def eval_omics_net(x_tr, age_tr, y_tr, delta_tr, \
x_va, age_va, y_va, delta_va, \
x_te, age_te, y_te, delta_te, \
gene_indices, pathway_indices, \
in_nodes, gene_nodes, pathway_nodes, hidden_nodes, \
LR, L2, max_epochs, dropout_rate, step = 100, tolerance = 0.02, sparse_coding = False):
net = omics_net(in_nodes, gene_nodes, pathway_nodes, hidden_nodes)
###if gpu is being used
if torch.cuda.is_available():
net = net.cuda()
###optimizer
opt = optim.Adam(net.parameters(), lr=LR, weight_decay = L2)
prev_sum = 0.0
for epoch in range(max_epochs):
net.train()
###reset gradients to zeros
opt.zero_grad()
###Randomize dropout masks
net.do_m1 = dropout_mask(pathway_nodes, dropout_rate[0])
net.do_m2 = dropout_mask(hidden_nodes[0], dropout_rate[1])
###Forward
pred = net(x_tr, age_tr, gene_indices, pathway_indices, dropout_rate)
###calculate loss
loss = neg_par_log_likelihood(pred, y_tr, delta_tr)
###calculate gradients
loss.backward()
###force the connections between omics layer and gene layer w.r.t. 'gene_mask'
net.omics.weight.grad = fixed_s_mask(net.omics.weight.grad, gene_indices)
###force the connections between gene layer and pathway layer w.r.t. 'pathway_mask'
net.gene.weight.grad = fixed_s_mask(net.gene.weight.grad, pathway_indices)
###update weights and biases
opt.step()
if sparse_coding == True:
net = sparse_func(net, x_tr, age_tr, y_tr, delta_tr, gene_indices, pathway_indices, dropout_rate)
if epoch % step == step - 1:
net.train()
pred = net(x_tr, age_tr, gene_indices, pathway_indices, dropout_rate)
train_cindex = c_index(pred.cpu(), y_tr.cpu(), delta_tr.cpu())
net.eval()
pred = net(x_va, age_va, gene_indices, pathway_indices, dropout_rate)
eval_cindex = c_index(pred.cpu(), y_va.cpu(), delta_va.cpu())
if ((eval_cindex.item() + train_cindex.item() + tolerance) < prev_sum):
print('Early stopping in [%d]' % (epoch + 1))
print('[%d] Best CIndex in Train: %.3f' % (epoch + 1, opt_cidx_tr))
print('[%d] Best CIndex in Valid: %.3f' % (epoch + 1, opt_cidx_va))
opt_net.eval()
pred = opt_net(x_te, age_te, gene_indices, pathway_indices, dropout_rate)
eval_cindex = c_index(pred.cpu(), y_te.cpu(), delta_te.cpu())
break
else:
prev_sum = eval_cindex.item() + train_cindex.item()
opt_cidx_tr = train_cindex
opt_cidx_va = eval_cindex
opt_net = copy.deepcopy(net)
print('[%d] CIndex in Train: %.3f' % (epoch + 1, train_cindex))
print('[%d] CIndex in Valid: %.3f' % (epoch + 1, eval_cindex))
return (opt_cidx_tr, opt_cidx_va, eval_cindex)
def forward(self, x_1, x_2, x_3, pathway_idx, Drop_Rate):
###force the connections between gene layer and pathway layer w.r.t. 'pathway_mask'
self.gene.weight.data = fixed_s_mask(self.gene.weight.data, pathway_idx)
x_1 = self.tanh(self.bn1(self.gene(x_1)))
if self.training == True:
x_1 = (1/(1-Drop_Rate[0])) * x_1.mul(self.do_m1)
x_1 = self.tanh(self.bn2(self.pathway(x_1)))
if self.training == True:
x_1 = (1 / (1 - Drop_Rate[1])) * x_1.mul(self.do_m2)
x_1 = self.tanh(self.bn3(self.hidden(x_1)))
if self.training == True:
x_3 = (1 / (1 - Drop_Rate[2])) * x_3.mul(self.do_m4)
x_3 = self.tanh(self.bn4(self.image(x_3)))
###integration
x_cat = torch.cat((x_1, x_2, x_3), 1)
lin_pred = self.integrative(x_cat)
return lin_pred
def forward(self, x_1, x_2, gene_idx, pathway_idx, Drop_Rate):
