-
Notifications
You must be signed in to change notification settings - Fork 9
/
HIST2ST.py
239 lines (227 loc) · 8.67 KB
/
HIST2ST.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
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
import torch
import numpy as np
import pytorch_lightning as pl
import torchvision.transforms as tf
from gcn import *
from NB_module import *
from transformer import *
from scipy.stats import pearsonr
from torch.utils.data import DataLoader
from copy import deepcopy as dcp
from collections import defaultdict as dfd
from sklearn.metrics import adjusted_rand_score as ari_score
from sklearn.metrics.cluster import normalized_mutual_info_score as nmi_score
class convmixer_block(nn.Module):
def __init__(self,dim,kernel_size):
super().__init__()
self.dw=nn.Sequential(
nn.Conv2d(dim, dim, kernel_size, groups=dim, padding="same"),
nn.BatchNorm2d(dim),
nn.GELU(),
nn.Conv2d(dim, dim, kernel_size, groups=dim, padding="same"),
nn.BatchNorm2d(dim),
nn.GELU(),
)
self.pw=nn.Sequential(
nn.Conv2d(dim, dim, kernel_size=1),
nn.GELU(),
nn.BatchNorm2d(dim),
)
def forward(self,x):
x=self.dw(x)+x
x=self.pw(x)
return x
class mixer_transformer(nn.Module):
def __init__(self,channel=32, kernel_size=5, dim=1024,
depth1=2, depth2=8, depth3=4,
heads=8, dim_head=64, mlp_dim=1024, dropout = 0.,
policy='mean',gcn=True
):
super().__init__()
self.layer1=nn.Sequential(
*[convmixer_block(channel,kernel_size) for i in range(depth1)],
)
self.layer2=nn.Sequential(*[attn_block(dim,heads,dim_head,mlp_dim,dropout) for i in range(depth2)])
self.layer3=nn.ModuleList([gs_block(dim,dim,policy,gcn) for i in range(depth3)])
self.jknet=nn.Sequential(
nn.LSTM(dim,dim,2),
SelectItem(0),
)
self.down=nn.Sequential(
nn.Conv2d(channel,channel//8,1,1),
nn.Flatten(),
)
def forward(self,x,ct,adj):
x=self.down(self.layer1(x))
g=x.unsqueeze(0)
g=self.layer2(g+ct).squeeze(0)
jk=[]
for layer in self.layer3:
g=layer(g,adj)
jk.append(g.unsqueeze(0))
g=torch.cat(jk,0)
g=self.jknet(g).mean(0)
return g
class ViT(nn.Module):
def __init__(self, channel=32,kernel_size=5,dim=1024,
depth1=2, depth2=8, depth3=4,
heads=8, mlp_dim=1024, dim_head = 64, dropout = 0.,
policy='mean',gcn=True
):
super().__init__()
self.dropout = nn.Dropout(dropout)
self.transformer = mixer_transformer(
channel, kernel_size, dim,
depth1, depth2, depth3,
heads, dim_head, mlp_dim, dropout,
policy,gcn,
)
def forward(self,x,ct,adj):
x = self.dropout(x)
x = self.transformer(x,ct,adj)
return x
class Hist2ST(pl.LightningModule):
def __init__(self, learning_rate=1e-5, fig_size=112, label=None,
dropout=0.2, n_pos=64, kernel_size=5, patch_size=7, n_genes=785,
depth1=2, depth2=8, depth3=4, heads=16, channel=32,
zinb=0, nb=False, bake=0, lamb=0, policy='mean',
):
super().__init__()
# self.save_hyperparameters()
dim=(fig_size//patch_size)**2*channel//8
self.learning_rate = learning_rate
self.nb=nb
self.zinb=zinb
self.bake=bake
self.lamb=lamb
self.label=label
self.patch_embedding = nn.Conv2d(3,channel,patch_size,patch_size)
self.x_embed = nn.Embedding(n_pos,dim)
self.y_embed = nn.Embedding(n_pos,dim)
self.vit = ViT(
channel=channel, kernel_size=kernel_size, heads=heads,
dim=dim, depth1=depth1,depth2=depth2, depth3=depth3,
mlp_dim=dim, dropout = dropout, policy=policy, gcn=True,
)
self.channel=channel
self.patch_size=patch_size
self.n_genes=n_genes
if self.zinb>0:
if self.nb:
self.hr=nn.Linear(dim, n_genes)
self.hp=nn.Linear(dim, n_genes)
else:
