP.modules.mix_modules.mix_module import MixModule | |||
__all__ = [ | |||
"TorchPaddleDriver", | |||
] | |||
class TorchPaddleDriver(Driver): | |||
""" | |||
针对torch和paddle混合模型的driver | |||
由于是两种不同的框架不方便实现多卡,暂时先实现CPU和GPU单卡的功能 | |||
""" | |||
def __init__(self, model, device: Optional[str] = None, **kwargs): | |||
super(TorchPaddleDriver, self).__init__(model) | |||
self.model_device = device | |||
self.torch_non_blocking = kwargs.get("torch_non_blocking", None) | |||
self.paddle_blocking = kwargs.get("paddle_blocking", None) | |||
self._data_device = kwargs.get("_data_device", None) | |||
if isinstance(self._data_device, int): | |||
# 将data_device设置为cuda:x的字符串形式 | |||
if self._data_device < 0: | |||
raise ValueError("Parameter `_data_device` can not be smaller than 0.") | |||
_could_use_device_num = paddle.device.cuda.device_count() | |||
if self._data_device >= _could_use_device_num: | |||
raise ValueError("The gpu device that parameter `device` specifies is not existed.") | |||
self._data_device = f"cuda:{self._data_device}" | |||
elif self._data_device is not None: | |||
raise ValueError("Parameter `device` is wrong type, please check our documentation for the right use.") | |||
def setup(self): | |||
if self.model_device is not None: | |||
paddle.device.set_device(self.model_device.replace("cuda", "gpu")) | |||
self.model.to(self.model_device) | |||
@staticmethod | |||
def check_dataloader_legality(dataloader, dataloader_name, is_train: bool = False): | |||
if is_train: | |||
if not isinstance(dataloader, (TorchDataLoader, PaddleDataLoader)): | |||
raise ValueError(f"Parameter `{dataloader_name}` should be 'torch.util.data.DataLoader' or `paddle.io.dataloader` type, not {type(dataloader)}.") | |||
else: | |||
if not isinstance(dataloader, Dict): | |||
raise ValueError(f"Parameter `{dataloader_name}` should be 'Dict' type, not {type(dataloader)}.") | |||
else: | |||
for each_dataloader in dataloader.values(): | |||
if not isinstance(each_dataloader, (TorchDataLoader, PaddleDataLoader)): | |||
raise ValueError(f"Each dataloader of parameter `{dataloader_name}` should be " | |||
f"'torch.util.data.DataLoader' or `paddle.io.dataloader` " | |||
f"type, not {type(each_dataloader)}.") | |||
@staticmethod | |||
def _check_optimizer_legality(optimizers): | |||
for each_optimizer in optimizers: | |||
if not isinstance(each_optimizer, (TorchOptimizer, PaddleOptimizer)): | |||
raise ValueError(f"Each optimizers of parameter `optimizers` should be " | |||
f"'torch.optim.Optimizer' or 'paddle.optimizers.Optimizer' type, " | |||
f"not {type(each_optimizer)}.") | |||
def step(self): | |||
for optimizer in self.optimizers: | |||
optimizer.step() | |||
def backward(self, loss): | |||
loss.backward() | |||
def zero_grad(self): | |||
for optimizer in self.optimizers: | |||
if isinstance(optimizer, TorchOptimizer): | |||
optimizer.zero_grad() | |||
elif isinstance(optimizer, PaddleOptimizer): | |||
optimizer.clear_grad() | |||
else: | |||
raise ValueError("Unknown optimizers type.") | |||
def model_call(self, batch, fn: Callable, signature_fn: Optional[Callable]) -> Dict: | |||
if isinstance(batch, Dict) and not self.wo_auto_param_call: | |||
return auto_param_call(fn, batch, signature_fn=signature_fn) | |||
else: | |||
return fn(batch) | |||
def get_model_call_fn(self, fn: str) -> Tuple: | |||
if hasattr(self.model, fn): | |||
fn = getattr(self.model, fn) | |||
if not callable(fn): | |||
raise RuntimeError(f"The `{fn}` attribute is not `Callable`.") | |||
logger.debug(f'Use {_get_fun_msg(fn, with_fp=False)}...') | |||
return fn, None | |||
elif fn in {"train_step", "evaluate_step"}: | |||
logger.debug(f'Use {_get_fun_msg(self.model.forward, with_fp=False)}...') | |||
return self.model, self.model.forward | |||
else: | |||
raise RuntimeError(f"There is no `{fn}` method in your {type(self.model)}.") | |||
def predict_step(self, batch): | |||
if isinstance(batch, Dict): | |||
return auto_param_call(self._predict_step, batch) | |||
else: | |||
return self._predict_step(batch) | |||
@rank_zero_call | |||
def save_model(self, filepath: str, only_state_dict: bool = True, model_save_fn: Optional[Callable] = None): | |||
r""" | |||
暂时不提供保存整个模型的方法 | |||
""" | |||
if only_state_dict == False: | |||
logger.warn("TorchPaddleModule only support saving state dicts now.") | |||
if model_save_fn is not None: | |||
model_save_fn(filepath) | |||
else: | |||
model = self.unwrap_model() | |||
self.move_model_to_device(model, "cpu") | |||
self.model.save(filepath) | |||
self.move_model_to_device(model, self.model_device) | |||
def load_model(self, filepath: str): | |||
""" | |||
加载模型的加载函数; | |||
:param filepath: 保存文件的文件位置(需要包括文件名); | |||
:return: | |||
""" | |||
return self.model.load(filepath) | |||
def save(self): | |||
... | |||
def load(self): | |||
... | |||
@staticmethod | |||
def move_model_to_device(model: MixModule, device: str): | |||
if device is not None: | |||
model.to(device) | |||
def unwrap_model(self): | |||
return self.model | |||
@staticmethod | |||
def tensor_to_numeric(tensor): | |||
if tensor is None: | |||
return None | |||
def _translate(_data): | |||
return _data.tolist() | |||
return apply_to_collection( | |||
data=tensor, | |||
dtype=(paddle.Tensor, torch.Tensor), | |||
function=_translate | |||
) | |||
def set_model_mode(self, mode: str): | |||
assert mode in {"train", "eval"} | |||
getattr(self.model, mode)() | |||
def get_model_device(self): | |||
return self.model_device | |||
@property | |||
def data_device(self): | |||
if self.model_device is not None: | |||
return self.model_device | |||
else: | |||
return self._data_device | |||
def set_model_mode(self, mode: str): | |||
assert mode in {"train", "eval"} | |||
getattr(self.model, mode)() | |||
def set_sampler_epoch(self, dataloader: Union['TorchDataLoader', 'PaddleDataLoader'], cur_epoch_idx): | |||
# 保证 ddp 训练时的 shuffle=True 时的正确性,因为需要保证每一个进程上的 sampler 的shuffle 的随机数种子是一样的; | |||
return dataloader |
@@ -1,4 +0,0 @@ | |||
from fastNLP.envs.imports import _NEED_IMPORT_PADDLE | |||
if _NEED_IMPORT_PADDLE: | |||
pass |
@@ -11,7 +11,6 @@ __all__ = [ | |||
'is_in_fnlp_paddle_dist', | |||
'is_in_paddle_launch_dist', | |||
'f_rich_progress', | |||
'torch_paddle_move_data_to_device', | |||
'torch_move_data_to_device', | |||
'get_fn_arg_names', | |||
'auto_param_call', | |||
@@ -32,7 +31,6 @@ from .jittor_utils import is_jittor_dataset, jittor_collate_wraps | |||
from .paddle_utils import get_device_from_visible, paddle_to, paddle_move_data_to_device, get_paddle_device_id, get_paddle_gpu_str, is_in_paddle_dist, \ | |||
is_in_fnlp_paddle_dist, is_in_paddle_launch_dist | |||
from .rich_progress import f_rich_progress | |||
from .torch_paddle_utils import torch_paddle_move_data_to_device | |||
from .torch_utils import torch_move_data_to_device | |||
from .utils import * | |||
@@ -1,4 +1,4 @@ | |||
import functools | |||
__all__ = [] | |||
class DummyClass: | |||
def __init__(self, *args, **kwargs): | |||
@@ -1,7 +1,6 @@ | |||
""" | |||
该文件用于为fastNLP提供一个统一的progress bar管理,通过共用一个Task对象,trainer中的progress bar和evaluation中的progress bar才能 | |||
不冲突 | |||
该文件用于为 ``fastNLP`` 提供一个统一的 ``progress bar`` 管理,通过共用一个``Task`` 对象, :class:`~fastNLP.core.Trainer` 中 | |||
的 ``progress bar`` 和 :class:`~fastNLP.core.Evaluator` 中的 ``progress bar`` 才能不冲突 | |||
""" | |||
import sys | |||
from typing import Any, Union, Optional | |||
@@ -1,49 +0,0 @@ | |||
from typing import Any, Optional | |||
from fastNLP.envs.imports import _NEED_IMPORT_PADDLE, _NEED_IMPORT_TORCH | |||
if _NEED_IMPORT_PADDLE: | |||
import paddle | |||
if _NEED_IMPORT_TORCH: | |||
import torch | |||
__all__ = [ | |||
"torch_paddle_move_data_to_device", | |||
] | |||
from .utils import apply_to_collection | |||
from .paddle_utils import paddle_to | |||
def torch_paddle_move_data_to_device(batch: Any, device: Optional[str] = None, non_blocking: Optional[bool] = True, | |||
data_device: Optional[str] = None) -> Any: | |||
r""" | |||
将数据集合传输到给定设备。只有paddle.Tensor和torch.Tensor对象会被传输到设备中,其余保持不变 | |||
:param batch: | |||
:param device: | |||
:param non_blocking: | |||
:param data_device: | |||
:return: 相同的集合,但所有包含的张量都驻留在新设备上; | |||
""" | |||
if device is None: | |||
if data_device is not None: | |||
device = data_device | |||
else: | |||
return batch | |||
torch_device = device.replace("gpu", "cuda") | |||
paddle_device = device.replace("cuda", "gpu") | |||
def batch_to(data: Any) -> Any: | |||
if isinstance(data, torch.Tensor): | |||
data = data.to(torch_device, non_blocking=non_blocking) | |||
elif isinstance(data, paddle.Tensor): | |||
data = paddle_to(data, paddle_device) | |||
return data | |||
return apply_to_collection(batch, dtype=(paddle.Tensor, torch.Tensor), function=batch_to) |
@@ -10,10 +10,6 @@ from typing import Callable, List, Any, Dict, AnyStr, Union, Mapping, Sequence | |||
from typing import Tuple, Optional | |||
from time import sleep | |||
try: | |||
from typing import Literal, Final | |||
except ImportError: | |||
from typing_extensions import Literal, Final | |||
import os | |||
from contextlib import contextmanager | |||
from functools import wraps | |||
@@ -22,7 +18,6 @@ import numpy as np | |||
from pathlib import Path | |||
from fastNLP.core.log import logger | |||
from ...envs import SUPPORT_BACKENDS | |||
__all__ = [ | |||
@@ -43,10 +38,10 @@ __all__ = [ | |||
def get_fn_arg_names(fn: Callable) -> List[str]: | |||
r""" | |||
返回一个函数的所有参数的名字; | |||
返回一个函数所有参数的名字 | |||
:param fn: 需要查询的函数; | |||