###force the connections between omics layer and gene layer w.r.t. 'gene_mask'
self.omics.weight.data = fixed_s_mask(self.omics.weight.data, gene_idx)
###force the connections between gene layer and pathway layer w.r.t. 'pathway_mask'
self.gene.weight.data = fixed_s_mask(self.gene.weight.data, pathway_idx)
# bath norm beofre activation
x_1 = self.relu(self.bn1(self.omics(x_1)))
x_1 = self.relu(self.bn2(self.gene(x_1)))
if self.training == True:
# inverted dropout
x_1 = (1/(1-Drop_Rate[0])) * x_1.mul(self.do_m1)
x_1 = self.relu(self.bn3(self.pathway(x_1)))
if self.training == True:
x_1 = (1 / (1 - Drop_Rate[1])) * x_1.mul(self.do_m2)
x_1 = self.relu(self.bn4(self.hidden(x_1)))
###add age
x_cat = torch.cat((x_1, x_2), 1)
lin_pred = self.hidden2(x_cat)
return lin_pred
def train_model(x_tr, age_tr, pt_tr, y_tr, delta_tr, \
x_va, age_va, pt_va, y_va, delta_va, \
x_te, age_te, pt_te, y_te, delta_te, \
pathway_indices, \
gene_nodes, pathway_nodes, image_nodes, hidden_nodes, \
lr, l2, max_epochs, dropout_rate, step = 100, tolerance = 0.02, sparse_coding = False, test_phrase = False):
net = cox_pasnet_pathology(gene_nodes, pathway_nodes, image_nodes,
hidden_nodes)
net = net.cuda()
###optimizer
opt = optim.Adam(net.parameters(), lr=lr, weight_decay=l2)
prev_sum = 0.0
temp_loss_list = []
for epoch in range(max_epochs):
torch.cuda.empty_cache()
net.train()
opt.zero_grad() ###reset gradients to zeros
###Randomize dropout masks
net.do_m1 = dropout_mask(pathway_nodes, dropout_rate[0])
net.do_m2 = dropout_mask(hidden_nodes[0], dropout_rate[1])
net.do_m4 = dropout_mask(image_nodes, dropout_rate[2])
pred = net(x_tr, age_tr, pt_tr, pathway_indices,
dropout_rate) ###Forward
loss = neg_par_log_likelihood(pred, y_tr, delta_tr) ###calculate loss
loss.backward() ###calculate gradients
###force the connections between gene layer and pathway layer w.r.t. 'pathway_mask'
net.gene.weight.grad = fixed_s_mask(net.gene.weight.grad,
pathway_indices)
opt.step() ###update weights and biases
if sparse_coding == True:
net = sparse_func(net, x_tr, age_tr, pt_tr, y_tr, delta_tr,
pathway_indices, dropout_rate)
torch.cuda.empty_cache()
net.eval()
valid_pred = net(x_va, age_va, pt_va, pathway_indices, dropout_rate)
valid_loss = neg_par_log_likelihood(valid_pred, y_va,
delta_va).detach().cpu()
temp_loss_list.append(valid_loss.detach().cpu())
if (epoch % step == step - 1) and (epoch + 1 >= 500):
torch.cuda.empty_cache()
print("Current LR: ", lr)
opt_temp_loss = np.min(temp_loss_list)
gl = valid_loss / opt_temp_loss - 1.0
net.train()
pred = net(x_tr, age_tr, pt_tr, pathway_indices, dropout_rate)
train_cindex = c_index(pred.cpu(), y_tr.cpu(), delta_tr.cpu())
valid_cindex = c_index(valid_pred.cpu(), y_va.cpu(),
delta_va.cpu())
del pred
if (gl > tolerance):
if epoch + 1 == 500:
opt_cidx_tr = train_cindex
opt_cidx_ev = valid_cindex
opt_net = copy.deepcopy(net)
print('Early stopping in [%d]' % (epoch + 1))
print('[%d] GL: %.4f' % (epoch + 1, gl))
print('[%d] Best CIndex in Train: %.3f' %
(epoch + 1, opt_cidx_tr))
print('[%d] Best CIndex in Valid: %.3f' %
(epoch + 1, opt_cidx_ev))
if (test_phrase == True):
opt_net.eval()
test_pred = opt_net(x_te, age_te, pt_te, pathway_indices,
dropout_rate)
opt_cidx_ev = c_index(test_pred.cpu(), y_te.cpu(),
