未验证 提交 c5af51ca 编写于 作者: M mjxs 提交者: GitHub

修复paddle.amp.decorate等API的文档 (#48983)

* 涉及到的api有
paddle.amp.decorate
paddle.static.npu_places
paddle.signal.istft
paddle.signal.stft
paddle.linalg.eigvalsh
paddle.randint_like

* change signal.stft

* randint_like的low增加optional

* ; test=docs_preview

* 修改了注解格式; test=docs_preview

* 修改了公式格式

* 修改了decorate的models等

* test=document_fix
Co-authored-by: NLigoml <39876205+Ligoml@users.noreply.github.com>
上级 2472d8f9
...@@ -94,9 +94,9 @@ def decorate( ...@@ -94,9 +94,9 @@ def decorate(
Commonly, it is used together with `auto_cast` to achieve Pure float16/bfloat16 in imperative mode. Commonly, it is used together with `auto_cast` to achieve Pure float16/bfloat16 in imperative mode.
Args: Args:
models(Layer|list of Layer, optional): The defined models by user, models must be either a single model or a list of models. Default is None. models(Layer|list of Layer): The defined models by user, models must be either a single model or a list of models. Default is None.
optimizers(Optimizer|list of Optimizer, optional): The defined optimizers by user, optimizers must be either a single optimizer or a list of optimizers. Default is None. optimizers(Optimizer|list of Optimizer, optional): The defined optimizers by user, optimizers must be either a single optimizer or a list of optimizers. Default is None.
level(str, optional): Auto mixed precision level. Accepted values are "O1" and "O2": O1 represent mixed precision, the decorator will do nothing; level(str, optional): Auto mixed precision level. Accepted values are 'O1' and 'O2': O1 represent mixed precision, the decorator will do nothing;
O2 represent Pure float16/bfloat16, the decorator will cast all parameters of models to float16/bfloat16, except BatchNorm and LayerNorm. Default is O1(amp) O2 represent Pure float16/bfloat16, the decorator will cast all parameters of models to float16/bfloat16, except BatchNorm and LayerNorm. Default is O1(amp)
dtype(str, optional): Whether to use 'float16' or 'bfloat16'. Default is 'float16'. dtype(str, optional): Whether to use 'float16' or 'bfloat16'. Default is 'float16'.
master_weight(bool, optinal): For level='O2', whether to use multi-precision during weight updating. If master_weight is None, in O2 level optimizer will use multi-precision. Default is None. master_weight(bool, optinal): For level='O2', whether to use multi-precision during weight updating. If master_weight is None, in O2 level optimizer will use multi-precision. Default is None.
......
...@@ -932,7 +932,8 @@ def xpu_places(device_ids=None): ...@@ -932,7 +932,8 @@ def xpu_places(device_ids=None):
def npu_places(device_ids=None): def npu_places(device_ids=None):
""" """
**Note**:
Note:
For multi-card tasks, please use `FLAGS_selected_npus` environment variable to set the visible NPU device. For multi-card tasks, please use `FLAGS_selected_npus` environment variable to set the visible NPU device.
This function creates a list of :code:`paddle.NPUPlace` objects. This function creates a list of :code:`paddle.NPUPlace` objects.
......
...@@ -4174,7 +4174,6 @@ class ModelAverage(Optimizer): ...@@ -4174,7 +4174,6 @@ class ModelAverage(Optimizer):
class ExponentialMovingAverage: class ExponentialMovingAverage:
r""" r"""
:api_attr: Static Graph
Compute the moving average of parameters with exponential decay. Compute the moving average of parameters with exponential decay.
Given a parameter :math:`\\theta`, its exponential moving average (EMA) Given a parameter :math:`\\theta`, its exponential moving average (EMA)
...@@ -4182,9 +4181,9 @@ class ExponentialMovingAverage: ...@@ -4182,9 +4181,9 @@ class ExponentialMovingAverage:
.. math:: .. math::
\\text{EMA}_0 & = 0 \text{EMA}_0 & = 0
