File size: 4,948 Bytes
081442b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

class VAE(nn.Module):
    def __init__(self, input_dim, hidden_dim, latent_dim, num_styles=2):
        super(VAE, self).__init__()
        self.input_dim = input_dim
        self.latent_dim = latent_dim
        self.hidden_dim = hidden_dim

        self.encode = Encoder(self.input_dim, self.hidden_dim, self.latent_dim)
        self.decode = Decoder(self.latent_dim, self.hidden_dim, self.input_dim)
        self.style_classifier = StyleClassifier(self.latent_dim, num_styles)

    def reparameterize(self, mu, logvar):
        std = torch.exp(0.5 * logvar)
        eps = torch.randn_like(std)
        return mu + eps * std

    def forward(self, x, right=None, left=None, check=False):
        mu, logvar, output = self.encode(x)
        z = self.reparameterize(mu, logvar)
        style_pred = self.style_classifier(z)
        decod = self.decode(z, output)
        return decod, mu, logvar, style_pred

class Encoder(nn.Module):
    def __init__(self, input_dim, hidden_dim, latent_dim):
        super(Encoder, self).__init__()
        self.hidden_dim = hidden_dim

        self.gru_piano_right = nn.GRU(input_dim, hidden_dim, batch_first=True)
        self.gru_piano_left = nn.GRU(input_dim, hidden_dim, batch_first=True)
        
        self.dense_layer = nn.Linear(hidden_dim * 2, hidden_dim, bias = True)

        self.fc_mu = nn.Linear(hidden_dim, latent_dim, bias = True)  
        self.fc_logvar = nn.Linear(hidden_dim, latent_dim, bias = True)  
    
    def forward(self, x):
        input_list = torch.chunk(x, 2, dim=1)
        right_input = input_list[0] # torch.Size([Batch Size, Sequence length, input length]) 
        left_input = input_list[1] 

        # initialize hidden state
        h0 = torch.zeros(1, right_input.size(0), self.hidden_dim, device=right_input.device)

        # Forward pass through GRU for each instrument
        o_r, h_r = self.gru_piano_right(right_input, h0)
        o_l, h_l = self.gru_piano_left(left_input, h0)
        
        output = torch.cat((o_r, o_l), dim=1)
        h = torch.cat((h_r[-1,], h_l[-1,]), dim=1)

        h = self.dense_layer(h)
        h = F.relu(h)
        mu = self.fc_mu(h)
        mu = F.relu(mu)
        logvar = self.fc_logvar(h)
        logvar = F.relu(logvar) + 1e-4
        return mu, logvar, output


class Decoder(nn.Module):
    def __init__(self, latent_dim, hidden_dim, output_dim):
        super(Decoder, self).__init__()
        self.latent_dim = latent_dim
        self.output_dim = output_dim

        self.latent_to_hidden = nn.Linear(latent_dim, latent_dim, bias = True)

        self.piano_right_layer = nn.Linear(latent_dim, hidden_dim, bias = True)
        self.piano_left_layer = nn.Linear(latent_dim, hidden_dim, bias = True)

        self.r_layer = nn.Linear(hidden_dim, output_dim, bias = True)
        self.l_layer = nn.Linear(hidden_dim, output_dim, bias = True)

        self.gru_piano_right_cell = nn.GRUCell(output_dim, hidden_dim)
        self.gru_piano_left_cell = nn.GRUCell(output_dim, hidden_dim)


        self.fr_layer = nn.Linear(hidden_dim, output_dim, bias = True)
        self.fl_layer = nn.Linear(hidden_dim , output_dim, bias = True) 

    def forward(self, z, output):
        
        h = self.latent_to_hidden(z)
        h = F.relu(h)
        
        right = torch.split(output, 150, dim=1)[0]
        left = torch.split(output, 150, dim=1)[1] 

        right = right.permute(1, 0, 2)
        left = left.permute(1, 0, 2)

        right = self.r_layer(right)
        right = F.tanh(right)
        left = self.l_layer(left)            
        left = F.tanh(left)


        piano_hidden = self.piano_right_layer(h) 
        left_hidden = self.piano_left_layer(h)

        right_outputs = []
        left_outputs = []

        for t in range(right.size(0)):
            piano_hidden = self.gru_piano_right_cell(right[t] , piano_hidden)
            left_hidden = self.gru_piano_left_cell(left[t], left_hidden)
            
            right_outputs.append(piano_hidden.unsqueeze(1))
            left_outputs.append(left_hidden.unsqueeze(1))

        # print(right_outputs[0].shape)
        right_outputs = torch.cat(right_outputs, dim=1)
        left_outputs = torch.cat(left_outputs, dim=1)

        right_outputs = self.fr_layer(right_outputs)
        left_outputs = self.fl_layer(left_outputs)
        
        right_outputs = F.sigmoid(right_outputs)
        left_outputs = F.sigmoid(left_outputs)

        output = torch.cat((right_outputs, left_outputs), dim=1)

        return output

class StyleClassifier(nn.Module):
    def __init__(self, latent_dim, num_styles):
        super(StyleClassifier, self).__init__()
        self.fc1 = nn.Linear(latent_dim, 128)
        self.fc2 = nn.Linear(128, num_styles)

    def forward(self, z):
        x = F.relu(self.fc1(z))
        x = self.fc2(x)
        return F.softmax(x, dim=-1)