:return: 一个列表,其中的元素则是查询函数的参数的字符串名字; | |||
:param fn: 需要查询的函数 | |||
:return: 一个列表,其中的元素是函数 ``fn`` 参数的字符串名字 | |||
""" | |||
return list(inspect.signature(fn).parameters) | |||
@@ -54,24 +49,18 @@ def get_fn_arg_names(fn: Callable) -> List[str]: | |||
def auto_param_call(fn: Callable, *args, signature_fn: Optional[Callable] = None, | |||
mapping: Optional[Dict[AnyStr, AnyStr]] = None) -> Any: | |||
r""" | |||
该函数会根据输入函数的形参名从*args(因此都需要是dict类型)中找到匹配的值进行调用,如果传入的数据与fn的形参不匹配,可以通过mapping | |||
参数进行转换。mapping参数中的一对(key,value)表示以这个key在*args中找到值,并将这个值传递给形参名为value的参数。 | |||
该函数会根据输入函数的形参名从 ``*args`` (因此都需要是 ``dict`` 类型)中找到匹配的值进行调用,如果传入的数据与 ``fn`` 的形参不匹配,可以通过 | |||
``mapping`` 参数进行转换。``mapping`` 参数中的一对 ``(key, value)`` 表示在 ``*args`` 中找到 ``key`` 对应的值,并将这个值传递给形参中名为 | |||
``value`` 的参数。 | |||
1.该函数用来提供给用户根据字符串匹配从而实现自动调用; | |||
2.注意 mapping 默认为 None,如果你希望指定输入和运行函数的参数的对应方式,那么你应当让 mapping 为一个这样的字典传入进来; | |||
如果 mapping 不为 None,那么我们一定会先使用 mapping 将输入的字典的 keys 修改过来,因此请务必亲自检查 mapping 的正确性; | |||
3.如果输入的函数的参数有默认值,那么如果之后的输入中没有该参数对应的值,我们就会使用该参数对应的默认值,否则也会使用之后的输入的值; | |||
4.如果输入的函数是一个 `partial` 函数,情况同 '3.',即和默认参数的情况相同; | |||
:param fn: 用来进行实际计算的函数,其参数可以包含有默认值; | |||
:param args: 一系列的位置参数,应当为一系列的字典,我们需要从这些输入中提取 `fn` 计算所需要的实际参数; | |||
:param signature_fn: 函数,用来替换 `fn` 的函数签名,如果该参数不为 None,那么我们首先会从该函数中提取函数签名,然后通过该函数签名提取 | |||
参数值后,再传给 `fn` 进行实际的运算; | |||
:param mapping: 一个字典,用来更改其前面的字典的键值; | |||
:return: 返回 `fn` 运行的结果; | |||
1. 该函数用来提供给用户根据字符串匹配从而实现自动调用; | |||
2. 注意 ``mapping`` 默认为 ``None``,如果你希望指定输入和运行函数的参数的对应方式,那么你应当让 ``mapping`` 为一个字典传入进来; | |||
如果 ``mapping`` 不为 ``None``,那么我们一定会先使用 ``mapping`` 将输入的字典的 ``keys`` 修改过来,因此请务必亲自检查 ``mapping`` 的正确性; | |||
3. 如果输入的函数的参数有默认值,那么如果之后的输入中没有该参数对应的值,我们就会使用该参数对应的默认值,否则也会使用之后的输入的值; | |||
4. 如果输入的函数是一个 ``partial`` 函数,情况同第三点,即和默认参数的情况相同; | |||
Examples:: | |||
>>> # 1 | |||
>>> loss_fn = CrossEntropyLoss() # 如果其需要的参数为 def CrossEntropyLoss(y, pred); | |||
>>> batch = {"x": 20, "y": 1} | |||
@@ -84,6 +73,14 @@ def auto_param_call(fn: Callable, *args, signature_fn: Optional[Callable] = None | |||
>>> print(auto_param_call(test_fn, {"x": 10}, {"y": 20, "a": 30})) # res: 70 | |||
>>> print(auto_param_call(partial(test_fn, a=100), {"x": 10}, {"y": 20})) # res: 140 | |||
>>> print(auto_param_call(partial(test_fn, a=100), {"x": 10}, {"y": 20, "a": 200})) # res: 240 | |||
:param fn: 用来进行实际计算的函数,其参数可以包含有默认值; | |||
:param args: 一系列的位置参数,应当为一系列的字典,我们需要从这些输入中提取 ``fn`` 计算所需要的实际参数; | |||
:param signature_fn: 函数,用来替换 ``fn`` 的函数签名,如果该参数不为 ``None``,那么我们首先会从该函数中提取函数签名,然后通过该函数签名提取 | |||
参数值后,再传给 ``fn`` 进行实际的运算; | |||
:param mapping: 一个字典,用来更改其前面的字典的键值; | |||
:return: 返回 ``fn`` 运行的结果; | |||
""" | |||
if signature_fn is not None: | |||
@@ -226,13 +223,13 @@ def _check_valid_parameters_number(fn, expected_params:List[str], fn_name=None): | |||
def check_user_specific_params(user_params: Dict, fn: Callable): | |||
""" | |||
该函数使用用户的输入来对指定函数的参数进行赋值; | |||
主要用于一些用户无法直接调用函数的情况; | |||
该函数主要的作用在于帮助检查用户对使用函数 fn 的参数输入是否有误; | |||
该函数使用用户的输入来对指定函数的参数进行赋值,主要用于一些用户无法直接调用函数的情况; | |||
该函数主要的作用在于帮助检查用户对使用函数 ``fn`` 的参数输入是否有误; | |||
:param user_params: 用户指定的参数的值,应当是一个字典,其中 key 表示每一个参数的名字,value 为每一个参数应当的值; | |||
:param fn: 会被调用的函数; | |||
:return: 返回一个字典,其中为在之后调用函数 fn 时真正会被传进去的参数的值; | |||
:param user_params: 用户指定的参数的值,应当是一个字典,其中 ``key`` 表示每一个参数的名字, | |||
``value`` 为每一个参数的值; | |||
:param fn: 将要被调用的函数; | |||
:return: 返回一个字典,其中为在之后调用函数 ``fn`` 时真正会被传进去的参数的值; | |||
""" | |||
fn_arg_names = get_fn_arg_names(fn) | |||
@@ -243,6 +240,9 @@ def check_user_specific_params(user_params: Dict, fn: Callable): | |||
def dataclass_to_dict(data: "dataclasses.dataclass") -> Dict: | |||
""" | |||
将传入的 `dataclass` 实例转换为字典。 | |||
""" | |||
if not is_dataclass(data): | |||
raise TypeError(f"Parameter `data` can only be `dataclass` type instead of {type(data)}.") | |||
_dict = dict() | |||
@@ -253,21 +253,31 @@ def dataclass_to_dict(data: "dataclasses.dataclass") -> Dict: | |||
def match_and_substitute_params(mapping: Optional[Union[Callable, Dict]] = None, data: Optional[Any] = None) -> Any: | |||
r""" | |||
用来实现将输入:batch,或者输出:outputs,通过 `mapping` 将键值进行更换的功能; | |||
该函数应用于 `input_mapping` 和 `output_mapping`; | |||
对于 `input_mapping`,该函数会在 `TrainBatchLoop` 中取完数据后立刻被调用; | |||
对于 `output_mapping`,该函数会在 `Trainer.train_step` 以及 `Evaluator.train_step` 中得到结果后立刻被调用; | |||
用来实现将输入的 ``batch``,或者输出的 ``outputs``,通过 ``mapping`` 将键值进行更换的功能; | |||
该函数应用于 ``input_mapping`` 和 ``output_mapping``; | |||
转换的逻辑按优先级依次为: | |||
对于 ``input_mapping``,该函数会在 :class:`~fastNLP.core.controllers.TrainBatchLoop` 中取完数据后立刻被调用; | |||
对于 ``output_mapping``,该函数会在 :class:`~fastNLP.core.Trainer` 的 :meth:`~fastNLP.core.Trainer.train_step` | |||
以及 :class:`~fastNLP.core.Evaluator` 的 :meth:`~fastNLP.core.Evaluator.train_step` 中得到结果后立刻被调用; | |||
1. 如果 `mapping` 是一个函数,那么会直接返回 `mapping(data)`; | |||
2. 如果 `mapping` 是一个 `Dict`,那么 `data` 的类型只能为以下三种: [`Dict`, `dataclass`, `Sequence`]; | |||
如果 `data` 是 `Dict`,那么该函数会将 `data` 的 key 替换为 mapping[key]; | |||
如果 `data` 是 `dataclass`,那么该函数会先使用 `dataclasses.asdict` 函数将其转换为 `Dict`,然后进行转换; | |||
如果 `data` 是 `Sequence`,那么该函数会先将其转换成一个对应的 `Dict`:{"_0": list[0], "_1": list[1], ...},然后使用 | |||
mapping对这个 `Dict` 进行转换,如果没有匹配上mapping中的key则保持"_number"这个形式。 | |||
转换的逻辑按优先级依次为: | |||
:param mapping: 用于转换的字典或者函数;mapping是函数时,返回值必须为字典类型。 | |||
1. 如果 ``mapping`` 是一个函数,那么会直接返回 ``mapping(data)``; | |||
2. 如果 ``mapping`` 是一个 ``Dict``,那么 ``data`` 的类型只能为以下三种: ``[Dict, dataclass, Sequence]``; | |||
* 如果 ``data`` 是 ``Dict``,那么该函数会将 ``data`` 的 ``key`` 替换为 ``mapping[key]``; | |||
* 如果 ``data`` 是 ``dataclass``,那么该函数会先使用 :func:`dataclasses.asdict` 函数将其转换为 ``Dict``,然后进行转换; | |||
* 如果 ``data`` 是 ``Sequence``,那么该函数会先将其转换成一个对应的字典:: | |||
{ | |||
"_0": list[0], | |||
"_1": list[1], | |||
... | |||
} | |||
然后使用 ``mapping`` 对这个 ``Dict`` 进行转换,如果没有匹配上 ``mapping`` 中的 ``key`` 则保持 ``\'\_number\'`` 这个形式。 | |||
:param mapping: 用于转换的字典或者函数;``mapping`` 是函数时,返回值必须为字典类型。 | |||
:param data: 需要被转换的对象; | |||
:return: 返回转换好的结果; | |||
""" | |||
@@ -320,21 +330,20 @@ def apply_to_collection( | |||
include_none: bool = True, | |||
**kwargs: Any, | |||
) -> Any: | |||
"""将函数 function 递归地在 data 中的元素执行,但是仅在满足元素为 dtype 时执行。 | |||
this function credit to: https://github.com/PyTorchLightning/pytorch-lightning | |||
Args: | |||
data: the collection to apply the function to | |||
dtype: the given function will be applied to all elements of this dtype | |||
function: the function to apply | |||
*args: positional arguments (will be forwarded to calls of ``function``) | |||
wrong_dtype: the given function won't be applied if this type is specified and the given collections | |||
is of the ``wrong_dtype`` even if it is of type ``dtype`` | |||
include_none: Whether to include an element if the output of ``function`` is ``None``. | |||
**kwargs: keyword arguments (will be forwarded to calls of ``function``) | |||
Returns: | |||
The resulting collection | |||
""" | |||
使用函数 ``function`` 递归地在 ``data`` 中的元素执行,但是仅在满足元素为 ``dtype`` 时执行。 | |||
该函数参考了 `pytorch-lightning <https://github.com/PyTorchLightning/pytorch-lightning>`_ 的实现 | |||
:param data: 需要进行处理的数据集合或数据 | |||
:param dtype: 数据的类型,函数 ``function`` 只会被应用于 ``data`` 中类型为 ``dtype`` 的数据 | |||
:param function: 对数据进行处理的函数 | |||
:param args: ``function`` 所需要的其它参数 | |||
:param wrong_dtype: ``function`` 一定不会生效的数据类型。如果数据既是 ``wrong_dtype`` 类型又是 ``dtype`` 类型 | |||
那么也不会生效。 | |||
:param include_none: 是否包含执行结果为 ``None`` 的数据,默认为 ``True``。 | |||
:param kwargs: ``function`` 所需要的其它参数 | |||
:return: 经过 ``function`` 处理后的数据集合 | |||
""" | |||
# Breaking condition | |||
if isinstance(data, dtype) and (wrong_dtype is None or not isinstance(data, wrong_dtype)): | |||
@@ -402,16 +411,18 @@ def apply_to_collection( | |||
@contextmanager | |||
def nullcontext(): | |||
r""" | |||
用来实现一个什么 dummy 的 context 上下文环境; | |||
实现一个什么都不做的上下文环境 | |||
""" | |||
yield | |||
def sub_column(string: str, c: int, c_size: int, title: str) -> str: | |||
r""" | |||
对传入的字符串进行截断,方便在命令行中显示 | |||
:param string: 要被截断的字符串 | |||
:param c: 命令行列数 | |||
:param c_size: instance或dataset field数 | |||
:param c_size: :class:`~fastNLP.core.Instance` 或 :class:`fastNLP.core.DataSet` 的 ``field`` 数目 | |||
:param title: 列名 | |||
:return: 对一个过长的列进行截断的结果 | |||
""" | |||
@@ -442,18 +453,17 @@ def _is_iterable(value): | |||
def pretty_table_printer(dataset_or_ins) -> PrettyTable: | |||
r""" | |||
:param dataset_or_ins: 传入一个dataSet或者instance | |||
.. code-block:: | |||
在 ``fastNLP`` 中展示数据的函数:: | |||
ins = Instance(field_1=[1, 1, 1], field_2=[2, 2, 2], field_3=["a", "b", "c"]) | |||
>>> ins = Instance(field_1=[1, 1, 1], field_2=[2, 2, 2], field_3=["a", "b", "c"]) | |||
+-----------+-----------+-----------------+ | |||
| field_1 | field_2 | field_3 | | |||
+-----------+-----------+-----------------+ | |||
| [1, 1, 1] | [2, 2, 2] | ['a', 'b', 'c'] | | |||
+-----------+-----------+-----------------+ | |||
:return: 以 pretty table的形式返回根据terminal大小进行自动截断 | |||
:param dataset_or_ins: 要展示的 :class:`~fastNLP.core.DataSet` 或者 :class:`~fastNLP.core.Instance` | |||
:return: 根据 ``terminal`` 大小进行自动截断的数据表格 | |||
""" | |||
x = PrettyTable() | |||
try: | |||
@@ -486,7 +496,7 @@ def pretty_table_printer(dataset_or_ins) -> PrettyTable: | |||
class Option(dict): | |||
r"""a dict can treat keys as attributes""" | |||
r"""将键转化为属性的字典类型""" | |||
def __getattr__(self, item): | |||
try: | |||
@@ -516,11 +526,10 @@ _emitted_deprecation_warnings = set() | |||
def deprecated(help_message: Optional[str] = None): | |||
"""Decorator to mark a function as deprecated. | |||
