@scrypted/libav 1.0.167 → 1.0.169

package/opus/dnn/torch/lpcnet/utils/sparsification/common.py

@@ -0,0 +1,121 @@
+"""
+/* Copyright (c) 2023 Amazon
+   Written by Jan Buethe */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+"""
+
+import torch
+
+def sparsify_matrix(matrix : torch.tensor, density : float, block_size, keep_diagonal : bool=False, return_mask : bool=False):
+    """ sparsifies matrix with specified block size
+
+        Parameters:
+        -----------
+        matrix : torch.tensor
+            matrix to sparsify
+        density : int
+            target density
+        block_size : [int, int]
+            block size dimensions
+        keep_diagonal : bool
+            If true, the diagonal will be kept. This option requires block_size[0] == block_size[1] and defaults to False
+    """
+
+    m, n   = matrix.shape
+    m1, n1 = block_size
+
+    if m % m1 or n % n1:
+        raise ValueError(f"block size {(m1, n1)} does not divide matrix size {(m, n)}")
+
+    # extract diagonal if keep_diagonal = True
+    if keep_diagonal:
+        if m != n:
+            raise ValueError("Attempting to sparsify non-square matrix with keep_diagonal=True")
+
+        to_spare = torch.diag(torch.diag(matrix))
+        matrix   = matrix - to_spare
+    else:
+        to_spare = torch.zeros_like(matrix)
+
+    # calculate energy in sub-blocks
+    x = torch.reshape(matrix, (m // m1, m1, n // n1, n1))
+    x = x ** 2
+    block_energies = torch.sum(torch.sum(x, dim=3), dim=1)
+
+    number_of_blocks = (m * n) // (m1 * n1)
+    number_of_survivors = round(number_of_blocks * density)
+
+    # masking threshold
+    if number_of_survivors == 0:
+        threshold = 0
+    else:
+        threshold = torch.sort(torch.flatten(block_energies)).values[-number_of_survivors]
+
+    # create mask
+    mask = torch.ones_like(block_energies)
+    mask[block_energies < threshold] = 0
+    mask = torch.repeat_interleave(mask, m1, dim=0)
+    mask = torch.repeat_interleave(mask, n1, dim=1)
+
+    # perform masking
+    masked_matrix = mask * matrix + to_spare
+
+    if return_mask:
+        return masked_matrix, mask
+    else:
+        return masked_matrix
+
+def calculate_gru_flops_per_step(gru, sparsification_dict=dict(), drop_input=False):
+    input_size = gru.input_size
+    hidden_size = gru.hidden_size
+    flops = 0
+
+    input_density = (
+        sparsification_dict.get('W_ir', [1])[0]
+        + sparsification_dict.get('W_in', [1])[0]
+        + sparsification_dict.get('W_iz', [1])[0]
+    ) / 3
+
+    recurrent_density = (
+        sparsification_dict.get('W_hr', [1])[0]
+        + sparsification_dict.get('W_hn', [1])[0]
+        + sparsification_dict.get('W_hz', [1])[0]
+    ) / 3
+
+    # input matrix vector multiplications
+    if not drop_input:
+        flops += 2 * 3 * input_size * hidden_size * input_density
+
+    # recurrent matrix vector multiplications
+    flops += 2 * 3 * hidden_size * hidden_size * recurrent_density
+
+    # biases
+    flops += 6 * hidden_size
+
+    # activations estimated by 10 flops per activation
+    flops += 30 * hidden_size
+
+    return flops

