-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
207 lines (176 loc) · 8.77 KB
/
model.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
# !/usr/bin/env python
# -*- coding: utf8 -*-
import torch,sys
import torch_scatter
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from utils import set_device,sparse_to_tuple,normalize_adj
from modules import MultiVGAE,ConstraintMatching,ConstraintBinning,ConstraintProcessing
import scipy.sparse as sp
from collections import Counter,defaultdict
import gc
gc.collect()
torch.cuda.empty_cache()
class UnitigBIN:
def __init__(self, ipt_dim, args):
self.args = args
self.ipt_dim = ipt_dim
self.device = set_device()
self.model = MultiVGAE(ipt_dim, args.hid_dim*2, args.hid_dim, args, self.device)
self.model.to(self.device)
def PPRDiffusionConv_Torch(self, A, alpha, eps):
# print('--- PPR-based Diffusion Convolution (Torch) ---')
N = A.shape[0]
A_loop = torch.tensor(torch.eye(N)+A,dtype=torch.float).to(self.device) ## Self-loops
D_loop_vec = torch.sum(A_loop, 0)
D_loop_vec_invsqrt = 1 / torch.sqrt(D_loop_vec)
D_loop_invsqrt = torch.diag(D_loop_vec_invsqrt)
DA = torch.matmul(D_loop_invsqrt, A_loop)
T_sym = torch.matmul(DA, D_loop_invsqrt)
I = torch.eye(N).to(self.device)
S = alpha * torch.linalg.inv(I - (1 - alpha) * T_sym)
S_tilde = torch.multiply(S, torch.ge(S, eps))
D_tilde_vec = torch.sum(S_tilde, 0)
T_S = S_tilde / D_tilde_vec
return T_S.to_sparse()
def learning(self, data):
# print("# Params:", sum(p.numel() for p in self.model.parameters() if p.requires_grad))
V = torch.tensor(data.edge_index[0]).to(self.device)
E = torch.tensor(data.edge_index[1]).to(self.device)
R = torch.tensor(data.batchReads[0]).to(self.device)
I = torch.tensor(data.batchReads[1]).to(self.device)
triplets = data.triplets
batches = data.batches #[readList,readAdj,readFeat]
pairs = torch.tensor(data.readsPair).to(self.device)
adj_norm,adj_labels,features = [],[],[]
adj_convs = []
norms,weight_tensors = [],[]
for idx,batch in batches.items(): #batch:[readList,readAdj,readFeat,readAdjFull]
diffConv = self.PPRDiffusionConv_Torch(batch[1].toarray(), self.args.alpha, self.args.eps)
adj_norm.append(diffConv) #diffusion
adj = sparse_to_tuple(batch[1])#sparse_to_tuple(sp.csr_matrix(batch[1]))
adjTensor = torch.sparse.FloatTensor(torch.LongTensor(adj[0].T),torch.FloatTensor(adj[1]),torch.Size(adj[2])).to(self.device)
adj_labels.append(adjTensor) #adjacency
feat = sparse_to_tuple(sp.csr_matrix(np.concatenate((batch[1].toarray(),batch[2].toarray()),axis=1)))
featTensor = torch.sparse.FloatTensor(torch.LongTensor(feat[0].T),torch.FloatTensor(feat[1]),torch.Size(feat[2])).to(self.device)
features.append(featTensor) #features
pos_weight = float(batch[1].shape[0]*batch[1].shape[0]-batch[1].sum())/batch[1].sum()
weight_mask = adjTensor.to_dense().view(-1) == 1
weight_tensor = torch.ones(weight_mask.size(0)).to(self.device)
weight_tensor[weight_mask] = pos_weight
weight_tensors.append(weight_tensor)
norm = batch[1].shape[0]*batch[1].shape[0]/float((batch[1].shape[0]*batch[1].shape[0]-batch[1].sum())*2)
norms.append(norm)
optimizer = torch.optim.Adam(self.model.parameters(), lr=self.args.lr, weight_decay=self.args.weight_decay)
print("\n### Learning: Representing Unitig-level Assembly Graph with Constraints.")
cnt_wait,best = 0,1e9
for epoch in range(self.args.epochs):
# Training loop
self.model.train()
optimizer.zero_grad()
adj_recons,means,logstds,meansAll,logstdsAll,meansRead,logstdsRead = self.model.forward(adj_norm, features, (R,I))
rec_loss,kl_div = 0,0
for i in range(self.args.nbatchGraph):
rec_loss += norm * F.binary_cross_entropy(adj_recons[i].view(-1), adj_labels[i].to_dense().view(-1), weight=weight_tensors[i])
kl_div += -0.5/adj_recons[i].size(0)*(1+2*logstds[i]-means[i]**2-torch.exp(logstds[i])**2).sum(1).mean()
rec_loss = rec_loss/self.args.nbatchGraph
kl_div = kl_div/self.args.nbatchGraph
rnd = np.unique(np.random.randint(0, triplets.shape[0], self.args.nbatchConst))
Tensor_triplets = torch.tensor(triplets[rnd]).to(self.device)
con_loss = 10*self.model.triplet_loss(meansRead,F.elu(logstdsRead)+1+1e-14,Tensor_triplets)
loss = rec_loss + kl_div
loss += self.args.lambda1*con_loss
if self.args.nbatchGraph != 1:
bat_loss = 10*self.model.batch_loss(meansAll,F.elu(logstdsAll)+1+1e-14,pairs)
loss += bat_loss
else:
bat_loss = torch.tensor(0.0)
if (epoch+1)%20==0 or epoch == 0:
print("Epoch:", '%04d' % (epoch+1), "loss=", "{:.5f}".format(loss.item()), "Rec_loss=", "{:.5f}".format(rec_loss.item()),
"KL_Div=", "{:.5f}".format(kl_div.item()), "Batch_loss=", "{:.5f}".format(bat_loss.item()), "Con_loss=", "{:.5f}".format(con_loss.item()))
if loss < best:
cnt_wait,best = 0,loss
torch.save(self.model.state_dict(), 'best_learning_'+str(self.args.runs)+'.pkl')
else:
cnt_wait += 1
if cnt_wait == self.args.patience:
print('Early stopping!')