""" | |||
标记当前功能已经过时的装饰器。 | |||
Args: | |||
help_message (`Optional[str]`): An optional message to guide the user on how to | |||
switch to non-deprecated usage of the library. | |||
:param help_message: 一段指引信息,告知用户如何将代码切换为当前版本提倡的用法。 | |||
""" | |||
def decorator(deprecated_function: Callable): | |||
@@ -549,11 +558,10 @@ def deprecated(help_message: Optional[str] = None): | |||
return decorator | |||
def seq_len_to_mask(seq_len, max_len=None): | |||
def seq_len_to_mask(seq_len, max_len: Optional[int]): | |||
r""" | |||
将一个表示sequence length的一维数组转换为二维的mask,不包含的位置为0。 | |||
转变 1-d seq_len到2-d mask. | |||
将一个表示 ``sequence length`` 的一维数组转换为二维的 ``mask`` ,不包含的位置为 **0**。 | |||
.. code-block:: | |||
@@ -570,10 +578,11 @@ def seq_len_to_mask(seq_len, max_len=None): | |||
>>>print(mask.size()) | |||
torch.Size([14, 100]) | |||
:param np.ndarray,torch.LongTensor seq_len: shape将是(B,) | |||
:param int max_len: 将长度pad到这个长度。默认(None)使用的是seq_len中最长的长度。但在nn.DataParallel的场景下可能不同卡的seq_len会有 | |||
区别,所以需要传入一个max_len使得mask的长度是pad到该长度。 | |||
:return: np.ndarray, torch.Tensor 。shape将是(B, max_length), 元素类似为bool或torch.uint8 | |||
:param seq_len: 大小为是 ``(B,)`` 的长度序列 | |||
:param int max_len: 将长度 ``pad`` 到 ``max_len``。默认情况(为 ``None``)使用的是 ``seq_len`` 中最长的长度。 | |||
但在 :class:`torch.nn.DataParallel` 等分布式的场景下可能不同卡的 ``seq_len`` 会有区别,所以需要传入 | |||
一个 ``max_len`` 使得 ``mask`` 的长度 ``pad`` 到该长度。 | |||
:return: 大小为 ``(B, max_len)`` 的 ``mask``, 元素类型为 ``bool`` 或 ``uint8`` | |||
""" | |||
if isinstance(seq_len, np.ndarray): | |||
assert len(np.shape(seq_len)) == 1, f"seq_len can only have one dimension, got {len(np.shape(seq_len))}." | |||
@@ -6,6 +6,7 @@ from packaging.version import Version | |||
import subprocess | |||
import pkg_resources | |||
__all__ = [] | |||
def _module_available(module_path: str) -> bool: | |||
"""Check if a path is available in your environment. | |||
@@ -48,10 +49,11 @@ def _compare_version(package: str, op: Callable, version: str, use_base_version: | |||
pkg_version = Version(pkg_version.base_version) | |||
return op(pkg_version, Version(version)) | |||
def get_gpu_count(): | |||
def get_gpu_count() -> int: | |||
""" | |||
利用命令行获取gpu数目的函数 | |||
:return: gpu数目,如果没有显卡设备则为-1 | |||
利用命令行获取 ``gpu`` 数目的函数 | |||
:return: 显卡数目,如果没有显卡设备则为-1 | |||
""" | |||
try: | |||
lines = subprocess.check_output(['nvidia-smi', '--query-gpu=memory.used', '--format=csv']) | |||
@@ -1,9 +0,0 @@ | |||
__all__ = [ | |||
"MixModule", | |||
"torch2paddle", | |||
"paddle2torch", | |||
"torch2jittor", | |||
"jittor2torch", | |||
] | |||
from .mix_modules import MixModule, torch2paddle, paddle2torch, torch2jittor, jittor2torch |
@@ -1,10 +0,0 @@ | |||
__all__ = [ | |||
"MixModule", | |||
"torch2paddle", | |||
"paddle2torch", | |||
"torch2jittor", | |||
"jittor2torch", | |||
] | |||
from .mix_module import MixModule | |||
from .utils import * |
@@ -1,310 +0,0 @@ | |||
import os | |||
import io | |||
import pickle | |||
from typing import Dict | |||
from collections import OrderedDict | |||
import numpy as np | |||
from fastNLP.envs.imports import _NEED_IMPORT_JITTOR, _NEED_IMPORT_PADDLE, _NEED_IMPORT_TORCH | |||
from fastNLP.core.utils.paddle_utils import paddle_to | |||
if _NEED_IMPORT_PADDLE: | |||
import paddle | |||
from paddle.nn import Layer as PaddleLayer | |||
if _NEED_IMPORT_TORCH: | |||
import torch | |||
from torch.nn import Module as TorchModule, Parameter as TorchParameter | |||
if _NEED_IMPORT_JITTOR: | |||
import jittor | |||
__all__ = [ | |||
"MixModule", | |||
] | |||
class MixModule: | |||
""" | |||
TODO: 支持不同的混合方式;添加state_dict的支持;如果参数里有List of Tensors该怎么处理; | |||
是否需要仿照Module那样在初始化的时候给各种模型分类 | |||
可以同时使用Torch和Paddle框架的混合模型 | |||
""" | |||
def __init__(self, *args, **kwargs): | |||
pass | |||
def __call__(self, *args, **kwargs): | |||
return self.forward(*args, **kwargs) | |||
def named_parameters(self, prefix='', recurse: bool=True, backend=None): | |||
""" | |||
返回模型的名字和参数 | |||
:param prefix: 输出时在参数名前加上的前缀 | |||
:param recurse: 是否递归地输出参数 | |||
:param backend: `backend`=`None`时,将所有模型和张量的参数返回; | |||
`backend`=`torch`时,返回`torch`的参数; | |||
`backend`=`paddle`时,返回`paddle`的参数。 | |||
""" | |||
if backend is None: | |||
generator = self.attributes(TorchModule, TorchParameter, PaddleLayer) | |||
elif backend == "torch": | |||
generator = self.attributes(TorchModule, TorchParameter) | |||
elif backend == "paddle": | |||
generator = self.attributes(PaddleLayer) | |||
else: | |||
raise ValueError("Unknown backend parameter.") | |||
for name, value in generator: | |||
name = prefix + ('.' if prefix else '') + name | |||
if isinstance(value, TorchParameter): | |||
# 非Module/Layer类型,直接输出名字和值 | |||
yield name, value | |||
elif recurse: | |||
# 递归地调用named_parameters | |||
for name_r, value_r in value.named_parameters(name, recurse): | |||
yield name_r, value_r | |||
def parameters(self, recurse: bool = True, backend: str = None): | |||
""" | |||
返回模型的参数 | |||
:param recurse: | |||
:param backend: `backend`=`None`时,将所有模型和张量的参数返回; | |||
`backend`=`torch`时,返回`torch`的参数; | |||
`backend`=`paddle`时,返回`paddle`的参数。 | |||
""" | |||
for name, value in self.named_parameters(recurse=recurse, backend=backend): | |||
yield value | |||
def forward(self, *args, **kwargs): | |||
raise NotImplementedError | |||
def train_step(self, batch): | |||
raise NotImplementedError | |||
def test_step(self, batch): | |||
raise NotImplementedError | |||
def evaluate_step(self, batch): | |||
raise NotImplementedError | |||
def train(self): | |||
for name, value in self.attributes(TorchModule, PaddleLayer): | |||
value.train() | |||
def eval(self): | |||
for name, value in self.attributes(TorchModule, PaddleLayer): | |||
value.eval() | |||
def to(self, device): | |||
""" | |||
:param device: 设备名 | |||
""" | |||
# 有jittor的话 warning | |||
if device == "cpu": | |||
paddle_device = device | |||
elif device.startswith("cuda"): | |||
paddle_device = device.replace("cuda", "gpu") | |||
elif device.startswith("gpu"): | |||
paddle_device = device | |||
device = device.replace("gpu", "cuda") | |||
else: | |||
raise ValueError("Device value error") | |||
for name, value in self.attributes(TorchModule): | |||
# torch的to函数不影响Tensor | |||
vars(self)[name] = value.to(device) | |||
for name, value in self.attributes(TorchParameter): | |||
# Parameter在经过to函数后会变成Tensor类型 | |||
vars(self)[name] = TorchParameter(value.to(device), requires_grad=value.requires_grad) | |||
for name, value in self.attributes(PaddleLayer): | |||
vars(self)[name] = value.to(paddle_device) | |||
for name, value in self.attributes(paddle.Tensor): | |||
# paddle的to函数会影响到Tensor | |||
vars(self)[name] = paddle_to(value, paddle_device) | |||
return self | |||
def state_dict(self, backend: str = None) -> Dict: | |||
""" | |||
返回模型的state_dict。 | |||
.. note:: torch的destination参数会在将来删除,因此不提供destination参数 | |||
:param backend: `backend`=`None`时,将所有模型和张量的state dict返回; | |||
`backend`=`torch`时,返回`torch`的state dict; | |||
`backend`=`paddle`时,返回`paddle`的state dict。 | |||
""" | |||
if backend is None: | |||
generator = self.attributes(TorchModule, TorchParameter, PaddleLayer) | |||
elif backend == "torch": | |||
generator = self.attributes(TorchModule, TorchParameter) | |||
elif backend == "paddle": | |||
generator = self.attributes(PaddleLayer) | |||
else: | |||
raise ValueError(f"Unknown backend {backend}.") | |||
destination = OrderedDict() | |||
for name, value in generator: | |||
if value is None: | |||
continue | |||
if isinstance(value, TorchParameter): | |||
destination[name] = value | |||
else: | |||
# 不同框架state_dict函数的参数名和顺序不同 | |||
if isinstance(value, PaddleLayer): | |||
kwargs = { | |||
"structured_name_prefix": name + ".", | |||
} | |||
elif isinstance(value, TorchModule): | |||
kwargs = { | |||
"prefix": name + ".", | |||
} | |||
else: | |||
raise ValueError(f"Unknown item type {type(value)}") | |||
destination.update(value.state_dict(**kwargs)) | |||
return destination | |||
def save_state_dict_to_file(self, path: str): | |||
""" | |||
保存模型的state dict到path | |||
""" | |||
# TODO 设备限制 | |||
filename = os.path.basename(path) | |||
if filename == "": | |||
raise ValueError("Received empty filename.") | |||
dirname = os.path.dirname(path) | |||
if dirname and not os.path.exists(dirname): | |||
os.makedirs(dirname) | |||
protocol = 4 | |||
saved = {} | |||
paddle_dict = self.state_dict(backend="paddle") | |||
torch_dict = self.state_dict(backend="torch") | |||
# 保存paddle部分 | |||
# 调用paddle保存时的处理函数 | |||
paddle_saved_obj = paddle.framework.io._build_saved_state_dict(paddle_dict) | |||
paddle_saved_obj = paddle.fluid.io._unpack_saved_dict(paddle_saved_obj, protocol) | |||
# 将返回的dict保存 | |||
saved["paddle"] = paddle_saved_obj | |||
# 保存torch部分 | |||
buffer = io.BytesIO() | |||
torch.save(torch_dict, buffer) | |||
saved["torch"] = buffer.getvalue() | |||
# 保存 | |||
with open(path, "wb") as f: | |||
pickle.dump(saved, f, protocol) | |||
def load_state_dict_from_file(self, path: str): | |||
""" | |||
从 `path` 中加载保存的state dict | |||
""" | |||
state_dict = {} | |||
with open(path, "rb") as f: | |||
loaded = pickle.load(f) | |||
# 加载paddle的数据 | |||
paddle_loaded_obj = loaded["paddle"] | |||
paddle_load_result = paddle.fluid.io._pack_loaded_dict(paddle_loaded_obj) | |||
if "StructuredToParameterName@@" in paddle_load_result: | |||
for key in paddle_load_result["StructuredToParameterName@@"]: | |||
if isinstance(paddle_load_result[key], np.ndarray): | |||
paddle_load_result[key] = paddle.to_tensor(paddle_load_result[key]) | |||
