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
| class mLSTMLayerConfig(UpProjConfigMixin):
conv1d_kernel_size: int = 4
qkv_proj_blocksize: int = 4
num_heads: int = 4
proj_factor: float = 2.0
embedding_dim: int = -1
bias: bool = False
dropout: float = 0.0
context_length: int = -1
_num_blocks: int = 1
_inner_embedding_dim: int = None
def __post_init__(self):
self._set_proj_up_dim(embedding_dim=self.embedding_dim)
self._inner_embedding_dim = self._proj_up_dim
class mLSTMLayer(nn.Module):
config_class = mLSTMLayerConfig
def __init__(self, config: mLSTMLayerConfig):
super().__init__()
self.config = config
self.proj_up = nn.Linear(
in_features=self.config.embedding_dim,
out_features=2 * self.config._inner_embedding_dim,
bias=self.config.bias,
)
num_proj_heads = round(self.config._inner_embedding_dim // self.config.qkv_proj_blocksize)
self.q_proj = LinearHeadwiseExpand(
config=LinearHeadwiseExpandConfig(
in_features=self.config._inner_embedding_dim,
num_heads=num_proj_heads,
bias=self.config.bias,
)
)
self.k_proj = LinearHeadwiseExpand(
config=LinearHeadwiseExpandConfig(
in_features=self.config._inner_embedding_dim,
num_heads=num_proj_heads,
bias=self.config.bias,
)
)
self.v_proj = LinearHeadwiseExpand(
config=LinearHeadwiseExpandConfig(
in_features=self.config._inner_embedding_dim,
num_heads=num_proj_heads,
bias=self.config.bias,
)
)
self.conv1d = CausalConv1d(
config=CausalConv1dConfig(
feature_dim=self.config._inner_embedding_dim,
kernel_size=self.config.conv1d_kernel_size,
)
)
self.conv_act_fn = nn.SiLU()
self.mlstm_cell = mLSTMCell(
config=mLSTMCellConfig(
context_length=self.config.context_length,
embedding_dim=self.config._inner_embedding_dim,
num_heads=self.config.num_heads,
)
)
self.ogate_act_fn = nn.SiLU()
self.learnable_skip = nn.Parameter(torch.ones(self.config._inner_embedding_dim, requires_grad=True))
self.proj_down = nn.Linear(
in_features=self.config._inner_embedding_dim,
out_features=self.config.embedding_dim,
bias=self.config.bias,
)
self.dropout = nn.Dropout(self.config.dropout)
self.reset_parameters()
def forward(self, x: torch.Tensor, **kwargs) -> torch.Tensor:
B, S, _ = x.shape
'''
batch：同dataloader输出的batch维度
time:时间序列
feature：每个序列的值，对于频谱则为特征频率的幅值
'''
x_inner = self.proj_up(x)
x_mlstm, z = torch.split(x_inner, split_size_or_sections=self.config._inner_embedding_dim, dim=-1)
x_mlstm_conv = self.conv1d(x_mlstm)
x_mlstm_conv_act = self.conv_act_fn(x_mlstm_conv)
q = self.q_proj(x_mlstm_conv_act)
k = self.k_proj(x_mlstm_conv_act)
v = self.v_proj(x_mlstm)
h_tilde_state = self.mlstm_cell(q=q, k=k, v=v)
h_tilde_state_skip = h_tilde_state + (self.learnable_skip * x_mlstm_conv_act)
h_state = h_tilde_state_skip * self.ogate_act_fn(z)
y = self.dropout(self.proj_down(h_state))
return y
def step(
self,
x: torch.Tensor,
mlstm_state: tuple[torch.Tensor, torch.Tensor, torch.Tensor] = None,
conv_state: tuple[torch.Tensor] = None,
) -> tuple[torch.Tensor, dict[str, tuple[torch.Tensor, ...]]]:
x_inner = self.proj_up(x)
x_mlstm, z = torch.split(x_inner, split_size_or_sections=self.config._inner_embedding_dim, dim=-1)
x_mlstm_conv, conv_state = self.conv1d.step(x_mlstm, conv_state=conv_state)
x_mlstm_conv_act = self.conv_act_fn(x_mlstm_conv)
q = self.q_proj(x_mlstm_conv_act)
k = self.k_proj(x_mlstm_conv_act)
v = self.v_proj(x_mlstm)
h_tilde_state, mlstm_state = self.mlstm_cell.step(q=q, k=k, v=v, mlstm_state=mlstm_state)
h_tilde_state_skip = h_tilde_state + (self.learnable_skip * x_mlstm_conv_act)
h_state = h_tilde_state_skip * self.ogate_act_fn(z)
y = self.dropout(self.proj_down(h_state))
return y, {"mlstm_state": mlstm_state, "conv_state": conv_state}
def reset_parameters(self):
small_init_init_(self.proj_up.weight, dim=self.config.embedding_dim)
if self.proj_up.bias is not None:
nn.init.zeros_(self.proj_up.bias)
wang_init_(self.proj_down.weight, dim=self.config.embedding_dim, num_blocks=self.config._num_blocks)
if self.proj_down.bias is not None:
nn.init.zeros_(self.proj_down.bias)
nn.init.ones_(self.learnable_skip)
def _init_qkv_proj(qkv_proj: LinearHeadwiseExpand):
small_init_init_(qkv_proj.weight, dim=self.config.embedding_dim)
if qkv_proj.bias is not None:
nn.init.zeros_(qkv_proj.bias)
_init_qkv_proj(self.q_proj)
_init_qkv_proj(self.k_proj)
_init_qkv_proj(self.v_proj)
self.mlstm_cell.reset_parameters()
|