break
loss.backward()
optimizer.step()
self.model.load_state_dict(torch.load('best_learning_'+str(self.args.runs)+'.pkl'))
self.model.eval()
_,_,_,_,_,meansRead,logstdsRead = self.model.forward(adj_norm, features, (R,I))
_embed = torch_scatter.scatter(meansRead[..., V, :], E, dim=-2, reduce='mean')
embed = _embed.cpu().detach().numpy()
composition_contig = data.featContig
print("\n### Matching: an adapted matching algorithm to initialize markered contigs.")
constraints = data.constraints
self.Matching = ConstraintMatching(self.args)
_embed = {i:emb for i,emb in enumerate(composition_contig)}
initBins = self.Matching.Estimation(constraints, _embed, THRESHOLD=self.args.threshold)
pred_bins_matching = dict()
for binid,contigs in initBins.items():
for contig in contigs:
pred_bins_matching[contig] = binid
self.Processing = ConstraintProcessing(data.constraints,data.neg_contigs,data.contigMap,self.args)
contigsBins_spliting,binsContigs_spliting = self.Processing.Spliting(pred_bins_matching)
contigsBins_merging,binsContigs_merging = self.Processing.Merging(binsContigs_spliting)
pred_bins_matching = contigsBins_merging #contigsBins_merging
print("\n### Propagating: annotate unmarked contigs while satisfying constraints.")
init_labels_contigs = pred_bins_matching
n_initBins = len(init_labels_contigs)
initalBins = {key:val for key,val in init_labels_contigs.items()} #contig-level
inits = [val for key,val in init_labels_contigs.items()]
# mask (contig-level)
idxs = [idx for idx,val in initalBins.items()]
mask = np.array([True if idx in idxs else False for idx in range(data.n_contig)])
init_labels = [initalBins[idx] if idx in idxs else 0 for idx in range(data.n_contig)]
init_labels = torch.LongTensor(init_labels).to(self.device)
mask = torch.LongTensor(mask).to(self.device)
mask = mask.float()
mask = mask / mask.mean()
adj = sparse_to_tuple(normalize_adj(data.readADJ))
adj_conv = torch.sparse.FloatTensor(torch.LongTensor(adj[0].T),torch.FloatTensor(adj[1]),torch.Size(adj[2])).to(self.device)
self.binner = ConstraintBinning(self.args.hid_dim,self.args.hid_dim,n_initBins,self.args).to(self.device)
# print(self.binner)
# print("# Params:", sum(p.numel() for p in self.binner.parameters() if p.requires_grad))
conflict = torch.Tensor(np.ones(n_initBins)-np.eye(n_initBins)).to(self.device)
neg_contigs = torch.LongTensor(data.neg_contigs).to(self.device) #negative = data.neg_contigs
opt = torch.optim.Adam(self.binner.parameters(), lr=0.001, weight_decay=self.args.weight_decay)
cnt_wait,best,loss_last = 0,1e9,1e9
min_violating = 1e9
min_preds = {}
patience = self.args.patience
for epoch in range(self.args.epochs):
self.binner.train()
opt.zero_grad()
out,_ = self.binner(adj_conv, meansRead, (V,E))
loss = F.cross_entropy(out, init_labels, reduction='none')
loss *= mask
loss = loss.mean()
loss_n = 10*self.binner.constraintloss(out, neg_contigs, conflict)
loss += self.args.lambda2*loss_n
if (epoch+1)%100==0 or epoch == 0:
print("Epoch:", '%04d' % (epoch+1), "loss=", "{:.5f}".format(loss.item()), "loss_c=", "{:.5f}".format(loss_n.item()))
if loss < best:
cnt_wait,best = 0,loss
torch.save(self.binner.state_dict(), 'best_binner_LP_'+str(self.args.runs)+'.pkl')
else:
cnt_wait += 1
if cnt_wait == patience:
print('Early stopping!')
break
loss.backward(retain_graph=True)
opt.step()
self.binner.load_state_dict(torch.load('best_binner_LP_'+str(self.args.runs)+'.pkl'))
self.binner.eval()
out,_emb = self.binner(adj_conv, meansRead, (V,E))
pred = out.argmax(dim=1)
pred_labels = {i:j.item() for i,j in enumerate(pred)} #contig
print("\n### Refining: fine-tune contigs binning assignments.")
contigsBins_spliting,binsContigs_spliting = self.Processing.Spliting(pred_labels)
contigsBins_merging,binsContigs_merging = self.Processing.Merging(binsContigs_spliting)
pred_labels = contigsBins_merging
return pred_labels