state_dict.update(paddle_load_result) | |||
# 加载torch的数据 | |||
torch_loaded_obj = loaded["torch"] | |||
torch_bytes = io.BytesIO(torch_loaded_obj) | |||
torch_load_result = torch.load(torch_bytes) | |||
state_dict.update(torch_load_result) | |||
self.load_state_dict(state_dict) | |||
def load_state_dict(self, state_dict): | |||
""" | |||
从state dict中加载数据 | |||
""" | |||
missing_keys = [] | |||
unexpected_keys = [] | |||
error_msgs = [] | |||
new_state = {} | |||
local_state = self.state_dict() | |||
# 对字典内容按前缀进行归类 | |||
for key, value in state_dict.items(): | |||
splited = key.split(".", 1) | |||
if len(splited) == 1: | |||
# 没有前缀,实际上只有torch.nn.Parameter会进入这种情况 | |||
new_state[key] = value | |||
else: | |||
prefix, name = splited | |||
if prefix not in new_state: | |||
new_state[prefix] = {} | |||
new_state[prefix][name] = value | |||
for key, param in self.attributes(TorchModule, TorchParameter, PaddleLayer): | |||
if key in new_state: | |||
# 在传入的字典中找到了对应的值 | |||
input_param = new_state[key] | |||
if not isinstance(input_param, dict): | |||
# 且不是字典,即上述没有前缀的情况 | |||
# 按照torch.nn.Module._load_from_state_dict进行赋值 | |||
if not torch.overrides.is_tensor_like(input_param): | |||
error_msgs.append('While copying the parameter named "{}", ' | |||
'expected torch.Tensor or Tensor-like object from checkpoint but ' | |||
'received {}' | |||
.format(key, type(input_param))) | |||
continue | |||
# This is used to avoid copying uninitialized parameters into | |||
# non-lazy modules, since they dont have the hook to do the checks | |||
# in such case, it will error when accessing the .shape attribute. | |||
is_param_lazy = torch.nn.parameter.is_lazy(param) | |||
# Backward compatibility: loading 1-dim tensor from 0.3.* to version 0.4+ | |||
if not is_param_lazy and len(param.shape) == 0 and len(input_param.shape) == 1: | |||
input_param = input_param[0] | |||
if not is_param_lazy and input_param.shape != param.shape: | |||
# local shape should match the one in checkpoint | |||
error_msgs.append('size mismatch for {}: copying a param with shape {} from checkpoint, ' | |||
'the shape in current model is {}.' | |||
.format(key, input_param.shape, param.shape)) | |||
continue | |||
try: | |||
with torch.no_grad(): | |||
param.copy_(input_param) | |||
except Exception as ex: | |||
error_msgs.append('While copying the parameter named "{}", ' | |||
'whose dimensions in the model are {} and ' | |||
'whose dimensions in the checkpoint are {}, ' | |||
'an exception occurred : {}.' | |||
.format(key, param.size(), input_param.size(), ex.args)) | |||
else: | |||
# 否则在子模块中 | |||
if isinstance(param, TorchModule): | |||
# torch模块 | |||
# 由于paddle没有提供类似strict的参数,因此也不对torch作要求 | |||
param.load_state_dict(input_param, strict=False) | |||
elif isinstance(param, PaddleLayer): | |||
# paddle模块 | |||
param.load_dict(input_param) | |||
else: | |||
missing_keys.append(key) | |||
if len(error_msgs) > 0: | |||
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( | |||
self.__class__.__name__, "\n\t".join(error_msgs))) | |||
def attributes(self, *types): | |||
""" | |||
查找对应类型的成员 | |||
""" | |||
for name, value in vars(self).items(): | |||
if isinstance(value, types): | |||
yield name, value |
@@ -1,233 +0,0 @@ | |||
import warnings | |||
import os | |||
from typing import Any, Optional, Union | |||
import numpy as np | |||
from fastNLP.core.utils.utils import apply_to_collection | |||
from fastNLP.core.utils.paddle_utils import paddle_to | |||
from fastNLP.envs.imports import _NEED_IMPORT_JITTOR, _NEED_IMPORT_TORCH, _NEED_IMPORT_PADDLE | |||
if _NEED_IMPORT_PADDLE: | |||
import paddle | |||
if _NEED_IMPORT_JITTOR: | |||
import jittor | |||
if _NEED_IMPORT_TORCH: | |||
import torch | |||
__all__ = [ | |||
"paddle2torch", | |||
"torch2paddle", | |||
"jittor2torch", | |||
"torch2jittor", | |||
] | |||
def _paddle2torch(paddle_tensor: 'paddle.Tensor', target_device: Optional[Union[str, int]] = None, no_gradient: bool = None) -> 'torch.Tensor': | |||
""" | |||
将paddle tensor转换为torch tensor,并且能够保留梯度进行反向传播 | |||
:param paddle_tensor: 要转换的paddle张量 | |||
:param target_device: 是否将转换后的张量迁移到特定设备上,输入为`None`时,和输入的张量相同。 | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 转换后的torch张量 | |||
""" | |||
no_gradient = paddle_tensor.stop_gradient if no_gradient is None else no_gradient | |||
paddle_numpy = paddle_tensor.numpy() | |||
if not np.issubdtype(paddle_numpy.dtype, np.inexact): | |||
no_gradient = True | |||
if target_device is None: | |||
if paddle_tensor.place.is_gpu_place(): | |||
# paddlepaddle有两种Place,对应不同的device id获取方式 | |||
if hasattr(paddle_tensor.place, "gpu_device_id"): | |||
# paddle.fluid.core_avx.Place | |||
# 在gpu环境下创建张量的话,张量的place是这一类型 | |||
target_device = f"cuda:{paddle_tensor.place.gpu_device_id()}" | |||
else: | |||
# paddle.CUDAPlace | |||
target_device = f"cuda:{paddle_tensor.place.get_device_id()}" | |||
else: | |||
# TODO: 可能需要支持xpu等设备 | |||
target_device = "cpu" | |||
if not no_gradient: | |||
# 保持梯度,并保持反向传播 | |||
# torch.tensor会保留numpy数组的类型 | |||
torch_tensor = torch.tensor(paddle_numpy, requires_grad=True, device=target_device) | |||
hook = torch_tensor.register_hook( | |||
lambda grad: paddle.autograd.backward(paddle_tensor, paddle.to_tensor(grad.cpu().numpy())) | |||
) | |||
else: | |||
# 不保留梯度 | |||
torch_tensor = torch.tensor(paddle_numpy, requires_grad=False, device=target_device) | |||
return torch_tensor | |||
def _torch2paddle(torch_tensor: 'torch.Tensor', target_device: str = None, no_gradient: bool = None) -> 'paddle.Tensor': | |||
""" | |||
将torch tensor转换为paddle tensor,并且能够保留梯度进行反向传播。 | |||
:param torch_tensor: 要转换的torch张量 | |||
:param target_device: 是否将转换后的张量迁移到特定设备上,输入为`None`时,和输入的张量相同。 | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 转换后的paddle张量 | |||
""" | |||
no_gradient = not torch_tensor.requires_grad if no_gradient is None else no_gradient | |||
if target_device is None: | |||
if torch_tensor.is_cuda: | |||
target_device = f"gpu:{torch_tensor.device.index}" | |||
else: | |||
target_device = "cpu" | |||
if not no_gradient: | |||
# 保持梯度并保持反向传播 | |||
# paddle的stop_gradient和torch的requires_grad表现是相反的 | |||
paddle_tensor = paddle.to_tensor(torch_tensor.detach().numpy(), stop_gradient=False) | |||
hook = paddle_tensor.register_hook( | |||
lambda grad: torch.autograd.backward(torch_tensor, torch.tensor(grad.numpy())) | |||
) | |||
else: | |||
paddle_tensor = paddle.to_tensor(torch_tensor.detach().numpy(), stop_gradient=True) | |||
paddle_tensor = paddle_to(paddle_tensor, target_device) | |||
return paddle_tensor | |||
def _jittor2torch(jittor_var: 'jittor.Var', target_device: Optional[Union[str, int]] = None, no_gradient: bool = None) -> 'torch.Tensor': | |||
""" | |||
将jittor Var转换为torch tensor,并且能够保留梯度进行反向传播 | |||
:param jittor_var: 要转换的jittor变量 | |||
:param target_device: 是否将转换后的张量迁移到特定设备上,输入为`None`时,根据jittor.flags.use_cuda决定。 | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 转换后的torch张量 | |||
""" | |||
# TODO: warning:无法保留梯度 | |||
# jittor的grad可以通过callback进行传递 | |||
# 如果outputs有_grad键,可以实现求导 | |||
no_gradient = not jittor_var.requires_grad if no_gradient is None else no_gradient | |||
if no_gradient == False: | |||
warnings.warn("The result tensor will not keep gradients due to differences between jittor and pytorch.") | |||
jittor_numpy = jittor_var.numpy() | |||
if not np.issubdtype(jittor_numpy.dtype, np.inexact): | |||
no_gradient = True | |||
if target_device is None: | |||
# jittor的设备分配是自动的 | |||
# 根据use_cuda判断 | |||
if jittor.flags.use_cuda: | |||
target_device = "cuda:0" | |||
else: | |||
target_device = "cpu" | |||
torch_tensor = torch.tensor(jittor_numpy, requires_grad=not no_gradient, device=target_device) | |||
return torch_tensor | |||
def _torch2jittor(torch_tensor: 'torch.Tensor', no_gradient: bool = None) -> 'jittor.Var': | |||
""" | |||
将torch tensor转换为jittor Var,并且能够保留梯度进行反向传播 | |||
:param torch_tensor: 要转换的torch张量 | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 转换后的jittor变量 | |||
""" | |||
no_gradient = not torch_tensor.requires_grad if no_gradient is None else no_gradient | |||
if not no_gradient: | |||
# 保持梯度并保持反向传播 | |||
jittor_var = jittor.Var(torch_tensor.detach().numpy()) | |||
jittor_var.requires_grad = True | |||
hook = jittor_var.register_hook( | |||
lambda grad: torch.autograd.backward(torch_tensor, torch.tensor(grad.numpy())) | |||
) | |||
else: | |||
jittor_var = jittor.Var(torch_tensor.detach().numpy()) | |||
jittor_var.requires_grad = False | |||
return jittor_var | |||
def torch2paddle(torch_in: Any, target_device: str = None, no_gradient: bool = None) -> Any: | |||
""" | |||
递归地将输入中包含的torch张量转换为paddle张量 | |||
:param torch_in: 要转换的包含torch.Tensor类型的变量 | |||
:param target_device: 是否将转换后的张量迁移到特定设备上, | |||
输入为`None`时,和输入的张量相同, | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 将所有torch.Tensor转换为paddle.Tensor的张量 | |||
""" | |||
return apply_to_collection( | |||
torch_in, | |||
dtype=torch.Tensor, | |||
function=_torch2paddle, | |||
target_device=target_device, | |||
no_gradient=no_gradient, | |||
) | |||
def paddle2torch(paddle_in: Any, target_device: str = None, no_gradient: bool = None) -> Any: | |||
""" | |||
递归地将输入中包含的paddle张量转换为torch张量 | |||
:param torch_in: 要转换的包含paddle.Tensor类型的变量 | |||
:param target_device: 是否将转换后的张量迁移到特定设备上, | |||
输入为`None`时,和输入的张量相同, | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 将所有paddle.Tensor转换为torch.Tensor后的变量 | |||
""" | |||
return apply_to_collection( | |||
