-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodels.py
137 lines (119 loc) · 6.88 KB
/
models.py
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
import torch
import torch.nn as nn
import torch.nn.functional as F
class SeqBaseModel(nn.Module):
def __init__(self, args):
super(SeqBaseModel, self).__init__()
self.num_tokens = len(args.num_tokens)
self.encoders = nn.ModuleList([nn.Embedding(num_tokens, n_emb) for n_emb, num_tokens in
zip(args.nembs, args.num_tokens)])
self.rnn = nn.GRU(sum(args.nembs), args.nhid, args.nlayers, dropout=args.dropout, batch_first=True)
def seq(self, part, part_length, recover_idx=None, bs=1):
if len(part) == 0:
return torch.zeros((bs, self.args.nhid)).cuda()
part = torch.cat([self.encoders[i](part[:, :, i]) for i in range(self.num_tokens)], -1)
if len(part) > 1:
part = torch.nn.utils.rnn.pack_padded_sequence(part, part_length, batch_first=True)
part, _ = self.rnn(part)
part, _ = torch.nn.utils.rnn.pad_packed_sequence(part, batch_first=True)
part_length = part_length.view(len(part), 1, 1).expand(-1, 1, part.shape[-1])
part = F.dropout(torch.gather(part, 1, part_length - 1), p=self.args.dropout).view(len(part), -1)
if bs > len(part):
part = torch.cat([part, torch.zeros((bs - len(part), part.shape[-1])).cuda()])
if recover_idx is not None:
part = part[recover_idx]
else:
part, _ = self.rnn(part)
part = F.dropout(part[:, -1, :], p=self.args.dropout).view(len(part), -1)
if bs > len(part):
part = torch.cat([part, torch.zeros((bs - len(part), part.shape[-1])).cuda()])
return part
class Generator(SeqBaseModel):
def __init__(self, args):
super(Generator, self).__init__(args)
self.meta_encoder = nn.Embedding(16, args.nemb_meta)
self.central_meta_fc = nn.Linear(args.nemb_meta, args.meta_nfc_cent)
self.central_fc = nn.Linear(args.nembs[0], args.nfc_cent)
self.left_fc = nn.Linear(2 * args.nhid, args.nfc_left)
self.fc = nn.Linear(args.nfc_left + args.nfc_cent + args.meta_nfc_cent, args.pred_nfc)
self.pred = nn.Linear(args.pred_nfc, args.num_pitches * args.num_durations)
self.value = nn.Linear(args.pred_nfc, 1)
self.args = args
def forward(self, inputs):
self_left, self_length, partner_left, partner_length, recover_idx, partner_central, meta_central = inputs
self_left = self.seq(self_left, self_length)
partner_left = self.seq(partner_left, partner_length, recover_idx, bs=len(self_left))
left_feature = self.left_fc(torch.cat([self_left, partner_left], -1))
meta_central_emb = self.meta_encoder(meta_central)
meta_central_feature = self.central_meta_fc(F.dropout(meta_central_emb, p=self.args.dropout))
central_emb = self.encoders[0](partner_central)
central_feature = self.central_fc(F.dropout(central_emb, p=self.args.dropout))
feature = torch.cat((left_feature, central_feature, meta_central_feature), dim=-1)
last_feature = F.relu(self.fc(feature))
pred = self.pred(F.dropout(last_feature, p=self.args.dropout))
value = self.value(F.dropout(last_feature, p=self.args.dropout))
return pred, value
class StyleRewarder(SeqBaseModel):
def __init__(self, args):
super(StyleRewarder, self).__init__(args)
self.left_fc = nn.Linear(args.nhid, args.nfc_left)
self.fc = nn.Linear(args.nfc_left, args.pred_nfc)
self.pred = nn.Linear(args.pred_nfc, args.num_pitches * args.num_durations)
self.args = args
def forward(self, inputs):
self_left, self_length, _ = inputs
self_left = self.seq(self_left, self_length)
feature = self.left_fc(self_left)
pred = self.pred(F.relu(self.fc(feature)))
return pred, None
def reward(self, self_part, return_seq=False):
self_emb = torch.cat([self.encoders[i](self_part[:, :, i]) for i in range(self.num_tokens)], -1)
self_part, _ = self.rnn(self_emb)
if not return_seq:
self_part = self_part[:, -1, :].view(len(self_part), -1)
feature = self.left_fc(self_part)
pred = self.pred(F.relu(self.fc(feature)))
return pred
class BachM(SeqBaseModel):
def __init__(self, args):
super(BachM, self).__init__(args)
self.meta_encoder = nn.Embedding(16, args.nemb_meta)
self.central_meta_fc = nn.Linear(args.nemb_meta, args.meta_nfc_cent)
self.left_fc = nn.Linear(args.nhid, args.nfc_left)
self.fc = nn.Linear(args.nfc_left + args.meta_nfc_cent, args.pred_nfc)
self.pred = nn.Linear(args.pred_nfc, args.num_pitches * args.num_durations)
self.args = args
def forward(self, inputs):
self_left, self_length, meta_central = inputs
self_left = self.seq(self_left, self_length)
left_feature = self.left_fc(self_left)
meta_central_emb = self.meta_encoder(meta_central)
meta_central_feature = self.central_meta_fc(F.dropout(meta_central_emb, p=self.args.dropout))
feature = torch.cat((left_feature, meta_central_feature), dim=-1)
pred = self.pred(F.dropout(F.relu(self.fc(feature)), p=self.args.dropout))
return pred, None
class BachHM(SeqBaseModel):
def __init__(self, args):
super(BachHM, self).__init__(args)
self.meta_encoder = nn.Embedding(16, args.nemb_meta)
self.central_meta_fc = nn.Linear(args.nemb_meta, args.meta_nfc_cent)
self.central_fc = nn.Linear(sum(args.nembs), args.nfc_cent)
self.rnn_fc = nn.Linear(3 * args.nhid, args.nfc_left)
self.fc = nn.Linear(args.nfc_left + args.nfc_cent + args.meta_nfc_cent, args.pred_nfc)
self.pred = nn.Linear(args.pred_nfc, args.num_pitches * args.num_durations)
self.args = args
def forward(self, inputs):
self_left, self_length, partner_left, p_left_length, p_left_recover_idx, \
partner_central, meta_central, partner_right, p_right_length, p_right_recover_idx = inputs
partner_left = self.seq(partner_left, p_left_length, p_left_recover_idx, bs=len(self_left))
partner_right = self.seq(partner_right, p_right_length, p_right_recover_idx, bs=len(self_left))
self_left = self.seq(self_left, self_length)
rnn_feature = torch.cat([self_left, partner_left, partner_right], -1)
rnn_feature = self.rnn_fc(rnn_feature)
central_emb = torch.cat([self.encoders[i](partner_central[..., i]) for i in range(self.num_tokens)], -1)
central_feature = self.central_fc(F.dropout(central_emb, p=self.args.dropout))
meta_central_emb = self.meta_encoder(meta_central)
meta_central_feature = self.central_meta_fc(F.dropout(meta_central_emb, p=self.args.dropout))
feature = torch.cat((rnn_feature, central_feature, meta_central_feature), dim=-1)
pred = self.pred(F.dropout(F.relu(self.fc(feature)), p=self.args.dropout))
return pred, None