-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
184 lines (155 loc) · 6.47 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
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
177
178
179
180
181
182
183
184
import numpy as np
import torch
from torch import nn
from torch.nn import functional as F
default_config = {
'cnn_ch_in': 1,
'cnn_ch_out': 384, # for now cnn_ch_out need to be the same as cnn_ch_in
'cnn_kernel_size': 15,
'cnn_stride': 1,
'cnn_padding': 0,
# output length : (ch_in - kernel + 2 * pad) / stride + 1
# to keep length: pad = ((ch_out - 1) * stride - ch_in + kernel) / 2
'cnn_num_layer': 4,
'cnn_drop_prob': 0.1,
'cnn_act_func': F.relu,
'cnn_signal_length': 100,
'trans_max_pos': 256,
# 16 * (100 - 14 * 4) # should be the same as cnn_ch_out * length
'trans_hidden': 704,
'trans_drop_prob': 0.1,
'trans_num_head': 8,
'trans_feedforward': 1024,
'trans_num_layer': 8,
'trans_act_func': 'relu',
# 'cls_ch_in': 256,
'cls_ch_out': 128, # 88 keys + not used. probabilities for each note
'cls_act_func': torch.sigmoid,
'num_batch': 10,
'num_epoch': 10,
}
debug_config = {
'cnn_ch_in': 1,
'cnn_ch_out': 8, # for now cnn_ch_out need to be the same as cnn_ch_in
'cnn_kernel_size': 15,
'cnn_stride': 1,
'cnn_padding': 0,
# output length : (ch_in - kernel + 2 * pad) / stride + 1
# to keep length: pad = ((ch_out - 1) * stride - ch_in + kernel) / 2
'cnn_num_layer': 4,
'cnn_drop_prob': 0.1,
'cnn_act_func': F.relu,
'cnn_signal_length': 100,
'trans_max_pos': 100,
'trans_hidden': 352, # should be the same as cnn_ch_out
'trans_drop_prob': 0.1,
'trans_num_head': 8,
'trans_feedforward': 512,
'trans_num_layer': 8,
'trans_act_func': 'relu', # F.relu,
# 'cls_ch_in': 8,
'cls_ch_out': 128, # 88 keys + not used. probabilities for each note
'cls_act_func': torch.sigmoid,
'num_batch': 10,
'num_epoch': 10,
}
class AcousticsCNN(nn.Module): # todo: different ch in ch out
def __init__(self, config):
super(AcousticsCNN, self).__init__()
self.config = config
self.convs = [nn.Conv1d(config['cnn_ch_in'],
config['cnn_ch_out'],
config['cnn_kernel_size'],
config['cnn_stride'],
config['cnn_padding'])]
self.convs += [nn.Conv1d(config['cnn_ch_out'],
config['cnn_ch_out'],
config['cnn_kernel_size'],
config['cnn_stride'],
config['cnn_padding'])
for i in range(config['cnn_num_layer'] - 1)]
self.convs = nn.ModuleList(self.convs)
# self.norm = nn.LayerNorm(todo)
# self.drop = nn.Dropout(config['cnn_drop_prob']) # will do dropout in position embedding
self.act = config['cnn_act_func']
def forward(self, x):
r"""
note that x should be shaped in [`numBatch`, `inChannel`, `signalLength`],
and output will be shaped in [`numBatch`, `outChannel`, `signalLength'`].
`signalLength'` is decided by `stride` and `padding`
(default: `signalLength'` = `signalLength`).
"""
for co in self.convs:
x = self.act(co(x))
# x = self.norm(x) # todo: how many times do i need to norm
# x = self.drop(x)
return x
class PositionEmbedding(nn.Module):
r"""
adpated from transformers package by huggingface.
"""
def __init__(self, config):
super(PositionEmbedding, self).__init__()
self.config = config
self.pos_embs = nn.Embedding(config['trans_max_pos'],
config['trans_hidden'])
self.LayerNorm = nn.LayerNorm(config['trans_hidden'])
self.dropout = nn.Dropout(config['trans_drop_prob'])
def forward(self, input_embs):
r"""
`input_embs` should be shaped as [`numBatch`, `seqLength`, `hiddenSize`]
"""
seq_length = input_embs.size(1)
position_ids = torch.arange(
seq_length, dtype=torch.long, device=input_embs.device)
position_ids = position_ids.unsqueeze(0).expand_as(input_embs[:,:,0])
position_embeddings = self.pos_embs(position_ids)
embeddings = input_embs + position_embeddings
embeddings = self.LayerNorm(embeddings)
embeddings = self.dropout(embeddings)
return embeddings
class TransformerEncoder(nn.Module):
def __init__(self, config):
super(TransformerEncoder, self).__init__()
self.config = config
self.position = PositionEmbedding(config)
self.encoder_layer = nn.TransformerEncoderLayer(d_model=config['trans_hidden'],
nhead=config['trans_num_head'],
dim_feedforward=config['trans_feedforward'],
dropout=config['trans_drop_prob'],
activation=config['trans_act_func'])
self.encoder = nn.TransformerEncoder(encoder_layer=self.encoder_layer,
num_layers=config['trans_num_layer'])
def forward(self, x):
x = self.position(x)
x = self.encoder_layer(x)
return x
class ClassificationLayer(nn.Module):
def __init__(self, config):
super(ClassificationLayer, self).__init__()
self.config = config
self.fcn = nn.Linear(config['trans_hidden'], config['cls_ch_out'])
self.act = config['cls_act_func']
def forward(self, x):
return self.act(self.fcn(x))
class MusicEye(nn.Module):
def __init__(self, config):
super(MusicEye, self).__init__()
self.config = config
self.acoustics = AcousticsCNN(config)
self.encoder = TransformerEncoder(config)
self.cls = ClassificationLayer(config)
self.hidden_states = torch.zeros([config['num_batch'],
config['trans_max_pos'],
config['trans_hidden']]).to('cuda')
# self.hidden_states = nn.Parameter(self.hidden_states)
def forward(self, x):
x = self.acoustics(x)
x = x.view([x.shape[0], 1, -1])
self.update_hidden_states(x)
x = self.encoder(self.hidden_states)
return self.cls(x)
def update_hidden_states(self, x):
self.hidden_states.detach_()
self.hidden_states = torch.cat([self.hidden_states[:, 1:, :],
x], dim=1)