paddle_in, | |||
dtype=paddle.Tensor, | |||
function=_paddle2torch, | |||
target_device=target_device, | |||
no_gradient=no_gradient, | |||
) | |||
def jittor2torch(jittor_in: Any, target_device: str = None, no_gradient: bool = None) -> Any: | |||
""" | |||
递归地将输入中包含的jittor变量转换为torch张量 | |||
:param jittor_in: 要转换的jittor变量 | |||
:param target_device: 是否将转换后的张量迁移到特定设备上,输入为`None`时,默认为cuda:0。 | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 转换后的torch张量 | |||
""" | |||
return apply_to_collection( | |||
jittor_in, | |||
dtype=jittor.Var, | |||
function=_jittor2torch, | |||
target_device=target_device, | |||
no_gradient=no_gradient, | |||
) | |||
def torch2jittor(torch_in: Any, no_gradient: bool = None) -> Any: | |||
""" | |||
递归地将输入中包含的torch张量转换为jittor变量 | |||
:param torch_tensor: 要转换的torch张量 | |||
:param no_gradient: 是否保留原张量的梯度。为`None`时,新的张量与输入张量保持一致; | |||
为`True`时,全部不保留梯度;为`False`时,全部保留梯度。 | |||
:return: 转换后的jittor变量 | |||
""" | |||
return apply_to_collection( | |||
torch_in, | |||
dtype=torch.Tensor, | |||
function=_torch2jittor, | |||
no_gradient=no_gradient, | |||
) |
@@ -334,9 +334,9 @@ def test_torch_dl(): | |||
dl = TorchDataLoader(ds, batch_size=2) | |||
batch = next(iter(dl)) | |||
assert 'x' in batch and 'y' in batch and 'z' in batch and 'i' in batch and 'j' in batch | |||
assert isinstance(batch['z'], torch.Tensor) | |||
assert isinstance(batch['z'], torch.FloatTensor) | |||
assert isinstance(batch['j'], list) | |||
assert isinstance(batch['i']['j'], torch.Tensor) | |||
assert isinstance(batch['i']['j'], torch.LongTensor) | |||
dl.set_ignore('x') | |||
batch = next(iter(dl)) | |||
@@ -11,6 +11,9 @@ if _NEED_IMPORT_JITTOR: | |||
import jittor as jt | |||
from jittor import nn, Module | |||
from jittor.dataset import Dataset | |||
else: | |||
from fastNLP.core.utils.dummy_class import DummyClass as Module | |||
from fastNLP.core.utils.dummy_class import DummyClass as Dataset | |||
class JittorNormalModel_Classification(Module): | |||
@@ -68,6 +71,7 @@ class TrainJittorConfig: | |||
@pytest.mark.parametrize("driver,device", [("jittor", None)]) | |||
@pytest.mark.parametrize("callbacks", [[RichCallback(100)]]) | |||
@pytest.mark.jittor | |||
def test_trainer_jittor( | |||
driver, | |||
device, | |||
@@ -1,122 +0,0 @@ | |||
import pytest | |||
from fastNLP.modules.mix_modules.mix_module import MixModule | |||
from fastNLP.core.drivers.torch_paddle_driver.torch_paddle_driver import TorchPaddleDriver | |||
from fastNLP.modules.mix_modules.utils import paddle2torch, torch2paddle | |||
import torch | |||
import paddle | |||
from paddle.io import Dataset, DataLoader | |||
import numpy as np | |||
############################################################################ | |||
# | |||
# 测试在MNIST数据集上的表现 | |||
# | |||
############################################################################ | |||
class MNISTDataset(Dataset): | |||
def __init__(self, dataset): | |||
self.dataset = [ | |||
( | |||
np.array(img).astype('float32').reshape(-1), | |||
label | |||
) for img, label in dataset | |||
] | |||
def __getitem__(self, idx): | |||
return self.dataset[idx] | |||
def __len__(self): | |||
return len(self.dataset) | |||
class MixMNISTModel(MixModule): | |||
def __init__(self): | |||
super(MixMNISTModel, self).__init__() | |||
self.fc1 = paddle.nn.Linear(784, 64) | |||
self.fc2 = paddle.nn.Linear(64, 32) | |||
self.fc3 = torch.nn.Linear(32, 10) | |||
self.fc4 = torch.nn.Linear(10, 10) | |||
def forward(self, x): | |||
paddle_out = self.fc1(x) | |||
paddle_out = self.fc2(paddle_out) | |||
torch_in = paddle2torch(paddle_out) | |||
torch_out = self.fc3(torch_in) | |||
torch_out = self.fc4(torch_out) | |||
return torch_out | |||
def train_step(self, x): | |||
return self.forward(x) | |||
def test_step(self, x): | |||
return self.forward(x) | |||
@pytest.mark.torchpaddle | |||
class TestMNIST: | |||
@classmethod | |||
def setup_class(self): | |||
self.train_dataset = paddle.vision.datasets.MNIST(mode='train') | |||
self.test_dataset = paddle.vision.datasets.MNIST(mode='test') | |||
self.train_dataset = MNISTDataset(self.train_dataset) | |||
self.lr = 0.0003 | |||
self.epochs = 20 | |||
self.dataloader = DataLoader(self.train_dataset, batch_size=100, shuffle=True) | |||
def setup_method(self): | |||
model = MixMNISTModel() | |||
self.torch_loss_func = torch.nn.CrossEntropyLoss() | |||
torch_opt = torch.optim.Adam(model.parameters(backend="torch"), self.lr) | |||
paddle_opt = paddle.optimizer.Adam(parameters=model.parameters(backend="paddle"), learning_rate=self.lr) | |||
self.driver = TorchPaddleDriver(model=model, device="cuda:0") | |||
self.driver.set_optimizers([torch_opt, paddle_opt]) | |||
def test_case1(self): | |||
epochs = 20 | |||
self.driver.setup() | |||
self.driver.zero_grad() | |||
# 开始训练 | |||
current_epoch_idx = 0 | |||
while current_epoch_idx < epochs: | |||
epoch_loss, batch = 0, 0 | |||
self.driver.set_model_mode("train") | |||
self.driver.set_sampler_epoch(self.dataloader, current_epoch_idx) | |||
for batch, (img, label) in enumerate(self.dataloader): | |||
img = paddle.to_tensor(img).cuda() | |||
torch_out = self.driver.train_step(img) | |||
label = torch.from_numpy(label.numpy()).reshape(-1) | |||
loss = self.torch_loss_func(torch_out.cpu(), label) | |||
epoch_loss += loss.item() | |||
self.driver.backward(loss) | |||
self.driver.step() | |||
self.driver.zero_grad() | |||
current_epoch_idx += 1 | |||
# 开始测试 | |||
correct = 0 | |||
for img, label in self.test_dataset: | |||
img = paddle.to_tensor(np.array(img).astype('float32').reshape(1, -1)) | |||
torch_out = self.driver.test_step(img) | |||
res = torch_out.softmax(-1).argmax().item() | |||
label = label.item() | |||
if res == label: | |||
correct += 1 | |||
acc = correct / len(self.test_dataset) | |||
assert acc > 0.85 |
@@ -1,204 +0,0 @@ | |||
import paddle | |||
import pytest | |||
import torch | |||
from fastNLP.core.utils.torch_paddle_utils import torch_paddle_move_data_to_device | |||
############################################################################ | |||
# | |||
# 测试将参数中包含的所有torch和paddle张量迁移到指定设备 | |||
# | |||
############################################################################ | |||
@pytest.mark.torchpaddle | |||
class TestTorchPaddleMoveDataToDevice: | |||
def check_gpu(self, tensor, idx): | |||
""" | |||
检查张量是否在指定显卡上的工具函数 | |||
""" | |||
if isinstance(tensor, paddle.Tensor): | |||
assert tensor.place.is_gpu_place() | |||
assert tensor.place.gpu_device_id() == idx | |||
elif isinstance(tensor, torch.Tensor): | |||
assert tensor.is_cuda | |||
assert tensor.device.index == idx | |||
def check_cpu(self, tensor): | |||
if isinstance(tensor, paddle.Tensor): | |||
assert tensor.place.is_cpu_place() | |||
elif isinstance(tensor, torch.Tensor): | |||
assert not tensor.is_cuda | |||
def test_tensor_transfer(self): | |||
""" | |||
测试迁移单个张量 | |||
""" | |||
paddle_tensor = paddle.rand((3, 4, 5)).cpu() | |||
res = torch_paddle_move_data_to_device(paddle_tensor, device=None, data_device=None) | |||
self.check_cpu(res) | |||
res = torch_paddle_move_data_to_device(paddle_tensor, device="gpu:0", data_device=None) | |||
self.check_gpu(res, 0) | |||
res = torch_paddle_move_data_to_device(paddle_tensor, device="gpu:1", data_device=None) | |||
self.check_gpu(res, 1) | |||
res = torch_paddle_move_data_to_device(paddle_tensor, device="cuda:0", data_device="cpu") | |||
self.check_gpu(res, 0) | |||
res = torch_paddle_move_data_to_device(paddle_tensor, device=None, data_device="gpu:0") | |||
self.check_gpu(res, 0) | |||
res = torch_paddle_move_data_to_device(paddle_tensor, device=None, data_device="cuda:1") | |||
self.check_gpu(res, 1) | |||
torch_tensor = torch.rand(3, 4, 5) | |||
res = torch_paddle_move_data_to_device(torch_tensor, device=None, data_device=None) | |||
self.check_cpu(res) | |||
res = torch_paddle_move_data_to_device(torch_tensor, device="gpu:0", data_device=None) | |||
self.check_gpu(res, 0) | |||
res = torch_paddle_move_data_to_device(torch_tensor, device="gpu:1", data_device=None) | |||
self.check_gpu(res, 1) | |||
res = torch_paddle_move_data_to_device(torch_tensor, device="gpu:0", data_device="cpu") | |||
self.check_gpu(res, 0) | |||
res = torch_paddle_move_data_to_device(torch_tensor, device=None, data_device="gpu:0") | |||
self.check_gpu(res, 0) | |||
res = torch_paddle_move_data_to_device(torch_tensor, device=None, data_device="gpu:1") | |||
self.check_gpu(res, 1) | |||
def test_list_transfer(self): | |||
""" | |||
测试迁移张量的列表 | |||
""" | |||
paddle_list = [paddle.rand((6, 4, 2)) for i in range(5)] + [torch.rand((6, 4, 2)) for i in range(5)] | |||
res = torch_paddle_move_data_to_device(paddle_list, device=None, data_device="gpu:1") | |||
assert isinstance(res, list) | |||
for r in res: | |||
self.check_gpu(r, 1) | |||
res = torch_paddle_move_data_to_device(paddle_list, device="cpu", data_device="gpu:1") | |||
assert isinstance(res, list) | |||
for r in res: | |||
self.check_cpu(r) | |||
res = torch_paddle_move_data_to_device(paddle_list, device="gpu:0", data_device=None) | |||
assert isinstance(res, list) | |||
for r in res: | |||
self.check_gpu(r, 0) | |||
res = torch_paddle_move_data_to_device(paddle_list, device="gpu:1", data_device="cpu") | |||
assert isinstance(res, list) | |||
for r in res: | |||
self.check_gpu(r, 1) | |||
def test_tensor_tuple_transfer(self): | |||
""" | |||
测试迁移张量的元组 | |||
""" | |||
paddle_list = [paddle.rand((6, 4, 2)) for i in range(10)] + [torch.rand((6, 4, 2)) for i in range(5)] | |||
