File size: 6,469 Bytes
4f1a5d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm

from components.semantic_extractor.ssl_model import get_ssl_model
from components.simcodec.model import SimCodec
from transformers import GPT2Config, GPT2LMHeadModel

class N2S(nn.Module):
    def __init__(self, hps):
        super().__init__()
        self.hps = hps
        self.wavlm, self.km = get_ssl_model(**hps.ssl_model)
        self.bos = 1
        self.eos = 2
        self.pad = 0
        self.shift_num = 3

        self.lm_conf = GPT2Config(
            vocab_size=self.hps.model['n2s_vocab_size'],  
            n_embd=self.hps.model['hidden_size'],  
            n_layer=self.hps.model['num_hidden_layers'],  
            n_head=self.hps.model['num_attention_heads'],  
            activation_function='gelu_new',
            n_positions=2048,  
            n_ctx=2048, 
            resid_pdrop=0.1,
            embd_pdrop=0.1,
            attn_pdrop=0.1,
            layer_norm_epsilon=1e-05,
            initializer_range=0.02,
            summary_type='mean',
            summary_use_proj=True,
            summary_activation=None,
            summary_proj_to_labels=True,
            summary_first_dropout=0.1,
            bos_token_id=self.bos,
            eos_token_id=self.eos,
        )
        self.lm = GPT2LMHeadModel(self.lm_conf)
    
    def extract_semantic(self, wavs, num_frames):
        padding_size = (0, 100)
        wavs = F.pad(wavs, padding_size, "constant", 0)
        num_frames += 100
        features = self.wavlm.extract_features(
            wavs,
            output_layer=6,
            ret_layer_results=False,
            input_length=num_frames
        )[0]
        b, t, d = features.shape
        tokens = self.km(features.reshape(-1, d), b=b, t=t)
        return tokens
    
    def inference(self, token_gen, pos_gen):
        predict_len = (token_gen.shape[1] - 1)
        truck_length = token_gen.shape[1]

        for j in tqdm(range(predict_len)):
            lm_outputs = self.lm(
                input_ids=token_gen,
                attention_mask=None,
                position_ids=pos_gen
            )
            logits = lm_outputs['logits']
            logits[:, :, 0:self.shift_num] = -1e5
            probs = logits[:, -1, :].softmax(dim=-1)

            dist = torch.distributions.categorical.Categorical(probs=probs)

            samples = dist.sample().unsqueeze(1).to(token_gen.device)
            token_gen = torch.cat([token_gen, samples], dim=1)
            pos_pad = torch.ones(pos_gen.shape[0]) * j
            pos_gen = torch.cat([pos_gen, pos_pad.unsqueeze(1).to(token_gen.device).long()], dim=1)
        
        return token_gen[:,truck_length:][0]


    def generate(self, mix):
        mix = mix.squeeze(1)
        num_frame = torch.LongTensor([mix.shape[1]]).to(mix.device)
        token_s = self.extract_semantic(mix, num_frames=num_frame)

        token_s += 3
        bos = torch.ones(token_s.shape[0],1).long().to(mix.device)
        token_gen = torch.cat([token_s, bos], dim=1)

        pos_gen_id = torch.from_numpy(np.asarray(list(range(token_s.shape[1] + 1)))).to(mix.device)
        pos_gen = []
        for i in range(token_s.shape[0]):
            pos_gen.append(pos_gen_id.unsqueeze(0))
        pos_gen = torch.cat(pos_gen, dim=0)

        clean_s = self.inference(token_gen, pos_gen) - self.shift_num
        token_s -= self.shift_num
        return token_s, clean_s


class S2S(nn.Module):
    def __init__(self, hps):
        super().__init__()
        self.hps = hps
        self.codec_tokenizer = SimCodec(hps.path['codec_config_path'])
        self.wavlm, self.km = get_ssl_model(**hps.ssl_model)
        self.bos = 1
        self.eos = 2
        self.pad = 0
        self.shift_num = 3 + self.hps.model['semantic_num']
        self.lm_conf = GPT2Config(
            vocab_size=self.hps.model['s2s_vocab_size'],  
            n_embd=self.hps.model['hidden_size'],  
            n_layer=self.hps.model['num_hidden_layers'],  
            n_head=self.hps.model['num_attention_heads'],  
            activation_function='gelu_new',
            n_positions=4096,
            n_ctx=4096,
            resid_pdrop=0.1,
            embd_pdrop=0.1,
            attn_pdrop=0.1,
            layer_norm_epsilon=1e-05,
            initializer_range=0.02,
            summary_type='mean',
            summary_use_proj=True,
            summary_activation=None,
            summary_proj_to_labels=True,
            summary_first_dropout=0.1,
            bos_token_id=self.bos,
            eos_token_id=self.eos,
        )
        self.lm = GPT2LMHeadModel(self.lm_conf)
    
    def inference(self, token_gen, pos_gen):
        predict_len = int((token_gen.shape[1] - 1) / 2)
        truck_length = token_gen.shape[1]
        for j in tqdm(range(predict_len)):
            lm_outputs = self.lm(
                input_ids=token_gen,
                attention_mask=None,
                position_ids=pos_gen
            )
            logits = lm_outputs['logits']
            logits[:, :, 0:self.shift_num] = -1e5
            probs = logits[:, -1, :].softmax(dim=-1)
            dist = torch.distributions.categorical.Categorical(probs=probs)
            samples = dist.sample().unsqueeze(1).to(token_gen.device)
            token_gen = torch.cat([token_gen, samples], dim=1)
            pos_pad = torch.ones(pos_gen.shape[0]) * (j + 1000)
            pos_gen = torch.cat([pos_gen, pos_pad.unsqueeze(1).to(token_gen.device).long()], dim=1)

        return token_gen[:,truck_length:][0]
    
    def generate(self, mix, mix_s, clean_s):
        mix_a = self.codec_tokenizer(mix).squeeze(-1)
        if len(clean_s.shape) == 1:
            clean_s = clean_s.unsqueeze(0)
        
        mix_s += 3
        clean_s += 3
        mix_a += self.shift_num

        bos = torch.ones(mix_s.shape[0],1).long().to(mix.device)
        token_gen = torch.cat([mix_s, clean_s, bos, mix_a], dim=1)

        pos_gen_id = torch.from_numpy(np.asarray(list(range(mix_s.shape[1] + clean_s.shape[1] + 1)) + list(range(mix_a.shape[1])))).to(mix.device)
        pos_gen = []
        for i in range(mix_s.shape[0]):
            pos_gen.append(pos_gen_id.unsqueeze(0))
        pos_gen = torch.cat(pos_gen, dim=0)

        pre_a = self.inference(token_gen, pos_gen) - self.shift_num
        gen_wav = self.codec_tokenizer.decode(pre_a.unsqueeze(0).unsqueeze(2)).squeeze(0).cpu()

        return gen_wav