def main():
# load data
data_path = '../data/output/records_final.pkl'
voc_path = '../data/output/voc_final.pkl'
ddi_adj_path = '../data/output/ddi_A_final.pkl'
device = torch.device('cuda:{}'.format(args.cuda))
ddi_adj = dill.load(open(ddi_adj_path, 'rb'))
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
np.random.seed(1203)
np.random.shuffle(data)
split_point = int(len(data) * 3 / 5)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
print(voc_size)
model = DualNN(voc_size, ddi_adj, emb_dim=args.dim, device=device)
# model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
if args.Test:
model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
model.to(device=device)
tic = time.time()
label_list, prob_add, prob_delete = eval(model, data_eval, voc_size, 0,
1)
threshold1, threshold2 = [], []
for i in range(label_list.shape[1]):
_, _, boundary_add = roc_curve(label_list[:, i],
prob_add[:, i],
pos_label=1)
_, _, boundary_delete = roc_curve(label_list[:, i],
prob_delete[:, i],
pos_label=0)
threshold1.append(boundary_add[min(round(len(boundary_add) * 0.05),
len(boundary_add) - 1)])
threshold2.append(boundary_delete[min(
round(len(boundary_delete) * 0.05),
len(boundary_delete) - 1)])
# threshold1 = np.ones(voc_size[2]) * np.mean(threshold1)
# threshold2 = np.ones(voc_size[2]) * np.mean(threshold2)
print(np.mean(threshold1), np.mean(threshold2))
eval(model, data_test, voc_size, 0, 0, threshold1, threshold2)
print('test time: {}'.format(time.time() - tic))
return
model.to(device=device)
print('parameters', get_n_params(model))
# exit()
optimizer = RMSprop(list(model.parameters()),
lr=args.lr,
weight_decay=args.weight_decay)
# start iterations
history = defaultdict(list)
best_epoch, best_ja = 0, 0
EPOCH = 40
for epoch in range(EPOCH):
t = 0
tic = time.time()
print('\nepoch {} --------------------------'.format(epoch + 1))
model.train()
for step, input in enumerate(data_train):
if len(input) < 2: continue
loss = 0
for adm_idx, adm in enumerate(input):
if adm_idx == 0: continue
seq_input = input[:adm_idx + 1]
loss_bce_target = np.zeros((1, voc_size[2]))
loss_bce_target[:, adm[2]] = 1
loss_bce_target_last = np.zeros((1, voc_size[2]))
loss_bce_target_last[:, input[adm_idx - 1][2]] = 1
add_target = np.zeros((1, voc_size[2]))
add_target[:, np.where(loss_bce_target == 1)[1]] = 1
delete_target = np.zeros((1, voc_size[2]))
delete_target[:, np.where(loss_bce_target == 0)[1]] = 1
loss_multi_target = np.full((1, voc_size[2]), -1)
for idx, item in enumerate(adm[2]):
loss_multi_target[0][idx] = item
loss_multi_target_last = np.full((1, voc_size[2]), -1)
for idx, item in enumerate(input[adm_idx - 1][2]):
loss_multi_target_last[0][idx] = item
loss_multi_add_target = np.full((1, voc_size[2]), -1)
for i, item in enumerate(np.where(add_target == 1)[0]):
loss_multi_add_target[0][i] = item
loss_multi_delete_target = np.full((1, voc_size[2]), -1)
for i, item in enumerate(np.where(delete_target == 1)[0]):
loss_multi_delete_target[0][i] = item
add_result, delete_result = model(seq_input)
loss_bce = F.binary_cross_entropy_with_logits(add_result, torch.FloatTensor(add_target).to(device)) + \
F.binary_cross_entropy_with_logits(delete_result, torch.FloatTensor(delete_target).to(device))
loss_multi = F.multilabel_margin_loss(F.sigmoid(add_result), torch.LongTensor(loss_multi_add_target).to(device)) + \
F.multilabel_margin_loss(F.sigmoid(delete_result), torch.LongTensor(loss_multi_delete_target).to(device))
# l2 = 0
# for p in model.parameters():
# l2 = l2 + (p ** 2).sum()
loss += 0.95 * loss_bce + 0.05 * loss_multi
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
llprint('\rtraining step: {} / {}'.format(step, len(data_train)))
print()
tic2 = time.time()
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, add, delete, avg_med = eval(
model, data_eval, voc_size, epoch)
print('training time: {}, test time: {}'.format(
time.time() - tic,
time.time() - tic2))