package/opus/dnn/torch/lpcnet/utils/sparsification/gru_sparsifier.py

@@ -0,0 +1,187 @@
+"""
+/* Copyright (c) 2023 Amazon
+   Written by Jan Buethe */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+"""
+
+import torch
+
+from .common import sparsify_matrix
+
+
+class GRUSparsifier:
+    def __init__(self, task_list, start, stop, interval, exponent=3):
+        """ Sparsifier for torch.nn.GRUs
+
+            Parameters:
+            -----------
+            task_list : list
+                task_list contains a list of tuples (gru, sparsify_dict), where gru is an instance
+                of torch.nn.GRU and sparsify_dic is a dictionary with keys in {'W_ir', 'W_iz', 'W_in',
+                'W_hr', 'W_hz', 'W_hn'} corresponding to the input and recurrent weights for the reset,
+                update, and new gate. The values of sparsify_dict are tuples (density, [m, n], keep_diagonal),
+                where density is the target density in [0, 1], [m, n] is the shape sub-blocks to which
+                sparsification is applied and keep_diagonal is a bool variable indicating whether the diagonal
+                should be kept.
+
+            start : int
+                training step after which sparsification will be started.
+
+            stop : int
+                training step after which sparsification will be completed.
+
+            interval : int
+                sparsification interval for steps between start and stop. After stop sparsification will be
+                carried out after every call to GRUSparsifier.step()
+
+            exponent : float
+                Interpolation exponent for sparsification interval. In step i sparsification will be carried out
+                with density (alpha + target_density * (1 * alpha)), where
+                alpha = ((stop - i) / (start - stop)) ** exponent
+
+            Example:
+            --------
+            >>> import torch
+            >>> gru = torch.nn.GRU(10, 20)
+            >>> sparsify_dict = {
+            ...         'W_ir' : (0.5, [2, 2], False),
+            ...         'W_iz' : (0.6, [2, 2], False),
+            ...         'W_in' : (0.7, [2, 2], False),
+            ...         'W_hr' : (0.1, [4, 4], True),
+            ...         'W_hz' : (0.2, [4, 4], True),
+            ...         'W_hn' : (0.3, [4, 4], True),
+            ...     }
+            >>> sparsifier = GRUSparsifier([(gru, sparsify_dict)], 0, 100, 50)
+            >>> for i in range(100):
+            ...         sparsifier.step()
+        """
+        # just copying parameters...
+        self.start      = start
+        self.stop       = stop
+        self.interval   = interval
+        self.exponent   = exponent
+        self.task_list  = task_list
+
+        # ... and setting counter to 0
+        self.step_counter = 0
+
+        self.last_masks = {key : None for key in ['W_ir', 'W_in', 'W_iz', 'W_hr', 'W_hn', 'W_hz']}
+
+    def step(self, verbose=False):
+        """ carries out sparsification step
+
+            Call this function after optimizer.step in your
+            training loop.
+
+            Parameters:
+            ----------
+            verbose : bool
+                if true, densities are printed out
+
+            Returns:
+            --------
+            None
+
+        """
+        # compute current interpolation factor
+        self.step_counter += 1
+
+        if self.step_counter < self.start:
+            return
+        elif self.step_counter < self.stop:
+            # update only every self.interval-th interval
+            if self.step_counter % self.interval:
+                return
+
+            alpha = ((self.stop - self.step_counter) / (self.stop - self.start)) ** self.exponent
+        else:
+            alpha = 0
+
+
+        with torch.no_grad():
+            for gru, params in self.task_list:
+                hidden_size = gru.hidden_size
+
+                # input weights
+                for i, key in enumerate(['W_ir', 'W_iz', 'W_in']):
+                    if key in params:
+                        density = alpha + (1 - alpha) * params[key][0]
+                        if verbose:
+                            print(f"[{self.step_counter}]: {key} density: {density}")
+
+                        gru.weight_ih_l0[i * hidden_size : (i+1) * hidden_size, : ], new_mask = sparsify_matrix(
+                            gru.weight_ih_l0[i * hidden_size : (i + 1) * hidden_size, : ],
+                            density, # density
+                            params[key][1], # block_size
+                            params[key][2], # keep_diagonal (might want to set this to False)
+                            return_mask=True
+                        )
+
+                        if type(self.last_masks[key]) != type(None):
+                            if not torch.all(self.last_masks[key] == new_mask) and self.step_counter > self.stop:
+                                print(f"sparsification mask {key} changed for gru {gru}")
+
+                        self.last_masks[key] = new_mask
+
+                # recurrent weights
+                for i, key in enumerate(['W_hr', 'W_hz', 'W_hn']):
+                    if key in params:
+                        density = alpha + (1 - alpha) * params[key][0]
+                        if verbose:
+                            print(f"[{self.step_counter}]: {key} density: {density}")
+                        gru.weight_hh_l0[i * hidden_size : (i+1) * hidden_size, : ], new_mask = sparsify_matrix(
+                            gru.weight_hh_l0[i * hidden_size : (i + 1) * hidden_size, : ],
+                            density,
+                            params[key][1], # block_size
+                            params[key][2], # keep_diagonal (might want to set this to False)
+                            return_mask=True
+                        )
+
+                        if type(self.last_masks[key]) != type(None):
+                            if not torch.all(self.last_masks[key] == new_mask) and self.step_counter > self.stop:
+                                print(f"sparsification mask {key} changed for gru {gru}")
+
+                        self.last_masks[key] = new_mask
+
+
+
+if __name__ == "__main__":
+    print("Testing sparsifier")
+
+    gru = torch.nn.GRU(10, 20)
+    sparsify_dict = {
+        'W_ir' : (0.5, [2, 2], False),
+        'W_iz' : (0.6, [2, 2], False),
+        'W_in' : (0.7, [2, 2], False),
+        'W_hr' : (0.1, [4, 4], True),
+        'W_hz' : (0.2, [4, 4], True),
+        'W_hn' : (0.3, [4, 4], True),
+    }
+
+    sparsifier = GRUSparsifier([(gru, sparsify_dict)], 0, 100, 10)
+
+    for i in range(100):
+        sparsifier.step(verbose=True)

