-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathvq_vae_patch_embedd.py
167 lines (128 loc) · 6.44 KB
/
vq_vae_patch_embedd.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
import torch
from torch import nn
from model.autencoder_lightning_base import Autoencoder
from model.vector_quantizer import VectorQuantizer, ResidualVQLightning
class PatchEmbedding(nn.Module):
def __init__(self, patch_size, embed_dim):
super().__init__()
self.patch_size = patch_size
self.proj = nn.Conv1d(1, embed_dim, kernel_size=patch_size, stride=patch_size)
def forward(self, x):
x = x.permute(0, 2, 1)
x = x.reshape(x.shape[0], -1).unsqueeze(1)
x = self.proj(x)
return x
class PatchEmbeddingInverse(nn.Module):
def __init__(self, patch_size, embed_dim, input_dim):
super().__init__()
self.patch_size = patch_size
if patch_size == 25:
kernel_size = 5
self.proj = nn.Sequential(
nn.ConvTranspose1d(embed_dim, embed_dim, kernel_size=kernel_size, stride=kernel_size),
nn.BatchNorm1d(embed_dim),
nn.GELU(),
nn.ConvTranspose1d(embed_dim, 1, kernel_size=kernel_size, stride=kernel_size),
)
elif patch_size == 10:
self.proj = nn.Sequential(
nn.ConvTranspose1d(embed_dim, embed_dim, kernel_size=2, stride=2),
nn.BatchNorm1d(embed_dim),
nn.GELU(),
nn.ConvTranspose1d(embed_dim, 1, kernel_size=5, stride=5),
)
elif patch_size == 50:
self.proj = nn.Sequential(
nn.ConvTranspose1d(embed_dim, embed_dim, kernel_size=10, stride=10),
nn.BatchNorm1d(embed_dim),
nn.GELU(),
nn.ConvTranspose1d(embed_dim, 1, kernel_size=5, stride=5),
)
else:
raise NotImplementedError(f"Patch size not implemented: {patch_size}")
self.input_dim = input_dim
def forward(self, x):
x = self.proj(x)
x = x.reshape(x.shape[0], -1, self.input_dim)
return x
class ResBlock(nn.Module):
def __init__(self, channels: int, kernel_size: int = 3, stride: int = 1, padding: int = 1, dropout_p: float = 0.1, batch_norm: bool = True):
super().__init__()
self.block = nn.Sequential(
nn.GELU(),
nn.Conv1d(channels, channels, kernel_size=kernel_size, stride=stride, padding=padding),
nn.BatchNorm1d(channels) if batch_norm else nn.Identity(),
nn.GELU(),
nn.Conv1d(channels, channels, kernel_size=kernel_size, stride=stride, padding=padding),
nn.BatchNorm1d(channels) if batch_norm else nn.Identity(),
nn.Dropout(p=dropout_p)
)
def forward(self, x):
return x + self.block(x)
class SepCNNBlock(nn.Module):
def __init__(self, hidden_dim: int, embedding_dim: int) -> None:
super().__init__()
self.shared_conv = nn.Conv1d(hidden_dim, embedding_dim, kernel_size=1, stride=1, padding=0)
def forward(self, x):
_, _, dims = x.shape
cnn_out = []
for i in range(dims):
x_t = self.shared_conv(x[:,:,i].unsqueeze(2))
cnn_out.append(x_t)
x = torch.cat(cnn_out, dim=2)
return x.permute(0, 2, 1)
class CNNBlock(nn.Module):
def __init__(self, embed_dim: int, seperate: bool = True, kernel_size: int = 3, stride: int = 1, padding: int = 1, dropout_p: float = 0.1, batch_norm: bool = True, n_resblocks : int = 1):
super(CNNBlock, self).__init__()
# Single convolutional layer blocks whose weights will be shared
self.seperate = seperate
self.shared_conv = nn.Sequential(
*[ResBlock(channels=embed_dim, kernel_size=kernel_size, stride=stride, padding=padding, dropout_p=dropout_p, batch_norm=batch_norm) for _ in range(n_resblocks)],
)
def forward(self, x):
_, _, dims = x.shape
if self.seperate:
cnn_out = []
for i in range(dims):
x_t = self.shared_conv(x[:,:,i].unsqueeze(2))
cnn_out.append(x_t)
x = torch.cat(cnn_out, dim=2)
else:
x = self.shared_conv(x)
return x
class VQVAEPatch(Autoencoder):
def __init__(self, hidden_dim: int, input_dim: int, num_embeddings: int, embedding_dim: int,
n_resblocks: int, learning_rate: float, dropout_p: float=0.1, patch_size: int=25, seq_len: int=200, batch_norm: bool = True, beta: float=0.25,
use_improved_vq: bool = False, kmeans_iters: int = 0, threshold_ema_dead_code: int = 2):
super().__init__(hidden_dim=hidden_dim, input_dim=input_dim, num_embeddings=num_embeddings,
embedding_dim=embedding_dim, n_resblocks=n_resblocks, learning_rate=learning_rate, seq_len=seq_len, dropout_p=dropout_p)
self.patch_embed = PatchEmbedding(patch_size=patch_size, embed_dim=hidden_dim)
self.encoder = nn.Sequential(
CNNBlock(embed_dim=hidden_dim, n_resblocks=n_resblocks, dropout_p=dropout_p, batch_norm=batch_norm),
SepCNNBlock(hidden_dim=hidden_dim, embedding_dim=embedding_dim)
)
if use_improved_vq:
self.vector_quantization = ResidualVQLightning(
num_quantizers=1, e_dim=embedding_dim, n_e=num_embeddings,
kmeans_init=True, kmeans_iters=kmeans_iters, threshold_ema_dead_code=threshold_ema_dead_code
)
else:
self.vector_quantization = VectorQuantizer(n_e=num_embeddings, e_dim=embedding_dim, beta=beta)
self.decoder = nn.Sequential(
nn.Conv1d(embedding_dim, hidden_dim, kernel_size=1, stride=1, padding=0),
CNNBlock(embed_dim=hidden_dim, seperate=False, n_resblocks=n_resblocks, dropout_p=dropout_p, batch_norm=batch_norm)
)
self.reverse_patch_embed = PatchEmbeddingInverse(patch_size=patch_size, embed_dim=hidden_dim, input_dim=input_dim)
self.enc_out_len = seq_len // patch_size * input_dim
self.patch_size = patch_size
self.apply(self.weights_init)
def forward(self, x):
# print("input", x.shape)
x = self.patch_embed(x)
z_e = self.encoder(x)
# print("z_e", z_e.shape)
embedding_loss, z_q, perplexity, _, _ = self.vector_quantization(z_e)
# print("z_q", z_q.shape)
x_hat = self.decoder(z_q.permute(0, 2, 1))
x_hat = self.reverse_patch_embed(x_hat)
return embedding_loss, x_hat, perplexity