\\text{EMA}_t & = \\text{decay} * \\text{EMA}_{t-1} + (1 - \\text{decay}) * \\theta_t \text{EMA}_t & = \text{decay} * \text{EMA}_{t-1} + (1 - \text{decay}) * \theta_t
The average results calculated by **update()** method will be saved in The average results calculated by **update()** method will be saved in
temporary variables which are created and maintained by the object, and can temporary variables which are created and maintained by the object, and can
...@@ -4193,12 +4192,12 @@ class ExponentialMovingAverage: ...@@ -4193,12 +4192,12 @@ class ExponentialMovingAverage:
**Bias correction**. All EMAs are initialized to :math:`0` and hence they will be **Bias correction**. All EMAs are initialized to :math:`0` and hence they will be
zero biased, which can be corrected by divided by a factor zero biased, which can be corrected by divided by a factor
:math:`(1 - \\text{decay}^t)` , i.e., the actual EMAs applied to parameters :math:`(1 - \text{decay}^t)` , i.e., the actual EMAs applied to parameters
when calling **apply()** method would be when calling **apply()** method would be
.. math:: .. math::
\\widehat{\\text{EMA}}_t = \\frac{\\text{EMA}_t}{1 - \\text{decay}^t} \widehat{\text{EMA}}_t = \frac{\text{EMA}_t}{1 - \text{decay}^t}
**Decay rate scheduling**. A large decay rate very close to 1 would result **Decay rate scheduling**. A large decay rate very close to 1 would result
in that the averages move very slowly. And a better strategy is to set a in that the averages move very slowly. And a better strategy is to set a
...@@ -4208,7 +4207,7 @@ class ExponentialMovingAverage: ...@@ -4208,7 +4207,7 @@ class ExponentialMovingAverage:
.. math:: .. math::
\\min(\\text{decay}, \\frac{1 + \\text{thres_steps}}{10 + \\text{thres_steps}}) \min(\text{decay}, \frac{1 + \text{thres_steps}}{10 + \text{thres_steps}})
Usually **thres_steps** can be the global training steps. Usually **thres_steps** can be the global training steps.
......
...@@ -28,7 +28,7 @@ def softmax_mask_fuse(x, mask, name=None): ...@@ -28,7 +28,7 @@ def softmax_mask_fuse(x, mask, name=None):
.. math:: .. math::
out = softmax(x + mask) out = softmax(x + mask)
**Note**: Note:
This API only supports GPU. This API only supports GPU.
Args: Args:
......
...@@ -29,7 +29,7 @@ def softmax_mask_fuse_upper_triangle(x): ...@@ -29,7 +29,7 @@ def softmax_mask_fuse_upper_triangle(x):
.. math:: .. math::
out = softmax(LowerTriangular(x)) out = softmax(LowerTriangular(x))
**Note**: Note:
This API only supports GPU. This API only supports GPU.
Args: Args:
......
...@@ -266,18 +266,13 @@ def stft( ...@@ -266,18 +266,13 @@ def stft(
windows of the input using this formula: windows of the input using this formula:
.. math:: .. math::
X_t[\omega] = \sum_{n = 0}^{N-1}% X_t[f] = \sum_{n = 0}^{N-1} \text{window}[n]\ x[t \times H + n]\ e^{-{2 \pi j f n}/{N}}
\text{window}[n]\ x[t \times H + n]\ %
e^{-{2 \pi j \omega n}/{N}}
Where: Where:
- :math:`t`: The :math:`t`-th input window. - :math:`t`: The :math:`t`-th input window.
- :math:`f`: Frequency :math:`0 \leq f < \text{n_fft}` for `onesided=False`,
- :math:`\omega`: Frequency :math:`0 \leq \omega < \text{n\_fft}` for `onesided=False`, or :math:`0 \leq f < \lfloor \text{n_fft} / 2 \rfloor + 1` for `onesided=True`.
or :math:`0 \leq \omega < \lfloor \text{n\_fft} / 2 \rfloor + 1` for `onesided=True`.
- :math:`N`: Value of `n_fft`. - :math:`N`: Value of `n_fft`.
- :math:`H`: Value of `hop_length`. - :math:`H`: Value of `hop_length`.
Args: Args:
...@@ -285,11 +280,11 @@ def stft( ...@@ -285,11 +280,11 @@ def stft(
shape `[..., seq_length]`. It can be a real-valued or a complex Tensor. shape `[..., seq_length]`. It can be a real-valued or a complex Tensor.
