o p i]@sddlmZddlmZmZddlZddlmZddlmZm Z m Z ddl m Z ddl mZdd lmZmZmZdd lmZerJdd lmZed Zd dgZ  d1d2ddZ d1d3ddZ       d4d5d-d Z       d6d7d0dZdS)8) annotations) TYPE_CHECKINGLiteralN)_C_ops)in_dynamic_modein_dynamic_or_pir_mode in_pir_mode)check_variable_and_dtype) LayerHelper)fft_c2cfft_c2rfft_r2c) is_complex)Tensorrstftistftrxr frame_lengthint hop_lengthaxis _SignalAxesname str | Nonereturnc Cs|dvr td|dt|tr|dkrtd|dt|tr&|dkr.td|dtrN||j|krFtd|d |j|d t||||StrYt||||Sd }t|d gd |t |fit }|j d d}|j |d}|j |d|i|||dd|id|S)ak Slice the N-dimensional (where N >= 1) input into (overlapping) frames. Args: x (Tensor): The input data which is a N-dimensional (where N >= 1) Tensor with shape `[..., seq_length]` or `[seq_length, ...]`. frame_length (int): Length of the frame and `0 < frame_length <= x.shape[axis]`. hop_length (int): Number of steps to advance between adjacent frames and `0 < hop_length`. axis (int, optional): Specify the axis to operate on the input Tensors. Its value should be 0(the first dimension) or -1(the last dimension). If not specified, the last axis is used by default. name (str|None, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Returns: The output frames tensor with shape `[..., frame_length, num_frames]` if `axis==-1`, otherwise `[num_frames, frame_length, ...]` where `num_frames = 1 + (x.shape[axis] - frame_length) // hop_length` Examples: .. code-block:: python >>> import paddle >>> from paddle import signal >>> # 1D >>> x = paddle.arange(8) >>> y0 = signal.frame(x, frame_length=4, hop_length=2, axis=-1) >>> print(y0) Tensor(shape=[4, 3], dtype=int64, place=Place(cpu), stop_gradient=True, [[0, 2, 4], [1, 3, 5], [2, 4, 6], [3, 5, 7]]) >>> y1 = signal.frame(x, frame_length=4, hop_length=2, axis=0) >>> print(y1) Tensor(shape=[3, 4], dtype=int64, place=Place(cpu), stop_gradient=True, [[0, 1, 2, 3], [2, 3, 4, 5], [4, 5, 6, 7]]) >>> # 2D >>> x0 = paddle.arange(16).reshape([2, 8]) >>> y0 = signal.frame(x0, frame_length=4, hop_length=2, axis=-1) >>> print(y0) Tensor(shape=[2, 4, 3], dtype=int64, place=Place(cpu), stop_gradient=True, [[[0 , 2 , 4 ], [1 , 3 , 5 ], [2 , 4 , 6 ], [3 , 5 , 7 ]], [[8 , 10, 12], [9 , 11, 13], [10, 12, 14], [11, 13, 15]]]) >>> x1 = paddle.arange(16).reshape([8, 2]) >>> y1 = signal.frame(x1, frame_length=4, hop_length=2, axis=0) >>> print(y1.shape) [3, 4, 2] >>> # > 2D >>> x0 = paddle.arange(32).reshape([2, 2, 8]) >>> y0 = signal.frame(x0, frame_length=4, hop_length=2, axis=-1) >>> print(y0.shape) [2, 2, 4, 3] >>> x1 = paddle.arange(32).reshape([8, 2, 2]) >>> y1 = signal.frame(x1, frame_length=4, hop_length=2, axis=0) >>> print(y1.shape) [3, 4, 2, 2] rUnexpected axis: . It should be 0 or -1.rzUnexpected frame_length: #. It should be an positive integer.Unexpected hop_length: zKAttribute frame_length should be less equal than sequence length, but got (z) > (z).framer)int32int64float16float32float64Zinput_param_namedtypeX)rrrOuttypeZinputsattrsZoutputs) ValueError isinstancerrshaperr"rr r locals input_dtype"create_variable_for_type_inference append_op) rrrrrop_typehelperr*outr:T/home/app/PaddleOCR-VL/.venv_paddleocr/lib/python3.10/site-packages/paddle/signal.pyr"*sNR      r"cCs|dvr td|dt|tr|dkrtd|dd}tr+t|||}|St|dgd |t|fit}|j dd }|j |d }|j |d |i||d d|id|S)aH Reconstructs a tensor consisted of overlap added sequences from input frames. Args: x (Tensor): The input data which is a N-dimensional (where N >= 2) Tensor with shape `[..., frame_length, num_frames]` or `[num_frames, frame_length ...]`. hop_length (int): Number of steps to advance between adjacent frames and `0 < hop_length <= frame_length`. axis (int, optional): Specify the axis to operate on the input Tensors. Its value should be 0(the first dimension) or -1(the last dimension). If not specified, the last axis is used by default. name (str|None, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Returns: The output frames tensor with shape `[..., seq_length]` if `axis==-1`, otherwise `[seq_length, ...]