xref: /aosp_15_r20/external/libopus/dnn/torch/plc/plc_dataset.py (revision a58d3d2adb790c104798cd88c8a3aff4fa8b82cc)
1import torch
2import numpy as np
3
4class PLCDataset(torch.utils.data.Dataset):
5    def __init__(self,
6                feature_file,
7                loss_file,
8                sequence_length=1000,
9                nb_features=20,
10                nb_burg_features=36,
11                lpc_order=16):
12
13        self.features_in = nb_features + nb_burg_features
14        self.nb_burg_features = nb_burg_features
15        total_features = self.features_in + lpc_order
16        self.sequence_length = sequence_length
17        self.nb_features = nb_features
18
19        self.features = np.memmap(feature_file, dtype='float32', mode='r')
20        self.lost = np.memmap(loss_file, dtype='int8', mode='r')
21        self.lost = self.lost.astype('float32')
22
23        self.nb_sequences = self.features.shape[0]//self.sequence_length//total_features
24
25        self.features = self.features[:self.nb_sequences*self.sequence_length*total_features]
26        self.features = self.features.reshape((self.nb_sequences, self.sequence_length, total_features))
27        self.features = self.features[:,:,:self.features_in]
28
29        #self.lost = self.lost[:(len(self.lost)//features.shape[1]-1)*features.shape[1]]
30        #self.lost = self.lost.reshape((-1, self.sequence_length))
31
32    def __len__(self):
33        return self.nb_sequences
34
35    def __getitem__(self, index):
36        features = self.features[index, :, :]
37        burg_lost = (np.random.rand(features.shape[0]) > .1).astype('float32')
38        burg_lost = np.reshape(burg_lost, (features.shape[0], 1))
39        burg_mask = np.tile(burg_lost, (1,self.nb_burg_features))
40
41        lost_offset = np.random.randint(0, high=self.lost.shape[0]-self.sequence_length)
42        lost = self.lost[lost_offset:lost_offset+self.sequence_length]
43        #randomly add a few 10-ms losses so that the model learns to handle them
44        lost = lost * (np.random.rand(lost.shape[-1]) > .02).astype('float32')
45        #randomly break long consecutive losses so we don't try too hard to predict them
46        lost = 1 - ((1-lost) * (np.random.rand(lost.shape[-1]) > .1).astype('float32'))
47        lost = np.reshape(lost, (features.shape[0], 1))
48        lost_mask = np.tile(lost, (1,features.shape[-1]))
49        in_features = features*lost_mask
50        in_features[:,:self.nb_burg_features] = in_features[:,:self.nb_burg_features]*burg_mask
51
52        #For the first frame after a loss, we don't have valid features, but the Burg estimate is valid.
53        #in_features[:,1:,self.nb_burg_features:] = in_features[:,1:,self.nb_burg_features:]*lost_mask[:,:-1,self.nb_burg_features:]
54        out_lost = np.copy(lost)
55        #out_lost[:,1:,:] = out_lost[:,1:,:]*out_lost[:,:-1,:]
56
57        out_features = np.concatenate([features[:,self.nb_burg_features:], 1.-out_lost], axis=-1)
58        burg_sign = 2*burg_lost - 1
59        # last dim is 1 for received packet, 0 for lost packet, and -1 when just the Burg info is missing
60        return in_features*lost_mask, lost*burg_sign, out_features
61