package/opus/dnn/torch/lpcnet/utils/templates.py

@@ -0,0 +1,157 @@
+"""
+/* Copyright (c) 2023 Amazon
+   Written by Jan Buethe */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+"""
+
+from models import multi_rate_lpcnet
+import copy
+
+setup_dict = dict()
+
+dataset_template_v2 = {
+    'version'               : 2,
+    'feature_file'          : 'features.f32',
+    'signal_file'           : 'data.s16',
+    'frame_length'          : 160,
+    'feature_frame_length'  : 36,
+    'signal_frame_length'   : 2,
+    'feature_dtype'         : 'float32',
+    'signal_dtype'          : 'int16',
+    'feature_frame_layout'  : {'cepstrum': [0,18], 'periods': [18, 19], 'pitch_corr': [19, 20], 'lpc': [20, 36]},
+    'signal_frame_layout'   : {'last_signal' : 0, 'signal': 1} # signal, last_signal, error, prediction
+}
+
+dataset_template_v1 = {
+    'version'               : 1,
+    'feature_file'          : 'features.f32',
+    'signal_file'           : 'data.u8',
+    'frame_length'          : 160,
+    'feature_frame_length'  : 55,
+    'signal_frame_length'   : 4,
+    'feature_dtype'         : 'float32',
+    'signal_dtype'          : 'uint8',
+    'feature_frame_layout'  : {'cepstrum': [0,18], 'periods': [36, 37], 'pitch_corr': [37, 38], 'lpc': [39, 55]},
+    'signal_frame_layout'   : {'last_signal' : 0, 'prediction' : 1, 'last_error': 2, 'error': 3} # signal, last_signal, error, prediction
+}
+
+# lpcnet
+
+lpcnet_config = {
+    'frame_size'                : 160,
+    'gru_a_units'               : 384,
+    'gru_b_units'               : 64,
+    'feature_conditioning_dim'  : 128,
+    'feature_conv_kernel_size'  : 3,
+    'period_levels'             : 257,
+    'period_embedding_dim'      : 64,
+    'signal_embedding_dim'      : 128,
+    'signal_levels'             : 256,
+    'feature_dimension'         : 19,
+    'output_levels'             : 256,
+    'lpc_gamma'                 : 0.9,
+    'features'                  : ['cepstrum', 'periods', 'pitch_corr'],
+    'signals'                   : ['last_signal', 'prediction', 'last_error'],
+    'input_layout'              : { 'signals'  : {'last_signal' : 0, 'prediction' : 1, 'last_error' : 2},
+                                    'features' : {'cepstrum' : [0, 18], 'pitch_corr' : [18, 19]} },
+    'target'                    : 'error',
+    'feature_history'           : 2,
+    'feature_lookahead'         : 2,
+    'sparsification'            :   {
+       'gru_a' :  {
+            'start'     : 10000,
+            'stop'      : 30000,
+            'interval'  : 100,
+            'exponent'  : 3,
+            'params'   :   {
+                'W_hr' : (0.05, [4, 8], True),
+                'W_hz' : (0.05, [4, 8], True),
+                'W_hn' : (0.2,  [4, 8], True)
+                },
+        },
+       'gru_b' :  {
+            'start'     : 10000,
+            'stop'      : 30000,
+            'interval'  : 100,
+            'exponent'  : 3,
+            'params'   :   {
+                'W_ir' : (0.5, [4, 8], False),
+                'W_iz' : (0.5, [4, 8], False),
+                'W_in' : (0.5,  [4, 8], False)
+                },
+        }
+    },
+    'add_reference_phase'       : False,
+    'reference_phase_dim'       : 0
+}
+
+
+
+# multi rate
+subconditioning = {
+        'subconditioning_a' : {
+            'number_of_subsamples'  : 2,
+            'method'                : 'modulative',
+            'signals'               : ['last_signal', 'prediction', 'last_error'],
+            'pcm_embedding_size'    : 64,
+            'kwargs'                : dict()
+
+        },
+        'subconditioning_b' : {
+            'number_of_subsamples'  : 2,
+            'method'                : 'modulative',
+            'signals'               : ['last_signal', 'prediction', 'last_error'],
+            'pcm_embedding_size'    : 64,
+            'kwargs'                : dict()
+        }
+}
+
+multi_rate_lpcnet_config = lpcnet_config.copy()
+multi_rate_lpcnet_config['subconditioning'] = subconditioning
+
+training_default = {
+    'batch_size'        : 256,
+    'epochs'            : 20,
+    'lr'                : 1e-3,
+    'lr_decay_factor'   : 2.5e-5,
+    'adam_betas'        : [0.9, 0.99],
+    'frames_per_sample' : 15
+}
+
+lpcnet_setup = {
+    'dataset'       : '/local/datasets/lpcnet_training',
+    'lpcnet'        : {'config' : lpcnet_config, 'model': 'lpcnet'},
+    'training'      : training_default
+}
+
+multi_rate_lpcnet_setup = copy.deepcopy(lpcnet_setup)
+multi_rate_lpcnet_setup['lpcnet']['config'] = multi_rate_lpcnet_config
+multi_rate_lpcnet_setup['lpcnet']['model'] = 'multi_rate'
+
+setup_dict = {
+    'lpcnet'     : lpcnet_setup,
+    'multi_rate' : multi_rate_lpcnet_setup
+}

