File size: 4,581 Bytes
e59f78e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F


# define a mlp encoder
# inputs: batch x num_genes (2446)
# outputs: batch x ecoding_dim 
class Encoder(nn.Module):

     def __init__(self, latent_dim, hidden_dims, num_genes=2446):
          super().__init__()
          
          layers = []

          prev_dim = num_genes
          for h_dim in hidden_dims:
               layers.append(nn.Linear(prev_dim, h_dim))
               layers.append(nn.BatchNorm1d(h_dim))
               layers.append(nn.ELU())
               layers.append(nn.Dropout(0.4))
               prev_dim = h_dim

          self.enc_net = nn.Sequential(*layers)

          self.fc_mean = nn.Linear(prev_dim, latent_dim)
          self.fc_std = nn.Linear(prev_dim, latent_dim)


     def forward(self, x_t):

          h = self.enc_net(x_t)

          mean = self.fc_mean(h)

          # Ensure minimum std to prevent posterior collapse
          # Higher minimum (1e-3) prevents std from collapsing to near-zero
          std = F.softplus(self.fc_std(h)) + 1e-3

          return mean, std


# define a corresponding mlp decoder
# input: batch x ecoding_dim + rnn_hidden_dim 
class Decoder(nn.Module):

     def __init__(self, latent_dim, rnn_hidden_dim, hidden_dims, num_genes=2446):
          super().__init__()

          layers = []

          prev_dim = latent_dim + rnn_hidden_dim

          for h_dim in hidden_dims:
               layers.append(nn.Linear(prev_dim, h_dim))
               layers.append(nn.BatchNorm1d(h_dim))
               layers.append(nn.ELU())
               layers.append(nn.Dropout(0.4))
               prev_dim = h_dim

          layers.append(nn.Linear(prev_dim, num_genes))
          self.dec_net = nn.Sequential(*layers)

     
     def forward(self, z, h):

          inps = torch.cat([z, h], dim=1)

          return self.dec_net(inps)
     
# define a gru-based rssm
# input: batch x ecoding_dim at t=0
# output: batch x 2*encoding_dim at t = 1 to get the mean and standard deviation

class RSSM(nn.Module):

     def __init__(self, latent_dim, rnn_hidden_dim):
          super().__init__()

          self.latent_dim = latent_dim
          self.hidden_dim = rnn_hidden_dim
          

          self.gru = nn.GRUCell(latent_dim, rnn_hidden_dim)
          self.mlp = nn.Sequential(
               nn.Linear(rnn_hidden_dim, rnn_hidden_dim), 
               nn.LayerNorm(rnn_hidden_dim),
               nn.ELU(), 
               nn.Linear(rnn_hidden_dim, 2 * latent_dim)
          )
          
          # Better initialization: larger std prevents weak prior
          # Use Xavier/Glorot initialization for better gradient flow
          nn.init.xavier_uniform_(self.mlp[3].weight, gain=0.1)
          nn.init.zeros_(self.mlp[3].bias)

     def forward(self, prev_r, prev_h):

          h_t_1 = self.gru(prev_r, prev_h)

          prev_stats = self.mlp(h_t_1)

          prev_mean, prev_std = torch.chunk(prev_stats, 2, dim=1)

          prev_std = F.softplus(prev_std) + 1e-3

          return h_t_1, prev_mean, prev_std
     

# create joint training architecture for dreamer
class CellDreamer(nn.Module):

     def __init__(
          self,
          device,
          latent_dim = 20,
          rnn_dim = 64,
          enc_hidden_dims = [128, 64, 32],
          dec_hidden_dims = [32, 64, 128],
          num_genes = 2446
     ):
          super().__init__()

          self.encoder = Encoder(latent_dim, enc_hidden_dims, num_genes)
          self.decoder = Decoder(latent_dim, rnn_dim, dec_hidden_dims, num_genes)
          self.rssm = RSSM(latent_dim, rnn_dim)

          self.rnn_dim = rnn_dim
          self.latent_dim = latent_dim
          self.input_dim = num_genes
          self.device = device

     def reparametrize(self, mean, std):

          eps = torch.randn_like(std)
          return mean + eps * std
     
     def forward(self, x_t):

          post_mean, post_std = self.encoder(x_t)
          z_t = self.reparametrize(post_mean, post_std)

          h_prev = torch.zeros(x_t.size(0), self.rnn_dim).to(self.device)

          h_next, velocity_mean, velocity_std = self.rssm(z_t, h_prev)
          prior_next_mean = z_t + velocity_mean
          prior_next_std = velocity_std

          rec_x = self.decoder(z_t, h_next)

          return {
               "recon_x": rec_x,
               "post_mean": post_mean,
               "post_std": post_std,
               "prior_next_mean": prior_next_mean,
               "prior_next_std": prior_next_std,
               "z_t": z_t,
               "h_next": h_next
          }