n_fft (int): The number of input samples to perform Fourier transform. n_fft (int): The number of input samples to perform Fourier transform.
hop_length (int, optional): Number of steps to advance between adjacent windows hop_length (int, optional): Number of steps to advance between adjacent windows
and `0 < hop_length`. Default: `None`(treated as equal to `n_fft//4`) and `0 < hop_length`. Default: `None` (treated as equal to `n_fft//4`)
win_length (int, optional): The size of window. Default: `None`(treated as equal win_length (int, optional): The size of window. Default: `None` (treated as equal
to `n_fft`) to `n_fft`)
window (Tensor, optional): A 1-dimensional tensor of size `win_length`. It will window (Tensor, optional): A 1-dimensional tensor of size `win_length`. It will
be center padded to length `n_fft` if `win_length < n_fft`. Default: `None`( be center padded to length `n_fft` if `win_length < n_fft`. Default: `None` (
treated as a rectangle window with value equal to 1 of size `win_length`). treated as a rectangle window with value equal to 1 of size `win_length`).
center (bool, optional): Whether to pad `x` to make that the center (bool, optional): Whether to pad `x` to make that the
:math:`t \times hop\_length` at the center of :math:`t`-th frame. Default: `True`. :math:`t \times hop\_length` at the center of :math:`t`-th frame. Default: `True`.
...@@ -441,8 +436,7 @@ def istft( ...@@ -441,8 +436,7 @@ def istft(
nonzero overlap-add (NOLA) condition is met: nonzero overlap-add (NOLA) condition is met:
.. math:: .. math::
\sum_{t = -\infty}^{\infty}% \sum_{t = -\infty}^{\infty} \text{window}^2[n - t \times H]\ \neq \ 0, \ \text{for } all \ n
\text{window}^2[n - t \times H]\ \neq \ 0, \ \text{for } all \ n
Where: Where:
- :math:`t`: The :math:`t`-th input window. - :math:`t`: The :math:`t`-th input window.
...@@ -452,7 +446,7 @@ def istft( ...@@ -452,7 +446,7 @@ def istft(
Result of `istft` expected to be the inverse of `paddle.signal.stft`, but it is Result of `istft` expected to be the inverse of `paddle.signal.stft`, but it is
not guaranteed to reconstruct a exactly realizible time-domain signal from a STFT not guaranteed to reconstruct a exactly realizible time-domain signal from a STFT
complex tensor which has been modified (via masking or otherwise). Therefore, `istft` complex tensor which has been modified (via masking or otherwise). Therefore, `istft`
gives the [Griffin-Lim optimal estimate](https://ieeexplore.ieee.org/document/1164317) gives the `[Griffin-Lim optimal estimate] <https://ieeexplore.ieee.org/document/1164317>`_
(optimal in a least-squares sense) for the corresponding signal. (optimal in a least-squares sense) for the corresponding signal.
Args: Args:
...@@ -460,9 +454,9 @@ def istft( ...@@ -460,9 +454,9 @@ def istft(
Tensor with shape `[..., n_fft, num_frames]`. Tensor with shape `[..., n_fft, num_frames]`.
n_fft (int): The size of Fourier transform. n_fft (int): The size of Fourier transform.
hop_length (int, optional): Number of steps to advance between adjacent windows hop_length (int, optional): Number of steps to advance between adjacent windows
from time-domain signal and `0 < hop_length < win_length`. Default: `None`( from time-domain signal and `0 < hop_length < win_length`. Default: `None` (
treated as equal to `n_fft//4`) treated as equal to `n_fft//4`)
win_length (int, optional): The size of window. Default: `None`(treated as equal win_length (int, optional): The size of window. Default: `None` (treated as equal
to `n_fft`) to `n_fft`)
window (Tensor, optional): A 1-dimensional tensor of size `win_length`. It will window (Tensor, optional): A 1-dimensional tensor of size `win_length`. It will
be center padded to length `n_fft` if `win_length < n_fft`. It should be a be center padded to length `n_fft` if `win_length < n_fft`. It should be a
