xref: /aosp_15_r20/external/libopus/dnn/training_tf2/plc_loader.py (revision a58d3d2adb790c104798cd88c8a3aff4fa8b82cc)
1*a58d3d2aSXin Li#!/usr/bin/python3
2*a58d3d2aSXin Li'''Copyright (c) 2021-2022 Amazon
3*a58d3d2aSXin Li
4*a58d3d2aSXin Li   Redistribution and use in source and binary forms, with or without
5*a58d3d2aSXin Li   modification, are permitted provided that the following conditions
6*a58d3d2aSXin Li   are met:
7*a58d3d2aSXin Li
8*a58d3d2aSXin Li   - Redistributions of source code must retain the above copyright
9*a58d3d2aSXin Li   notice, this list of conditions and the following disclaimer.
10*a58d3d2aSXin Li
11*a58d3d2aSXin Li   - Redistributions in binary form must reproduce the above copyright
12*a58d3d2aSXin Li   notice, this list of conditions and the following disclaimer in the
13*a58d3d2aSXin Li   documentation and/or other materials provided with the distribution.
14*a58d3d2aSXin Li
15*a58d3d2aSXin Li   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
16*a58d3d2aSXin Li   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
17*a58d3d2aSXin Li   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
18*a58d3d2aSXin Li   A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
19*a58d3d2aSXin Li   CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
20*a58d3d2aSXin Li   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
21*a58d3d2aSXin Li   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
22*a58d3d2aSXin Li   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
23*a58d3d2aSXin Li   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
24*a58d3d2aSXin Li   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
25*a58d3d2aSXin Li   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26*a58d3d2aSXin Li'''
27*a58d3d2aSXin Li
28*a58d3d2aSXin Liimport numpy as np
29*a58d3d2aSXin Lifrom tensorflow.keras.utils import Sequence
30*a58d3d2aSXin Li
31*a58d3d2aSXin Liclass PLCLoader(Sequence):
32*a58d3d2aSXin Li    def __init__(self, features, lost, nb_burg_features, batch_size):
33*a58d3d2aSXin Li        self.batch_size = batch_size
34*a58d3d2aSXin Li        self.nb_batches = features.shape[0]//self.batch_size
35*a58d3d2aSXin Li        self.features = features[:self.nb_batches*self.batch_size, :, :]
36*a58d3d2aSXin Li        self.lost = lost.astype('float')
37*a58d3d2aSXin Li        self.lost = self.lost[:(len(self.lost)//features.shape[1]-1)*features.shape[1]]
38*a58d3d2aSXin Li        self.nb_burg_features = nb_burg_features
39*a58d3d2aSXin Li        self.on_epoch_end()
40*a58d3d2aSXin Li
41*a58d3d2aSXin Li    def on_epoch_end(self):
42*a58d3d2aSXin Li        self.indices = np.arange(self.nb_batches*self.batch_size)
43*a58d3d2aSXin Li        np.random.shuffle(self.indices)
44*a58d3d2aSXin Li        offset = np.random.randint(0, high=self.features.shape[1])
45*a58d3d2aSXin Li        self.lost_offset = np.reshape(self.lost[offset:-self.features.shape[1]+offset], (-1, self.features.shape[1]))
46*a58d3d2aSXin Li        self.lost_indices = np.random.randint(0, high=self.lost_offset.shape[0], size=self.nb_batches*self.batch_size)
47*a58d3d2aSXin Li
48*a58d3d2aSXin Li    def __getitem__(self, index):
49*a58d3d2aSXin Li        features = self.features[self.indices[index*self.batch_size:(index+1)*self.batch_size], :, :]
50*a58d3d2aSXin Li        burg_lost = (np.random.rand(features.shape[0], features.shape[1]) > .1).astype('float')
51*a58d3d2aSXin Li        burg_lost = np.reshape(burg_lost, (features.shape[0], features.shape[1], 1))
52*a58d3d2aSXin Li        burg_mask = np.tile(burg_lost, (1,1,self.nb_burg_features))
53*a58d3d2aSXin Li
54*a58d3d2aSXin Li        lost = self.lost_offset[self.lost_indices[index*self.batch_size:(index+1)*self.batch_size], :]
55*a58d3d2aSXin Li        lost = np.reshape(lost, (features.shape[0], features.shape[1], 1))
56*a58d3d2aSXin Li        lost_mask = np.tile(lost, (1,1,features.shape[2]))
57*a58d3d2aSXin Li        in_features = features*lost_mask
58*a58d3d2aSXin Li        in_features[:,:,:self.nb_burg_features] = in_features[:,:,:self.nb_burg_features]*burg_mask
59*a58d3d2aSXin Li
60*a58d3d2aSXin Li        #For the first frame after a loss, we don't have valid features, but the Burg estimate is valid.
61*a58d3d2aSXin Li        #in_features[:,1:,self.nb_burg_features:] = in_features[:,1:,self.nb_burg_features:]*lost_mask[:,:-1,self.nb_burg_features:]
62*a58d3d2aSXin Li        out_lost = np.copy(lost)
63*a58d3d2aSXin Li        #out_lost[:,1:,:] = out_lost[:,1:,:]*out_lost[:,:-1,:]
64*a58d3d2aSXin Li
65*a58d3d2aSXin Li        out_features = np.concatenate([features[:,:,self.nb_burg_features:], 1.-out_lost], axis=-1)
66*a58d3d2aSXin Li        burg_sign = 2*burg_lost - 1
67*a58d3d2aSXin Li        # last dim is 1 for received packet, 0 for lost packet, and -1 when just the Burg info is missing
68*a58d3d2aSXin Li        inputs = [in_features*lost_mask, lost*burg_sign]
69*a58d3d2aSXin Li        outputs = [out_features]
70*a58d3d2aSXin Li        return (inputs, outputs)
71*a58d3d2aSXin Li
72*a58d3d2aSXin Li    def __len__(self):
73*a58d3d2aSXin Li        return self.nb_batches
74