paddle_tuple = tuple(paddle_list) | |||
res = torch_paddle_move_data_to_device(paddle_tuple, device=None, data_device="gpu:1") | |||
assert isinstance(res, tuple) | |||
for r in res: | |||
self.check_gpu(r, 1) | |||
res = torch_paddle_move_data_to_device(paddle_tuple, device="cpu", data_device="gpu:1") | |||
assert isinstance(res, tuple) | |||
for r in res: | |||
self.check_cpu(r) | |||
res = torch_paddle_move_data_to_device(paddle_tuple, device="gpu:0", data_device=None) | |||
assert isinstance(res, tuple) | |||
for r in res: | |||
self.check_gpu(r, 0) | |||
res = torch_paddle_move_data_to_device(paddle_tuple, device="gpu:1", data_device="cpu") | |||
assert isinstance(res, tuple) | |||
for r in res: | |||
self.check_gpu(r, 1) | |||
def test_dict_transfer(self): | |||
""" | |||
测试迁移复杂的字典结构 | |||
""" | |||
paddle_dict = { | |||
"torch_tensor": torch.rand((3, 4)), | |||
"torch_list": [torch.rand((6, 4, 2)) for i in range(10)], | |||
"dict":{ | |||
"list": [paddle.rand((6, 4, 2)) for i in range(5)] + [torch.rand((6, 4, 2)) for i in range(5)], | |||
"torch_tensor": torch.rand((3, 4)), | |||
"paddle_tensor": paddle.rand((3, 4)) | |||
}, | |||
"paddle_tensor": paddle.rand((3, 4)), | |||
"list": [paddle.rand((6, 4, 2)) for i in range(10)] , | |||
"int": 2, | |||
"string": "test string" | |||
} | |||
res = torch_paddle_move_data_to_device(paddle_dict, device="gpu:0", data_device=None) | |||
assert isinstance(res, dict) | |||
self.check_gpu(res["torch_tensor"], 0) | |||
self.check_gpu(res["paddle_tensor"], 0) | |||
assert isinstance(res["torch_list"], list) | |||
for t in res["torch_list"]: | |||
self.check_gpu(t, 0) | |||
assert isinstance(res["list"], list) | |||
for t in res["list"]: | |||
self.check_gpu(t, 0) | |||
assert isinstance(res["int"], int) | |||
assert isinstance(res["string"], str) | |||
assert isinstance(res["dict"], dict) | |||
assert isinstance(res["dict"]["list"], list) | |||
for t in res["dict"]["list"]: | |||
self.check_gpu(t, 0) | |||
self.check_gpu(res["dict"]["torch_tensor"], 0) | |||
self.check_gpu(res["dict"]["paddle_tensor"], 0) | |||
res = torch_paddle_move_data_to_device(paddle_dict, device=None, data_device="gpu:1") | |||
assert isinstance(res, dict) | |||
self.check_gpu(res["torch_tensor"], 1) | |||
self.check_gpu(res["paddle_tensor"], 1) | |||
assert isinstance(res["torch_list"], list) | |||
for t in res["torch_list"]: | |||
self.check_gpu(t, 1) | |||
assert isinstance(res["list"], list) | |||
for t in res["list"]: | |||
self.check_gpu(t, 1) | |||
assert isinstance(res["int"], int) | |||
assert isinstance(res["string"], str) | |||
assert isinstance(res["dict"], dict) | |||
assert isinstance(res["dict"]["list"], list) | |||
for t in res["dict"]["list"]: | |||
self.check_gpu(t, 1) | |||
self.check_gpu(res["dict"]["torch_tensor"], 1) | |||
self.check_gpu(res["dict"]["paddle_tensor"], 1) | |||
res = torch_paddle_move_data_to_device(paddle_dict, device="cpu", data_device="gpu:0") | |||
assert isinstance(res, dict) | |||
self.check_cpu(res["torch_tensor"]) | |||
self.check_cpu(res["paddle_tensor"]) | |||
assert isinstance(res["torch_list"], list) | |||
for t in res["torch_list"]: | |||
self.check_cpu(t) | |||
assert isinstance(res["list"], list) | |||
for t in res["list"]: | |||
self.check_cpu(t) | |||
assert isinstance(res["int"], int) | |||
assert isinstance(res["string"], str) | |||
assert isinstance(res["dict"], dict) | |||
assert isinstance(res["dict"]["list"], list) | |||
for t in res["dict"]["list"]: | |||
self.check_cpu(t) | |||
self.check_cpu(res["dict"]["torch_tensor"]) | |||
self.check_cpu(res["dict"]["paddle_tensor"]) |
@@ -1,378 +0,0 @@ | |||
import pytest | |||
import os | |||
from itertools import chain | |||
import torch | |||
import paddle | |||
from paddle.io import Dataset, DataLoader | |||
import numpy as np | |||
from fastNLP.modules.mix_modules.mix_module import MixModule | |||
from fastNLP.modules.mix_modules.utils import paddle2torch, torch2paddle | |||
from fastNLP.envs.distributed import rank_zero_rm | |||
############################################################################ | |||
# | |||
# 测试类的基本功能 | |||
# | |||
############################################################################ | |||
class MixModuleForTest(MixModule): | |||
def __init__(self): | |||
super(MixModuleForTest, self).__init__() | |||
self.torch_fc1 = torch.nn.Linear(10, 10) | |||
self.torch_softmax = torch.nn.Softmax(0) | |||
self.torch_conv2d1 = torch.nn.Conv2d(10, 10, 3) | |||
self.torch_tensor = torch.ones(3, 3) | |||
self.torch_param = torch.nn.Parameter(torch.ones(4, 4)) | |||
self.paddle_fc1 = paddle.nn.Linear(10, 10) | |||
self.paddle_softmax = paddle.nn.Softmax(0) | |||
self.paddle_conv2d1 = paddle.nn.Conv2D(10, 10, 3) | |||
self.paddle_tensor = paddle.ones((4, 4)) | |||
class TorchModuleForTest(torch.nn.Module): | |||
def __init__(self): | |||
super(TorchModuleForTest, self).__init__() | |||
self.torch_fc1 = torch.nn.Linear(10, 10) | |||
self.torch_softmax = torch.nn.Softmax(0) | |||
self.torch_conv2d1 = torch.nn.Conv2d(10, 10, 3) | |||
self.torch_tensor = torch.ones(3, 3) | |||
self.torch_param = torch.nn.Parameter(torch.ones(4, 4)) | |||
class PaddleModuleForTest(paddle.nn.Layer): | |||
def __init__(self): | |||
super(PaddleModuleForTest, self).__init__() | |||
self.paddle_fc1 = paddle.nn.Linear(10, 10) | |||
self.paddle_softmax = paddle.nn.Softmax(0) | |||
self.paddle_conv2d1 = paddle.nn.Conv2D(10, 10, 3) | |||
self.paddle_tensor = paddle.ones((4, 4)) | |||
@pytest.mark.torchpaddle | |||
class TestTorchPaddleMixModule: | |||
def setup_method(self): | |||
self.model = MixModuleForTest() | |||
self.torch_model = TorchModuleForTest() | |||
self.paddle_model = PaddleModuleForTest() | |||
def test_to(self): | |||
""" | |||
测试混合模型的to函数 | |||
""" | |||
self.model.to("cuda") | |||
self.torch_model.to("cuda") | |||
self.paddle_model.to("gpu") | |||
self.if_device_correct("cuda") | |||
self.model.to("cuda:2") | |||
self.torch_model.to("cuda:2") | |||
self.paddle_model.to("gpu:2") | |||
self.if_device_correct("cuda:2") | |||
self.model.to("gpu:1") | |||
self.torch_model.to("cuda:1") | |||
self.paddle_model.to("gpu:1") | |||
self.if_device_correct("cuda:1") | |||
self.model.to("cpu") | |||
self.torch_model.to("cpu") | |||
self.paddle_model.to("cpu") | |||
self.if_device_correct("cpu") | |||
def test_train_eval(self): | |||
""" | |||
测试train和eval函数 | |||
""" | |||
self.model.eval() | |||
self.if_training_correct(False) | |||
self.model.train() | |||
self.if_training_correct(True) | |||
def test_parameters(self): | |||
""" | |||
测试parameters()函数,由于初始化是随机的,目前仅比较得到结果的长度 | |||
""" | |||
mix_params = [] | |||
params = [] | |||
for value in self.model.named_parameters(): | |||
mix_params.append(value) | |||
for value in chain(self.torch_model.named_parameters(), self.paddle_model.named_parameters()): | |||
params.append(value) | |||
assert len(params) == len(mix_params) | |||
def test_named_parameters(self): | |||
""" | |||
测试named_parameters函数 | |||
""" | |||
mix_param_names = [] | |||
param_names = [] | |||
for name, value in self.model.named_parameters(): | |||
mix_param_names.append(name) | |||
for name, value in chain(self.torch_model.named_parameters(), self.paddle_model.named_parameters()): | |||
param_names.append(name) | |||
assert sorted(param_names) == sorted(mix_param_names) | |||
def test_torch_named_parameters(self): | |||
""" | |||
测试对torch参数的提取 | |||
""" | |||
mix_param_names = [] | |||
param_names = [] | |||
for name, value in self.model.named_parameters(backend="torch"): | |||
mix_param_names.append(name) | |||
for name, value in self.torch_model.named_parameters(): | |||
param_names.append(name) | |||
assert sorted(param_names) == sorted(mix_param_names) | |||
def test_paddle_named_parameters(self): | |||
""" | |||
测试对paddle参数的提取 | |||
""" | |||
mix_param_names = [] | |||
param_names = [] | |||
for name, value in self.model.named_parameters(backend="paddle"): | |||
mix_param_names.append(name) | |||
for name, value in self.paddle_model.named_parameters(): | |||
param_names.append(name) | |||
assert sorted(param_names) == sorted(mix_param_names) | |||
def test_torch_state_dict(self): | |||
""" | |||
测试提取torch的state dict | |||
""" | |||
torch_dict = self.torch_model.state_dict() | |||
mix_dict = self.model.state_dict(backend="torch") | |||
assert sorted(torch_dict.keys()) == sorted(mix_dict.keys()) | |||
def test_paddle_state_dict(self): | |||
""" | |||
测试提取paddle的state dict | |||
""" | |||
paddle_dict = self.paddle_model.state_dict() | |||
mix_dict = self.model.state_dict(backend="paddle") | |||
# TODO 测试程序会显示passed后显示paddle的异常退出信息 | |||
assert sorted(paddle_dict.keys()) == sorted(mix_dict.keys()) | |||
def test_state_dict(self): | |||
""" | |||
测试提取所有的state dict | |||
""" | |||
all_dict = self.torch_model.state_dict() | |||
all_dict.update(self.paddle_model.state_dict()) | |||
mix_dict = self.model.state_dict() | |||
# TODO 测试程序会显示passed后显示paddle的异常退出信息 | |||
assert sorted(all_dict.keys()) == sorted(mix_dict.keys()) | |||
def test_load_state_dict(self): | |||
""" | |||
测试load_state_dict函数 | |||
""" | |||
state_dict = self.model.state_dict() | |||
new_model = MixModuleForTest() | |||
new_model.load_state_dict(state_dict) | |||
new_state_dict = new_model.state_dict() | |||
for name, value in state_dict.items(): | |||
state_dict[name] = value.tolist() | |||
for name, value in new_state_dict.items(): | |||