...@@ -470,7 +464,7 @@ def istft( ...@@ -470,7 +464,7 @@ def istft(
a rectangle window with value equal to 1 of size `win_length`). a rectangle window with value equal to 1 of size `win_length`).
center (bool, optional): It means that whether the time-domain signal has been center (bool, optional): It means that whether the time-domain signal has been
center padded. Default: `True`. center padded. Default: `True`.
normalized (bool, optional): Control whether to scale the output by `1/sqrt(n_fft)`. normalized (bool, optional): Control whether to scale the output by :math:`1/sqrt(n_{fft})`.
Default: `False` Default: `False`
onesided (bool, optional): It means that whether the input STFT tensor is a half onesided (bool, optional): It means that whether the input STFT tensor is a half
of the conjugate symmetry STFT tensor transformed from a real-valued signal of the conjugate symmetry STFT tensor transformed from a real-valued signal
......
...@@ -3261,7 +3261,7 @@ def eigvalsh(x, UPLO='L', name=None): ...@@ -3261,7 +3261,7 @@ def eigvalsh(x, UPLO='L', name=None):
complex Hermitian (conjugate symmetric) or a real symmetric matrix. complex Hermitian (conjugate symmetric) or a real symmetric matrix.
Args: Args:
x (Tensor): A tensor with shape :math:`[_, M, M]` , The data type of the input Tensor x x (Tensor): A tensor with shape :math:`[*, M, M]` , where * is zero or greater batch dimension. The data type of the input Tensor x
should be one of float32, float64, complex64, complex128. should be one of float32, float64, complex64, complex128.
UPLO(str, optional): Lower triangular part of a (‘L’, default) or the upper triangular part (‘U’). UPLO(str, optional): Lower triangular part of a (‘L’, default) or the upper triangular part (‘U’).
name(str, optional): The default value is None. Normally there is no need for user to set this name(str, optional): The default value is None. Normally there is no need for user to set this
......
...@@ -878,14 +878,14 @@ def randint_like(x, low=0, high=None, dtype=None, name=None): ...@@ -878,14 +878,14 @@ def randint_like(x, low=0, high=None, dtype=None, name=None):
If ``high`` is None (the default), the range is [0, ``low``). If ``high`` is None (the default), the range is [0, ``low``).
Args: Args:
x (Tensor): The input tensor which specifies shape. The dtype of ``x`` x (Tensor): The input multi-dimensional tensor which specifies shape. The dtype of ``x``
can be bool, int32, int64, float16, float32, float64. can be bool, int32, int64, float16, float32, float64.
low (int): The lower bound on the range of random values to generate. low (int, optional): The lower bound on the range of random values to generate.
The ``low`` is included in the range. If ``high`` is None, the The ``low`` is included in the range. If ``high`` is None, the
range is [0, ``low``). Default is 0. range is [0, ``low``). Default is 0.
high (int, optional): The upper bound on the range of random values to high (int, optional): The upper bound on the range of random values to
generate, the ``high`` is excluded in the range. Default is None generate, the ``high`` is excluded in the range. Default is None.
(see above for behavior if high = None). Default is None. If ``high`` is None, the range is [0, ``low``).
dtype (str|np.dtype, optional): The data type of the dtype (str|np.dtype, optional): The data type of the
output tensor. Supported data types: bool, int32, int64, float16, output tensor. Supported data types: bool, int32, int64, float16,
float32, float64. If ``dytpe`` is None, the data type is the float32, float64. If ``dytpe`` is None, the data type is the
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册