-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainer.py
205 lines (174 loc) · 9.46 KB
/
trainer.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
import torch
from tqdm import tqdm
import random
import numpy as np
from sklearn.cluster import KMeans
from torch.nn.parameter import Parameter
from utils import *
import gc
class train():
def __init__(self, img_tensor, rna_tensor, edge_index, model, custom_decoder = True,n_epochs=100, opt="adam", lr=0.0001, weight_decay=0.0001, save_att=False, verbose=True):#, attn_out = None
self.img_tensor = img_tensor
self.rna_tensor = rna_tensor
self.edge_index = edge_index
self.model = model
self.custom_decoder = custom_decoder
self.n_epochs = n_epochs
self.save_att = save_att
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if opt=="sgd":
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=lr, momentum=0.9)
elif opt=="adam":
self.optimizer = torch.optim.Adam(self.model.parameters(),lr=lr, weight_decay=weight_decay)
def train_ae(self, gradient_clipping=5.):
print('tarining AE')
loss_list = []
pbar = tqdm(range(1, self.n_epochs+1),desc='Training model...')
current_memory = torch.cuda.memory_allocated() / 1024**2 # 转换为MB
for epoch in pbar:
self.model.train()
self.optimizer.zero_grad()
z, mean, logvar, x_hat, img_emb,rna_emb, attn_weight = self.model(self.img_tensor, self.rna_tensor, self.edge_index)
if self.custom_decoder == True:
concat_tensor = torch.cat((self.img_tensor,self.rna_tensor),1)
recon_loss = self.model.recon_loss(x_hat,concat_tensor)
# inner_loss = self.model.innerproduct_loss(z,self.edge_index)
kl_loss = self.model.kl_loss(mean, logvar)
loss = recon_loss + (1 / self.img_tensor.shape[0]) * kl_loss #+ inner_loss
else:
inner_loss = self.model.innerproduct_loss(z,self.edge_index)
kl_loss = self.model.kl_loss(mean, logvar)
loss = inner_loss + (1 / self.img_tensor.shape[0]) * kl_loss
loss_list.append(loss.item())
loss.backward()
#print(loss)
torch.nn.utils.clip_grad_norm_(self.model.parameters(), gradient_clipping)
self.optimizer.step()
gc.collect()
torch.cuda.empty_cache()
if epoch % 2 == 0:
pbar.set_postfix_str("loss: {:.4f}".format(loss.item()))
if self.save_att == True:
# print('attention weiget saved at: %s' % (attn_out))
# if not os.path.exists(attn_out):
# os.makedirs(attn_out)
self.save_attn_weight(epoch, attn_weight, inter=10, out_dir='./')
#tqdm.set_description("loss: {:.4f}".format(loss.item()))
#tqdm.set_postfix(loss=loss.item())
#tqdm.set_postfix_str("loss: {:.4f}".format(loss.item()))
# self.model.eval()
# with torch.no_grad():
# z, mean, logvar, x_hat, attn_weight = self.model(self.img_tensor, self.rna_tensor, self.edge_index)
# vae_emb = z.to('cpu').detach().numpy()
# return vae_emb,loss_list
return loss_list, img_emb, rna_emb, attn_weight
def train_dec(self, init = "kmeans", init_spa = True, n_cluster = 10, n_neighbors = 20, max_epochs = 100, update_interval = 3, tol = 1e-5, alpha = 0.9):
loss_list, img_emb, rna_emb, attn_weight = self.train_ae()
x = get_embedding(self.img_tensor, self.rna_tensor, self.edge_index, self.model)
self.model.dec = Parameter(torch.Tensor(n_cluster, x.shape[1])).to(self.device)
print('training use DEC')
# self.model.dec = Parameter(torch.Tensor(n_cluster, x.shape[0])).to(self.device)
# torch.nn.init.xavier_normal_(self.dec.data)
if torch.is_tensor(x) :
x_tensor = x.to(self.device)
elif isinstance(x, pd.DataFrame):
x_array = x.values.astype(np.float32)
x_tensor = torch.tensor(x_array).to(self.device)
else:
x_tensor = torch.tensor(x).to(self.device)
if init=="kmeans":
print(str("Initializing cluster centers with kmeans, n_clusters known: "+str(n_cluster)))
# self.n_clusters=n_clusters
kmeans = KMeans(n_cluster, n_init=20)
if init_spa:
#------Kmeans use exp and spatial
y_pred = kmeans.fit_predict(x)
else:
#------Kmeans only use exp info, no spatial
concat_tensor = torch.cat((self.img_tensor,self.rna_tensor),1)
y_pred = kmeans.fit_predict(concat_tensor.numpy())
elif init=="leiden":
print("Initializing cluster centers with leiden, resolution = ", res)
if init_spa:
adata=sc.AnnData(x)
else:
concat_tensor = torch.cat((self.img_tensor,self.rna_tensor),1)
adata=sc.AnnData(concat_tensor.numpy())
sc.pp.neighbors(adata, n_neighbors=n_neighbors)
res = choose_res(adata, cluster_num = n_cluster ,res_range = np.around(np.arange(0.3,0.9,0.04), 3), determine_clus_num = True)
adata = definite_res(adata,res,'./',plot_file='init_spatial_plot.png',title='init_leiden')
try :
select_res(adata,res,method='leiden',plot=True,title='init_plot')
except :
print('try select_res error')
y_pred=adata.obs['leiden'].astype(int).to_numpy()
y_pred_last = y_pred
Group=pd.Series(y_pred,index=np.arange(0,x.shape[0]),name="Group")
Mergefeature=pd.concat([pd.DataFrame(x),Group],axis=1) #detach().numpy()
cluster_centers=np.asarray(Mergefeature.groupby("Group").mean())
#print(cluster_centers.shape,cluster_centers)
self.model.dec.data.copy_(torch.Tensor(cluster_centers)).to(self.device)
self.model.train()
for epoch in range(max_epochs):
if epoch%update_interval == 0:
zq, _, _, _, _, _, _ = self.model.forward(self.img_tensor, self.rna_tensor, self.edge_index, atten_model = True)
q = 1.0 / ((1.0 + torch.sum((zq.unsqueeze(1) - self.model.dec)**2, dim=2) / alpha) + 1e-8)
q = q**(alpha+1.0)/2.0
q = q / torch.sum(q, dim=1, keepdim=True)
p = self.target_distribution(q).data
# else :
# p = p
self.optimizer.zero_grad()
z, mean, logvar, x_hat, img_emb,rna_emb, attn_weight = self.model.forward(self.img_tensor, self.rna_tensor, self.edge_index, atten_model = True)
if self.custom_decoder == True:
concat_tensor = torch.cat((self.img_tensor,self.rna_tensor),1)
recon_loss = self.model.recon_loss(x_hat,concat_tensor)
kl_loss = self.model.kl_loss(mean, logvar)
loss = 10*recon_loss + (1 / self.img_tensor.shape[0]) * kl_loss
else:
inner_loss = self.model.innerproduct_loss(z,self.edge_index)
kl_loss = self.model.kl_loss(mean, logvar)
loss = inner_loss + (1 / self.img_tensor.shape[0]) * kl_loss
# total_loss = loss + dec_kl_loss
loss.backward()
self.optimizer.step()
if epoch%10==0:
print("Epoch ", epoch, " loss:",loss)
#Check stop criterion
y_pred = torch.argmax(q, dim=1).data.cpu().numpy()
delta_label = np.sum(y_pred != y_pred_last).astype(np.float32) / x.shape[0]
y_pred_last = y_pred
if epoch>0 and (epoch-1)%update_interval == 0 and delta_label < tol:
print('delta_label ', delta_label, '< tol ', tol)
print("Reach tolerance threshold. Stopping training.")
print("Total epoch:", epoch)
break
return loss_list, img_emb, rna_emb, attn_weight
def target_distribution(self, q):
#weight = q ** 2 / q.sum(0)
#return torch.transpose((torch.transpose(weight,0,1) / weight.sum(1)),0,1)e
p = q**2 / torch.sum(q, dim=0)
p = p / torch.sum(p, dim=1, keepdim=True)
return p
def save_attn_weight(self, epoch, attn_weight, inter=10, out_dir='./'):
import pickle
"""Save all the networks to the disk.
Parameters:
epoch (int) -- current epoch; used in the file name '%s_net_%s.pth' % (epoch, name)
"""
if epoch % inter == 0:
save_filename = 'epoch_%s_att_weight.pkl' % (epoch)
with open(os.path.join(out_dir, save_filename), 'wb') as file:
pickle.dump(attn_weight, file)
def predict(self, ):
self.model.eval()
with torch.no_grad():
z, mean, logvar, x_hat, img_emb,rna_emb, attn_weight = self.model(self.img_tensor, self.rna_tensor, self.edge_index)
emb = z.to('cpu').detach().numpy()
return emb
def get_embedding(img_tensor, rna_tensor, edge_index, model):
model.eval()
with torch.no_grad():
z, mean, logvar, x_hat, img_emb,rna_emb, attn_weight = model(img_tensor, rna_tensor, edge_index)
vae_emb = z.to('cpu').detach().numpy()
return vae_emb