` where `seq_length = (n_frames - 1) * hop_length + frame_length` Examples: .. code-block:: python >>> import paddle >>> from paddle.signal import overlap_add >>> # 2D >>> x0 = paddle.arange(16).reshape([8, 2]) >>> print(x0) Tensor(shape=[8, 2], dtype=int64, place=Place(cpu), stop_gradient=True, [[0 , 1 ], [2 , 3 ], [4 , 5 ], [6 , 7 ], [8 , 9 ], [10, 11], [12, 13], [14, 15]]) >>> y0 = overlap_add(x0, hop_length=2, axis=-1) >>> print(y0) Tensor(shape=[10], dtype=int64, place=Place(cpu), stop_gradient=True, [0 , 2 , 5 , 9 , 13, 17, 21, 25, 13, 15]) >>> x1 = paddle.arange(16).reshape([2, 8]) >>> print(x1) Tensor(shape=[2, 8], dtype=int64, place=Place(cpu), stop_gradient=True, [[0 , 1 , 2 , 3 , 4 , 5 , 6 , 7 ], [8 , 9 , 10, 11, 12, 13, 14, 15]]) >>> y1 = overlap_add(x1, hop_length=2, axis=0) >>> print(y1) Tensor(shape=[10], dtype=int64, place=Place(cpu), stop_gradient=True, [0 , 1 , 10, 12, 14, 16, 18, 20, 14, 15]) >>> # > 2D >>> x0 = paddle.arange(32).reshape([2, 1, 8, 2]) >>> y0 = overlap_add(x0, hop_length=2, axis=-1) >>> print(y0.shape) [2, 1, 10] >>> x1 = paddle.arange(32).reshape([2, 8, 1, 2]) >>> y1 = overlap_add(x1, hop_length=2, axis=0) >>> print(y1.shape) [10, 1, 2] rrrrr!r overlap_addr)r#r$r%r&r'Zuint16r(r)r+)rrr,r-) r0r1rrrr<r r r3r4r5r6)rrrrr7r9r8r*r:r:r;r<s4J   r<TreflectFn_fft int | None win_lengthwindow Tensor | Nonecenterboolpad_modeLiteral['reflect', 'constant'] normalizedonesided bool | Nonec  Csdt|j} | dvsJd| | dkr|d}|dur#t|d}|dks/Jd|d|dur5|}trTd|krE|jd ksTnJd |jd d |dd|kr^|ksjnJd |d |d|durt|jdkr{t||ksJd |d|jdn tj|f|jd}||kr||d} ||| } tjj j || | gdd}|r|dvsJd|d|d} tjj j |d | | g|dd d }t |||d d}|j gdd}t||}|rdnd}|durt| }t|r|rJdt|st|dd |d|| d}n t|dd |d| d }|j gdd}| dkr0|d|S)!aT Short-time Fourier transform (STFT). The STFT computes the discrete Fourier transforms (DFT) of short overlapping windows of the input using this formula: .. math:: X_t[f] = \sum_{n = 0}^{N-1} \text{window}[n]\ x[t \times H + n]\ e^{-{2 \pi j f n}/{N}} Where: - :math:`t`: The :math:`t`-th input window. - :math:`f`: Frequency :math:`0 \leq f < \text{n_fft}` for `onesided=False`, or :math:`0 \leq f < \lfloor \text{n_fft} / 2 \rfloor + 1` for `onesided=True`. - :math:`N`: Value of `n_fft`. - :math:`H`: Value of `hop_length`. Args: x (Tensor): The input data which is a 1-dimensional or 2-dimensional Tensor with 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. hop_length (int|None, optional): Number of steps to advance between adjacent windows and `0 < hop_length`. Default: `None` (treated as equal to `n_fft//4`) win_length (int|None, optional): The size of window. Default: `None` (treated as equal to `n_fft`) window (Tensor|None, 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` ( 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 :math:`t \times hop\_length` at the center of :math:`t`-th frame. Default: `True`. pad_mode (str, optional): Choose padding pattern when `center` is `True`. See `paddle.nn.functional.pad` for all padding options. Default: `"reflect"` normalized (bool, optional): Control whether to scale the output by `1/sqrt(n_fft)`. Default: `False` onesided (bool, optional): Control whether to return half of the Fourier transform output that satisfies the conjugate symmetry condition when input is a real-valued tensor. It can not be `True` if input is a complex tensor. Default: `None` name (str|None, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Returns: The complex STFT output tensor with shape `[..., n_fft//2 + 1, num_frames]` (real-valued input and `onesided` is `True`) or `[..., n_fft, num_frames]` (`onesided` is `False`) Examples: .. code-block:: python >>> import paddle >>> from paddle.signal import stft >>> # real-valued input >>> x = paddle.randn([8, 48000], dtype=paddle.float64) >>> y1 = stft(x, n_fft=512) >>> print(y1.shape) [8, 257, 376] >>> y2 = stft(x, n_fft=512, onesided=False) >>> print(y2.shape) [8, 512, 376] >>> # complex input >>> x = paddle.randn([8, 48000], dtype=paddle.float64) + \ ... paddle.randn([8, 48000], dtype=paddle.float64)*1j >>> print(x.shape) [8, 48000] >>> print(x.dtype) paddle.complex128 >>> y1 = stft(x, n_fft=512, center=False, onesided=False) >>> print(y1.shape) [8, 512, 372] )r z9x