package/opus/dnn/torch/lpcnet/utils/ulaw.py

@@ -0,0 +1,58 @@
+"""
+/* Copyright (c) 2023 Amazon
+   Written by Jan Buethe */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+"""
+
+import math as m
+
+import torch
+
+
+
+def ulaw2lin(u):
+    scale_1 = 32768.0 / 255.0
+    u = u - 128
+    s = torch.sign(u)
+    u = torch.abs(u)
+    return s * scale_1 * (torch.exp(u / 128. * m.log(256)) - 1)
+
+
+def lin2ulawq(x):
+    scale = 255.0 / 32768.0
+    s = torch.sign(x)
+    x = torch.abs(x)
+    u = s * (128 * torch.log(1 + scale * x) / m.log(256))
+    u = torch.clip(128 + torch.round(u), 0, 255)
+    return u
+
+def lin2ulaw(x):
+    scale = 255.0 / 32768.0
+    s = torch.sign(x)
+    x = torch.abs(x)
+    u = s * (128 * torch.log(1 + scale * x) / torch.log(256))
+    u = torch.clip(128 + u, 0, 255)
+    return u

package/opus/dnn/torch/lpcnet/utils/wav.py

@@ -0,0 +1,43 @@
+"""
+/* Copyright (c) 2023 Amazon
+   Written by Jan Buethe */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+"""
+
+import wave
+
+def wavwrite16(filename, x, fs):
+    """ writes x as int16 to file with name filename
+
+        If x.dtype is int16 x is written as is. Otherwise,
+        it is scaled by 2**15 - 1 and converted to int16.
+    """
+    if x.dtype != 'int16':
+        x = ((2**15 - 1) * x).astype('int16')
+
+    with wave.open(filename, 'wb') as f:
+        f.setparams((1, 2, fs, len(x), 'NONE', ""))
+        f.writeframes(x.tobytes())

package/opus/dnn/torch/neural-pitch/README.md

@@ -0,0 +1,18 @@
+## Neural Pitch Estimation
+
+- Dataset Installation
+    1. Download and unzip PTDB Dataset:
+        wget https://www2.spsc.tugraz.at/databases/PTDB-TUG/SPEECH_DATA_ZIPPED.zip
+        unzip SPEECH_DATA_ZIPPED.zip
+
+    2. Inside "SPEECH DATA" above, run ptdb_process.sh to combine male/female
+
+    3. To Download and combine demand, simply run download_demand.sh
+
+- LPCNet preparation
+    1. To extract xcorr, add lpcnet_extractor.c and add relevant functions to lpcnet_enc.c, add source for headers/c files and Makefile.am, and compile to generate ./lpcnet_xcorr_extractor object
+
+- Dataset Augmentation and training (check out arguments to each of the following)
+    1. Run data_augmentation.py
+    2. Run training.py using augmented data
+    3. Run experiments.py