new_state_dict[name] = value.tolist() | |||
# self.assertDictEqual(state_dict, new_state_dict) | |||
def test_save_and_load_state_dict(self): | |||
""" | |||
测试save_state_dict_to_file和load_state_dict_from_file函数 | |||
""" | |||
path = "model" | |||
try: | |||
self.model.save_state_dict_to_file(path) | |||
new_model = MixModuleForTest() | |||
new_model.load_state_dict_from_file(path) | |||
state_dict = self.model.state_dict() | |||
new_state_dict = new_model.state_dict() | |||
for name, value in state_dict.items(): | |||
state_dict[name] = value.tolist() | |||
for name, value in new_state_dict.items(): | |||
new_state_dict[name] = value.tolist() | |||
# self.assertDictEqual(state_dict, new_state_dict) | |||
finally: | |||
rank_zero_rm(path) | |||
def if_device_correct(self, device): | |||
assert self.model.torch_fc1.weight.device == self.torch_model.torch_fc1.weight.device | |||
assert self.model.torch_conv2d1.weight.device == self.torch_model.torch_fc1.bias.device | |||
assert self.model.torch_conv2d1.bias.device == self.torch_model.torch_conv2d1.bias.device | |||
assert self.model.torch_tensor.device == self.torch_model.torch_tensor.device | |||
assert self.model.torch_param.device == self.torch_model.torch_param.device | |||
if device == "cpu": | |||
assert self.model.paddle_fc1.weight.place.is_cpu_place() | |||
assert self.model.paddle_fc1.bias.place.is_cpu_place() | |||
assert self.model.paddle_conv2d1.weight.place.is_cpu_place() | |||
assert self.model.paddle_conv2d1.bias.place.is_cpu_place() | |||
assert self.model.paddle_tensor.place.is_cpu_place() | |||
elif device.startswith("cuda"): | |||
assert self.model.paddle_fc1.weight.place.is_gpu_place() | |||
assert self.model.paddle_fc1.bias.place.is_gpu_place() | |||
assert self.model.paddle_conv2d1.weight.place.is_gpu_place() | |||
assert self.model.paddle_conv2d1.bias.place.is_gpu_place() | |||
assert self.model.paddle_tensor.place.is_gpu_place() | |||
assert self.model.paddle_fc1.weight.place.gpu_device_id() == self.paddle_model.paddle_fc1.weight.place.gpu_device_id() | |||
assert self.model.paddle_fc1.bias.place.gpu_device_id() == self.paddle_model.paddle_fc1.bias.place.gpu_device_id() | |||
assert self.model.paddle_conv2d1.weight.place.gpu_device_id() == self.paddle_model.paddle_conv2d1.weight.place.gpu_device_id() | |||
assert self.model.paddle_conv2d1.bias.place.gpu_device_id() == self.paddle_model.paddle_conv2d1.bias.place.gpu_device_id() | |||
assert self.model.paddle_tensor.place.gpu_device_id() == self.paddle_model.paddle_tensor.place.gpu_device_id() | |||
else: | |||
raise NotImplementedError | |||
def if_training_correct(self, training): | |||
assert self.model.torch_fc1.training == training | |||
assert self.model.torch_softmax.training == training | |||
assert self.model.torch_conv2d1.training == training | |||
assert self.model.paddle_fc1.training == training | |||
assert self.model.paddle_softmax.training == training | |||
assert self.model.paddle_conv2d1.training == training | |||
############################################################################ | |||
# | |||
# 测试在MNIST数据集上的表现 | |||
# | |||
############################################################################ | |||
class MNISTDataset(Dataset): | |||
def __init__(self, dataset): | |||
self.dataset = [ | |||
( | |||
np.array(img).astype('float32').reshape(-1), | |||
label | |||
) for img, label in dataset | |||
] | |||
def __getitem__(self, idx): | |||
return self.dataset[idx] | |||
def __len__(self): | |||
return len(self.dataset) | |||
class MixMNISTModel(MixModule): | |||
def __init__(self): | |||
super(MixMNISTModel, self).__init__() | |||
self.fc1 = paddle.nn.Linear(784, 64) | |||
self.fc2 = paddle.nn.Linear(64, 32) | |||
self.fc3 = torch.nn.Linear(32, 10) | |||
self.fc4 = torch.nn.Linear(10, 10) | |||
def forward(self, x): | |||
paddle_out = self.fc1(x) | |||
paddle_out = self.fc2(paddle_out) | |||
torch_in = paddle2torch(paddle_out) | |||
torch_out = self.fc3(torch_in) | |||
torch_out = self.fc4(torch_out) | |||
return torch_out | |||
@pytest.mark.torchpaddle | |||
class TestMNIST: | |||
@classmethod | |||
def setup_class(self): | |||
self.train_dataset = paddle.vision.datasets.MNIST(mode='train') | |||
self.test_dataset = paddle.vision.datasets.MNIST(mode='test') | |||
self.train_dataset = MNISTDataset(self.train_dataset) | |||
self.lr = 0.0003 | |||
self.epochs = 20 | |||
self.dataloader = DataLoader(self.train_dataset, batch_size=100, shuffle=True) | |||
def setup_method(self): | |||
self.model = MixMNISTModel().to("cuda") | |||
self.torch_loss_func = torch.nn.CrossEntropyLoss() | |||
self.torch_opt = torch.optim.Adam(self.model.parameters(backend="torch"), self.lr) | |||
self.paddle_opt = paddle.optimizer.Adam(parameters=self.model.parameters(backend="paddle"), learning_rate=self.lr) | |||
def test_case1(self): | |||
# 开始训练 | |||
for epoch in range(self.epochs): | |||
epoch_loss, batch = 0, 0 | |||
for batch, (img, label) in enumerate(self.dataloader): | |||
img = paddle.to_tensor(img).cuda() | |||
torch_out = self.model(img) | |||
label = torch.from_numpy(label.numpy()).reshape(-1) | |||
loss = self.torch_loss_func(torch_out.cpu(), label) | |||
epoch_loss += loss.item() | |||
loss.backward() | |||
self.torch_opt.step() | |||
self.paddle_opt.step() | |||
self.torch_opt.zero_grad() | |||
self.paddle_opt.clear_grad() | |||
else: | |||
assert epoch_loss / (batch + 1) < 0.3 | |||
# 开始测试 | |||
correct = 0 | |||
for img, label in self.test_dataset: | |||
img = paddle.to_tensor(np.array(img).astype('float32').reshape(1, -1)) | |||
torch_out = self.model(img) | |||
res = torch_out.softmax(-1).argmax().item() | |||
label = label.item() | |||
if res == label: | |||
correct += 1 | |||
acc = correct / len(self.test_dataset) | |||
assert acc > 0.85 | |||
############################################################################ | |||
# | |||
# 测试在ERNIE中文数据集上的表现 | |||
# | |||
############################################################################ |
@@ -1,435 +0,0 @@ | |||
import unittest | |||
import os | |||
os.environ["log_silent"] = "1" | |||
import torch | |||
import paddle | |||
import jittor | |||
from fastNLP.modules.mix_modules.utils import ( | |||
paddle2torch, | |||
torch2paddle, | |||
jittor2torch, | |||
torch2jittor, | |||
) | |||
############################################################################ | |||
# | |||
# 测试paddle到torch的转换 | |||
# | |||
############################################################################ | |||
class Paddle2TorchTestCase(unittest.TestCase): | |||
def check_torch_tensor(self, tensor, device, requires_grad): | |||
""" | |||
检查张量设备和梯度情况的工具函数 | |||
""" | |||
assert isinstance(tensor, torch.Tensor) | |||
assert tensor.device == torch.device(device) | |||
assert tensor.requires_grad == requires_grad | |||
def test_gradient(self): | |||
""" | |||
测试张量转换后的反向传播是否正确 | |||
""" | |||
x = paddle.to_tensor([1.0, 2.0, 3.0, 4.0, 5.0], stop_gradient=False) | |||
y = paddle2torch(x) | |||
z = 3 * (y ** 2) | |||
z.sum().backward() | |||
assert y.grad.tolist() == [6, 12, 18, 24, 30] | |||
def test_tensor_transfer(self): | |||
""" | |||
测试单个张量的设备和梯度转换是否正确 | |||
""" | |||
paddle_tensor = paddle.rand((3, 4, 5)).cpu() | |||
res = paddle2torch(paddle_tensor) | |||
self.check_torch_tensor(res, "cpu", not paddle_tensor.stop_gradient) | |||
res = paddle2torch(paddle_tensor, target_device="cuda:2", no_gradient=None) | |||
self.check_torch_tensor(res, "cuda:2", not paddle_tensor.stop_gradient) | |||
res = paddle2torch(paddle_tensor, target_device="cuda:1", no_gradient=True) | |||
self.check_torch_tensor(res, "cuda:1", False) | |||
res = paddle2torch(paddle_tensor, target_device="cuda:1", no_gradient=False) | |||
self.check_torch_tensor(res, "cuda:1", True) | |||
def test_list_transfer(self): | |||
""" | |||
测试张量列表的转换 | |||
""" | |||
paddle_list = [paddle.rand((6, 4, 2)).cuda(1) for i in range(10)] | |||
res = paddle2torch(paddle_list) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_torch_tensor(t, "cuda:1", False) | |||
res = paddle2torch(paddle_list, target_device="cpu", no_gradient=False) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_torch_tensor(t, "cpu", True) | |||
def test_tensor_tuple_transfer(self): | |||
""" | |||
测试张量元组的转换 | |||
""" | |||
paddle_list = [paddle.rand((6, 4, 2)).cuda(1) for i in range(10)] | |||
paddle_tuple = tuple(paddle_list) | |||
res = paddle2torch(paddle_tuple) | |||
assert isinstance(res, tuple) | |||
for t in res: | |||
self.check_torch_tensor(t, "cuda:1", False) | |||
def test_dict_transfer(self): | |||
""" | |||
测试包含复杂结构的字典的转换 | |||
""" | |||
paddle_dict = { | |||
"tensor": paddle.rand((3, 4)).cuda(0), | |||
"list": [paddle.rand((6, 4, 2)).cuda(0) for i in range(10)], | |||
"dict":{ | |||
"list": [paddle.rand((6, 4, 2)).cuda(0) for i in range(10)], | |||
"tensor": paddle.rand((3, 4)).cuda(0) | |||
}, | |||
"int": 2, | |||
"string": "test string" | |||
} | |||
res = paddle2torch(paddle_dict) | |||
assert isinstance(res, dict) | |||
self.check_torch_tensor(res["tensor"], "cuda:0", False) | |||
assert isinstance(res["list"], list) | |||
for t in res["list"]: | |||
self.check_torch_tensor(t, "cuda:0", False) | |||
assert isinstance(res["int"], int) | |||
assert isinstance(res["string"], str) | |||
assert isinstance(res["dict"], dict) | |||
assert isinstance(res["dict"]["list"], list) | |||
for t in res["dict"]["list"]: | |||
self.check_torch_tensor(t, "cuda:0", False) | |||