self.mean = nn.Sequential(nn.Linear(dim, n_genes), MeanAct())
self.disp = nn.Sequential(nn.Linear(dim, n_genes), DispAct())
self.pi = nn.Sequential(nn.Linear(dim, n_genes), nn.Sigmoid())
if self.bake>0:
self.coef=nn.Sequential(
nn.Linear(dim,dim),
nn.ReLU(),
nn.Linear(dim,1),
)
self.gene_head = nn.Sequential(
nn.LayerNorm(dim),
nn.Linear(dim, n_genes),
)
self.tf=tf.Compose([
tf.RandomGrayscale(0.1),
tf.RandomRotation(90),
tf.RandomHorizontalFlip(0.2),
])
def forward(self, patches, centers, adj, aug=False):
B,N,C,H,W=patches.shape
patches=patches.reshape(B*N,C,H,W)
patches = self.patch_embedding(patches)
centers_x = self.x_embed(centers[:,:,0])
centers_y = self.y_embed(centers[:,:,1])
ct=centers_x + centers_y
h = self.vit(patches,ct,adj)
x = self.gene_head(h)
extra=None
if self.zinb>0:
if self.nb:
r=self.hr(h)
p=self.hp(h)
extra=(r,p)
else:
m = self.mean(h)
d = self.disp(h)
p = self.pi(h)
extra=(m,d,p)
if aug:
h=self.coef(h)
return x,extra,h
def aug(self,patch,center,adj):
bake_x=[]
for i in range(self.bake):
new_patch=self.tf(patch.squeeze(0)).unsqueeze(0)
x,_,h=self(new_patch,center,adj,True)
bake_x.append((x.unsqueeze(0),h.unsqueeze(0)))
return bake_x
def distillation(self,bake_x):
new_x,coef=zip(*bake_x)
coef=torch.cat(coef,0)
new_x=torch.cat(new_x,0)
coef=F.softmax(coef,dim=0)
new_x=(new_x*coef).sum(0)
return new_x
def training_step(self, batch, batch_idx):
patch, center, exp, adj, oris, sfs, *_ = batch
adj=adj.squeeze(0)
exp=exp.squeeze(0)
pred,extra,h = self(patch, center, adj)
mse_loss = F.mse_loss(pred, exp)
self.log('mse_loss', mse_loss,on_epoch=True, prog_bar=True, logger=True)
bake_loss=0
if self.bake>0:
bake_x=self.aug(patch,center,adj)
new_pred=self.distillation(bake_x)
bake_loss+=F.mse_loss(new_pred,pred)
self.log('bake_loss', bake_loss,on_epoch=True, prog_bar=True, logger=True)
zinb_loss=0
if self.zinb>0:
if self.nb:
r,p=extra
zinb_loss = NB_loss(oris.squeeze(0),r,p)
else:
m,d,p=extra
zinb_loss = ZINB_loss(oris.squeeze(0),m,d,p,sfs.squeeze(0))
self.log('zinb_loss', zinb_loss,on_epoch=True, prog_bar=True, logger=True)
loss=mse_loss+self.zinb*zinb_loss+self.lamb*bake_loss
return loss
def validation_step(self, batch, batch_idx):
patch, center, exp, adj, oris, sfs, *_ = batch
def cluster(pred,cls):
sc.pp.pca(pred)
sc.tl.tsne(pred)
kmeans = KMeans(n_clusters=cls, init="k-means++", random_state=0).fit(pred.obsm['X_pca'])
pred.obs['kmeans'] = kmeans.labels_.astype(str)
p=pred.obs['kmeans'].to_numpy()
return p
pred,extra,h = self(patch, center, adj.squeeze(0))
if self.label is not None:
adata=ann.AnnData(pred.squeeze().cpu().numpy())
idx=self.label!='undetermined'
cls=len(set(self.label))
x=adata[idx]
l=self.label[idx]
predlbl=cluster(x,cls-1)
self.log('nmi',nmi_score(predlbl,l))
self.log('ari',ari_score(predlbl,l))
loss = F.mse_loss(pred.squeeze(0), exp.squeeze(0))
self.log('valid_loss', loss,on_epoch=True, prog_bar=True, logger=True)
pred=pred.squeeze(0).cpu().numpy().T
exp=exp.squeeze(0).cpu().numpy().T
r=[]
for g in range(self.n_genes):
r.append(pearsonr(pred[g],exp[g])[0])
R=torch.Tensor(r).mean()
self.log('R', R, on_epoch=True, prog_bar=True, logger=True)
return loss
def configure_optimizers(self):
# self.hparams available because we called self.save_hyperparameters()
optim=torch.optim.Adam(self.parameters(), lr=self.learning_rate)
StepLR = torch.optim.lr_scheduler.StepLR(optim, step_size=50, gamma=0.9)
optim_dict = {'optimizer': optim, 'lr_scheduler': StepLR}
return optim_dict