delta_te.cpu())
print('[%d] Final CIndex in Test: %.3f' %
(epoch + 1, opt_cidx_ev))
break
else:
opt_cidx_tr = train_cindex
opt_cidx_ev = valid_cindex
opt_net = copy.deepcopy(net)
print('[%d] GL: %.4f' % (epoch + 1, gl))
print('[%d] CIndex in Train: %.3f' % (epoch + 1, train_cindex))
print('[%d] CIndex in Valid: %.3f' % (epoch + 1, valid_cindex))
torch.cuda.empty_cache()
lr = lr * 0.8
if (test_phrase == True):
del opt_cidx_ev
opt_net.eval()
test_pred = opt_net(x_te, age_te, pt_te, pathway_indices,
dropout_rate)
opt_cidx_ev = c_index(test_pred.cpu(), y_te.cpu(),
delta_te.cpu())
print('[%d] Final CIndex in Test: %.3f' %
(epoch + 1, opt_cidx_ev))
return (opt_cidx_tr, opt_cidx_ev)
def interpret_net(outpath, x, age, pt, y, delta, pathway_indices, \
gene_nodes, pathway_nodes, image_nodes, hidden_nodes, \
lr, l2, max_epochs, dropout_rate, step = 100, tolerance = 0.05, sparse_coding = False):
net = cox_pasnet_pathology(gene_nodes, pathway_nodes, image_nodes,
hidden_nodes)
net = net.cuda()
###optimizer
opt = optim.Adam(net.parameters(), lr=lr, weight_decay=l2)
prev_sum = 0.0
temp_loss_list = []
cidx_tr = []
loss_tr = []
last_loss = np.inf
sp_pathway = []
sp_hidden = []
for epoch in range(max_epochs):
torch.cuda.empty_cache()
net.train()
opt.zero_grad() ###reset gradients to zeros
###Randomize dropout masks
net.do_m1 = dropout_mask(pathway_nodes, dropout_rate[0])
net.do_m2 = dropout_mask(hidden_nodes[0], dropout_rate[1])
net.do_m4 = dropout_mask(image_nodes, dropout_rate[2])
pred = net(x, age, pt, pathway_indices, dropout_rate) ###Forward
loss = neg_par_log_likelihood(pred, y, delta) ###calculate loss
loss.backward() ###calculate gradients
###force the connections between gene layer and pathway layer w.r.t. 'pathway_mask'
net.gene.weight.grad = fixed_s_mask(net.gene.weight.grad,
pathway_indices)
opt.step() ###update weights and biases
if sparse_coding == True:
net = sparse_func(net, x, age, pt, y, delta, pathway_indices,
dropout_rate)
torch.cuda.empty_cache()
if epoch % step == step - 1:
net.eval()
pred = net(x, age, pt, pathway_indices, dropout_rate)
train_loss = neg_par_log_likelihood(pred, y, delta).detach().cpu()
train_cindex = c_index(pred.cpu(), y.cpu(), delta.cpu())
if train_loss.item() - tolerance > last_loss:
break
else:
last_loss = train_loss.item()
opt_net = copy.deepcopy(net)
loss_tr.append(train_loss.item())
cidx_tr.append(train_cindex.item())
sp_pathway.append(net.pathway.weight.nonzero().size(0))
sp_hidden.append(net.hidden.weight.nonzero().size(0))
print('[%d] Loss in Train: %.4f' % (epoch + 1, train_loss))
print('[%d] CIndex in Train: %.3f' % (epoch + 1, train_cindex))
print('Connections between pathway and hidden layer: ',
net.pathway.weight.nonzero().size(0))
print('Connections between hidden and last hidden layer: ',
net.hidden.weight.nonzero().size(0))
print(step)
print(epoch)
x_axis = np.arange(0, epoch + 1, step)
fig = plt.figure()
print(len(loss_tr))
print(len(x_axis))
plt.plot(x_axis, loss_tr)
fig.savefig("cindex_" + str(lr) + "_" + str(l2) + ".png")
np.savetxt("sparsity_pathway.txt", sp_pathway, delimiter=',')
np.savetxt("sparsity_hidden.txt", sp_hidden, delimiter=',')
np.savetxt("cindex_entire_data.txt", cidx_tr, delimiter=',')
torch.save(opt_net.state_dict(), outpath)
return