self.check_torch_tensor(res["dict"]["tensor"], "cuda:0", False) | |||
############################################################################ | |||
# | |||
# 测试torch到paddle的转换 | |||
# | |||
############################################################################ | |||
class Torch2PaddleTestCase(unittest.TestCase): | |||
def check_paddle_tensor(self, tensor, device, stop_gradient): | |||
""" | |||
检查得到的paddle张量设备和梯度情况的工具函数 | |||
""" | |||
assert isinstance(tensor, paddle.Tensor) | |||
if device == "cpu": | |||
assert tensor.place.is_cpu_place() | |||
elif device.startswith("gpu"): | |||
paddle_device = paddle.device._convert_to_place(device) | |||
assert tensor.place.is_gpu_place() | |||
if hasattr(tensor.place, "gpu_device_id"): | |||
# paddle中,有两种Place | |||
# paddle.fluid.core.Place是创建Tensor时使用的类型 | |||
# 有函数gpu_device_id获取设备 | |||
assert tensor.place.gpu_device_id() == paddle_device.get_device_id() | |||
else: | |||
# 通过_convert_to_place得到的是paddle.CUDAPlace | |||
# 通过get_device_id获取设备 | |||
assert tensor.place.get_device_id() == paddle_device.get_device_id() | |||
else: | |||
raise NotImplementedError | |||
assert tensor.stop_gradient == stop_gradient | |||
def test_gradient(self): | |||
""" | |||
测试转换后梯度的反向传播 | |||
""" | |||
x = torch.tensor([1.0, 2.0, 3.0, 4.0, 5.0], requires_grad=True) | |||
y = torch2paddle(x) | |||
z = 3 * (y ** 2) | |||
z.sum().backward() | |||
assert y.grad.tolist() == [6, 12, 18, 24, 30] | |||
def test_tensor_transfer(self): | |||
""" | |||
测试单个张量的转换 | |||
""" | |||
torch_tensor = torch.rand((3, 4, 5)) | |||
res = torch2paddle(torch_tensor) | |||
self.check_paddle_tensor(res, "cpu", True) | |||
res = torch2paddle(torch_tensor, target_device="gpu:2", no_gradient=None) | |||
self.check_paddle_tensor(res, "gpu:2", True) | |||
res = torch2paddle(torch_tensor, target_device="gpu:2", no_gradient=True) | |||
self.check_paddle_tensor(res, "gpu:2", True) | |||
res = torch2paddle(torch_tensor, target_device="gpu:2", no_gradient=False) | |||
self.check_paddle_tensor(res, "gpu:2", False) | |||
def test_tensor_list_transfer(self): | |||
""" | |||
测试张量列表的转换 | |||
""" | |||
torch_list = [torch.rand(6, 4, 2) for i in range(10)] | |||
res = torch2paddle(torch_list) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_paddle_tensor(t, "cpu", True) | |||
res = torch2paddle(torch_list, target_device="gpu:1", no_gradient=False) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_paddle_tensor(t, "gpu:1", False) | |||
def test_tensor_tuple_transfer(self): | |||
""" | |||
测试张量元组的转换 | |||
""" | |||
torch_list = [torch.rand(6, 4, 2) for i in range(10)] | |||
torch_tuple = tuple(torch_list) | |||
res = torch2paddle(torch_tuple, target_device="cpu") | |||
assert isinstance(res, tuple) | |||
for t in res: | |||
self.check_paddle_tensor(t, "cpu", True) | |||
def test_dict_transfer(self): | |||
""" | |||
测试复杂的字典结构的转换 | |||
""" | |||
torch_dict = { | |||
"tensor": torch.rand((3, 4)), | |||
"list": [torch.rand(6, 4, 2) for i in range(10)], | |||
"dict":{ | |||
"list": [torch.rand(6, 4, 2) for i in range(10)], | |||
"tensor": torch.rand((3, 4)) | |||
}, | |||
"int": 2, | |||
"string": "test string" | |||
} | |||
res = torch2paddle(torch_dict) | |||
assert isinstance(res, dict) | |||
self.check_paddle_tensor(res["tensor"], "cpu", True) | |||
assert isinstance(res["list"], list) | |||
for t in res["list"]: | |||
self.check_paddle_tensor(t, "cpu", True) | |||
assert isinstance(res["int"], int) | |||
assert isinstance(res["string"], str) | |||
assert isinstance(res["dict"], dict) | |||
assert isinstance(res["dict"]["list"], list) | |||
for t in res["dict"]["list"]: | |||
self.check_paddle_tensor(t, "cpu", True) | |||
self.check_paddle_tensor(res["dict"]["tensor"], "cpu", True) | |||
############################################################################ | |||
# | |||
# 测试jittor到torch的转换 | |||
# | |||
############################################################################ | |||
class Jittor2TorchTestCase(unittest.TestCase): | |||
def check_torch_tensor(self, tensor, device, requires_grad): | |||
""" | |||
检查得到的torch张量的工具函数 | |||
""" | |||
assert isinstance(tensor, torch.Tensor) | |||
if device == "cpu": | |||
assert not tensor.is_cuda | |||
else: | |||
assert tensor.device == torch.device(device) | |||
assert tensor.requires_grad == requires_grad | |||
def test_var_transfer(self): | |||
""" | |||
测试单个Jittor Var的转换 | |||
""" | |||
jittor_var = jittor.rand((3, 4, 5)) | |||
res = jittor2torch(jittor_var) | |||
self.check_torch_tensor(res, "cpu", True) | |||
res = jittor2torch(jittor_var, target_device="cuda:2", no_gradient=None) | |||
self.check_torch_tensor(res, "cuda:2", True) | |||
res = jittor2torch(jittor_var, target_device="cuda:2", no_gradient=True) | |||
self.check_torch_tensor(res, "cuda:2", False) | |||
res = jittor2torch(jittor_var, target_device="cuda:2", no_gradient=False) | |||
self.check_torch_tensor(res, "cuda:2", True) | |||
def test_var_list_transfer(self): | |||
""" | |||
测试Jittor列表的转换 | |||
""" | |||
jittor_list = [jittor.rand((6, 4, 2)) for i in range(10)] | |||
res = jittor2torch(jittor_list) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_torch_tensor(t, "cpu", True) | |||
res = jittor2torch(jittor_list, target_device="cuda:1", no_gradient=False) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_torch_tensor(t, "cuda:1", True) | |||
def test_var_tuple_transfer(self): | |||
""" | |||
测试Jittor变量元组的转换 | |||
""" | |||
jittor_list = [jittor.rand((6, 4, 2)) for i in range(10)] | |||
jittor_tuple = tuple(jittor_list) | |||
res = jittor2torch(jittor_tuple, target_device="cpu") | |||
assert isinstance(res, tuple) | |||
for t in res: | |||
self.check_torch_tensor(t, "cpu", True) | |||
def test_dict_transfer(self): | |||
""" | |||
测试字典结构的转换 | |||
""" | |||
jittor_dict = { | |||
"tensor": jittor.rand((3, 4)), | |||
"list": [jittor.rand(6, 4, 2) for i in range(10)], | |||
"dict":{ | |||
"list": [jittor.rand(6, 4, 2) for i in range(10)], | |||
"tensor": jittor.rand((3, 4)) | |||
}, | |||
"int": 2, | |||
"string": "test string" | |||
} | |||
res = jittor2torch(jittor_dict) | |||
assert isinstance(res, dict) | |||
self.check_torch_tensor(res["tensor"], "cpu", True) | |||
assert isinstance(res["list"], list) | |||
for t in res["list"]: | |||
self.check_torch_tensor(t, "cpu", True) | |||
assert isinstance(res["int"], int) | |||
assert isinstance(res["string"], str) | |||
assert isinstance(res["dict"], dict) | |||
assert isinstance(res["dict"]["list"], list) | |||
for t in res["dict"]["list"]: | |||
self.check_torch_tensor(t, "cpu", True) | |||
self.check_torch_tensor(res["dict"]["tensor"], "cpu", True) | |||
############################################################################ | |||
# | |||
# 测试torch到jittor的转换 | |||
# | |||
############################################################################ | |||
class Torch2JittorTestCase(unittest.TestCase): | |||
def check_jittor_var(self, var, requires_grad): | |||
""" | |||
检查得到的Jittor Var梯度情况的工具函数 | |||
""" | |||
assert isinstance(var, jittor.Var) | |||
assert var.requires_grad == requires_grad | |||
def test_gradient(self): | |||
""" | |||
测试反向传播的梯度 | |||
""" | |||
x = torch.tensor([1.0, 2.0, 3.0, 4.0, 5.0], requires_grad=True) | |||
y = torch2jittor(x) | |||
z = 3 * (y ** 2) | |||
grad = jittor.grad(z, y) | |||
assert grad.tolist() == [6.0, 12.0, 18.0, 24.0, 30.0] | |||
def test_tensor_transfer(self): | |||
""" | |||
测试单个张量转换为Jittor | |||
""" | |||
torch_tensor = torch.rand((3, 4, 5)) | |||
res = torch2jittor(torch_tensor) | |||
self.check_jittor_var(res, False) | |||
res = torch2jittor(torch_tensor, no_gradient=None) | |||
self.check_jittor_var(res, False) | |||
res = torch2jittor(torch_tensor, no_gradient=True) | |||
self.check_jittor_var(res, False) | |||
res = torch2jittor(torch_tensor, no_gradient=False) | |||
self.check_jittor_var(res, True) | |||
def test_tensor_list_transfer(self): | |||
""" | |||
测试张量列表的转换 | |||
""" | |||
torch_list = [torch.rand((6, 4, 2)) for i in range(10)] | |||
res = torch2jittor(torch_list) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_jittor_var(t, False) | |||
res = torch2jittor(torch_list, no_gradient=False) | |||
assert isinstance(res, list) | |||
for t in res: | |||
self.check_jittor_var(t, True) | |||
def test_tensor_tuple_transfer(self): | |||
""" | |||
测试张量元组的转换 | |||
""" | |||
torch_list = [torch.rand((6, 4, 2)) for i in range(10)] | |||
torch_tuple = tuple(torch_list) | |||
res = torch2jittor(torch_tuple) | |||
assert isinstance(res, tuple) | |||
for t in res: | |||
self.check_jittor_var(t, False) | |||
def test_dict_transfer(self): | |||
""" | |||
测试字典结构的转换 | |||
""" | |||
torch_dict = { | |||
"tensor": torch.rand((3, 4)), | |||
"list": [torch.rand(6, 4, 2) for i in range(10)], | |||
"dict":{ | |||
"list": [torch.rand(6, 4, 2) for i in range(10)], | |||
"tensor": torch.rand((3, 4)) | |||
}, | |||
"int": 2, | |||
"string": "test string" | |||
} | |||
res = torch2jittor(torch_dict) | |||
assert isinstance(res, dict) | |||
self.check_jittor_var(res["tensor"], False) | |||
assert isinstance(res["list"], list) | |||
for t in res["list"]: | |||
self.check_jittor_var(t, False) | |||
assert isinstance(res["int"], int) | |||
assert isinstance(res["string"], str) | |||
assert isinstance(res["dict"], dict) | |||
assert isinstance(res["dict"]["list"], list) | |||
for t in res["dict"]["list"]: | |||
self.check_jittor_var(t, False) | |||
self.check_jittor_var(res["dict"]["tensor"], False) |