should be a 1D or 2D real tensor, but got rank of x is r rNz"hop_length should be > 0, but got .rz"n_fft should be in (0, seq_length( )], but got "win_length should be in (0, n_fft(8expected a 1D window tensor of size equal to win_length(), but got window with shape r2r*rJconstantpadmode)rRr=z5pad_mode should be "reflect" or "constant", but got "z".ZNLC)rTrUZ data_format)rrrrrrJr permorthobackwardzBonesided should be False when input or window is a complex Tensor.T)rnrnormforwardrHrrr[rr\r]r)lenr2 unsqueezerrpaddleonesr*nn functionalrTsqueezer" transposemultiplyrrr Zsqueeze_)rr>rr@rArCrErGrHrx_rankpad_left pad_rightZ pad_lengthZx_framesr\r9r:r:r;rs W                lengthreturn_complexc Cst|dddgdt|j} | dvsJd| | dkr"|d}|d ur,t|d }|d ur2|}d|kr<|ksHnJd |d |d d|krR|ks^nJd|d |d |jd} |jd} tr|jdkstJd|r| |ddksJd|ddd| d n| |ksJd|d| d |d urt|jdkrt||ksJd|d|jd n|jtj tj fvrtj ntj }tj |f|d}||kr||d}|||}tj jj|||gdd}|jgdd}|rdnd}| r|rJd t|d d|d!d d"}n)t|rJd#|d!ur0|d d d d d |ddf}t|d d|d!d d"}t||jgdd}t||dd$}ttjt||d| dgd%jddgd|dd$}|d ur|r|d d |d|d f}||d|d }n|r|d}nd}|d d |||f}||||}tr|d&krtd'||}| dkrtj|dd(}|S))a  Inverse short-time Fourier transform (ISTFT). Reconstruct time-domain signal from the giving complex input and window tensor when nonzero overlap-add (NOLA) condition is met: .. math:: \sum_{t = -\infty}^{\infty} \text{window}^2[n - t \times H]\ \neq \ 0, \ \text{for } all \ n Where: - :math:`t`: The :math:`t`-th input window. - :math:`N`: Value of `n_fft`. - :math:`H`: Value of `hop_length`. Result of `istft` expected to be the inverse of `paddle.signal.stft`, but it is not guaranteed to reconstruct a exactly realizable time-domain signal from a STFT complex tensor which has been modified (via masking or otherwise). Therefore, `istft` gives the `[Griffin-Lim optimal estimate] `_ (optimal in a least-squares sense) for the corresponding signal. Args: x (Tensor): The input data which is a 2-dimensional or 3-dimensional **complex** Tensor with shape `[..., n_fft, num_frames]`. n_fft (int): The size of Fourier transform. hop_length (int|None, optional): Number of steps to advance between adjacent windows from time-domain signal and `0 < hop_length < win_length`. Default: `None` ( treated as equal to `n_fft//4`) win_length (int|None, optional): The size of window. Default: `None` (treated as equal to `n_fft`) window (Tensor|None, 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 real-valued tensor if `return_complex` is False. Default: `None`(treated as 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 padded. Default: `True`. normalized (bool, optional): Control whether to scale the output by :math:`1/sqrt(n_{fft})`. Default: `False` 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 and `istft` will return a real-valued tensor when it is set to `True`. Default: `True`. length (int|None, optional): Specify the length of time-domain signal. Default: `None`( treated as the whole length of signal). return_complex (bool, optional): It means that whether the time-domain signal is real-valued. If `return_complex` is set to `True`, `onesided` should be set to `False` cause the output is complex. name (str|None, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Returns: A tensor of least squares estimation of the reconstructed signal(s) with shape `[..., seq_length]` Examples: .. code-block:: python >>> import numpy as np >>> import paddle >>> from paddle.signal import stft, istft >>> paddle.seed(0) >>> # STFT >>> x = paddle.randn([8, 48000], dtype=paddle.float64) >>> y = stft(x, n_fft=512) >>> print(y.shape) [8, 257, 376] >>> # ISTFT >>> x_ = istft(y, n_fft=512) >>> print(x_.shape) [8, 48000] >>> np.allclose(x, x_) True r complex64Z complex128r)rJzrr@rArCrGrHrkrlrrhZn_framesZfft_sizeZ window_dtyperirjr\r9Zwindow_envelopstartr:r:r;rsY             "    )rN) rrrrrrrrrrrr) rrrrrrrrrr)NNNTr=FNN)rrr>rrr?r@r?rArBrCrDrErFrGrDrHrIrrrr) NNNTFTNFN)rrr>rrr?r@r?rArBrCrDrGrDrHrDrkr?rlrDrrrr) __future__rtypingrrrarZpaddle.frameworkrrrZbase.data_feederr Zbase.layer_helperr Zfftr r rZtensor.attributerrr__all__r"r<rrr:r:r:r;sP        ~ l 9