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['prauc'].append(prauc)
history['add'].append(add)
history['delete'].append(delete)
history['med'].append(avg_med)
if epoch >= 5:
print(
'ddi: {}, Med: {}, Ja: {}, F1: {}, Add: {}, Delete: {}'.format(
np.mean(history['ddi_rate'][-5:]),
np.mean(history['med'][-5:]), np.mean(history['ja'][-5:]),
np.mean(history['avg_f1'][-5:]),
np.mean(history['add'][-5:]),
np.mean(history['delete'][-5:])))
torch.save(model.state_dict(), open(os.path.join('saved', args.model_name, \
'Epoch_{}_JA_{:.4}_DDI_{:.4}.model'.format(epoch, ja, ddi_rate)), 'wb'))
if epoch != 0 and best_ja < ja:
best_epoch = epoch
best_ja = ja
print('best_epoch: {}'.format(best_epoch))
dill.dump(
history,
open(
os.path.join('saved', args.model_name,
'history_{}.pkl'.format(args.model_name)), 'wb'))
def main():
# load data
data_path = '../data/output/records_final.pkl'
voc_path = '../data/output/voc_final.pkl'
ddi_adj_path = '../data/output/ddi_A_final.pkl'
device = torch.device('cuda:{}'.format(args.cuda))
ddi_adj = dill.load(open(ddi_adj_path, 'rb'))
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
np.random.seed(1203)
np.random.shuffle(data)
split_point = int(len(data) * 3 / 5)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
print(voc_size)
model = SimNN(voc_size, ddi_adj, emb_dim=args.dim, device=device)
# model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
if args.Test:
model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
model.to(device=device)
tic = time.time()
eval(model, data_test, voc_size, 0)
print('test time: {}'.format(time.time() - tic))
return
model.to(device=device)
print('parameters', get_n_params(model))
# exit()
optimizer = RMSprop(list(model.parameters()),
lr=args.lr,
weight_decay=args.weight_decay)
# start iterations
history = defaultdict(list)
best_epoch, best_ja = 0, 0
criterion = torch.nn.CrossEntropyLoss()
EPOCH = 40
for epoch in range(EPOCH):
t = 0
tic = time.time()
print('\nepoch {} --------------------------'.format(epoch + 1))
model.train()
for step, input in enumerate(data_train):
if len(input) < 2: continue
loss = 0
for adm_idx, adm in enumerate(input):
if adm_idx == 0: continue
seq_input = input[:adm_idx + 1]
loss_bce_target = np.zeros((1, voc_size[2]))
loss_bce_target[:, adm[2]] = 1
loss_bce_target_last = np.zeros((1, voc_size[2]))
loss_bce_target_last[:, input[adm_idx - 1][2]] = 1
target_list = loss_bce_target - loss_bce_target_last
target_list[target_list == -1] = 2
result = model(seq_input)
loss += criterion(result,
torch.LongTensor(target_list[0]).to(device))
# l2 = 0
# for p in model.parameters():
# l2 = l2 + (p ** 2).sum()
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
llprint('\rtraining step: {} / {}'.format(step, len(data_train)))
print()
tic2 = time.time()
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, add, delete, avg_med = eval(
model, data_eval, voc_size, epoch)
print('training time: {}, test time: {}'.format(
time.time() - tic,
time.time() - tic2))
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['prauc'].append(prauc)
history['add'].append(add)
history['delete'].append(delete)
history['med'].append(avg_med)
if epoch >= 5:
print(
'ddi: {}, Med: {}, Ja: {}, F1: {}, Add: {}, Delete: {}'.format(
np.mean(history['ddi_rate'][-5:]),
np.mean(history['med'][-5:]), np.mean(history['ja'][-5:]),
np.mean(history['avg_f1'][-5:]),
np.mean(history['add'][-5:]),
np.mean(history['delete'][-5:])))
torch.save(model.state_dict(), open(os.path.join('saved', args.model_name, \
'Epoch_{}_JA_{:.4}_DDI_{:.4}.model'.format(epoch, ja, ddi_rate)), 'wb'))
if epoch != 0 and best_ja < ja:
best_epoch = epoch
best_ja = ja
print('best_epoch: {}'.format(best_epoch))
dill.dump(
history,
open(
os.path.join('saved', args.model_name,
'history_{}.pkl'.format(args.model_name)), 'wb'))
def main():
# load data
data_path = '../data/output/records_final.pkl'
voc_path = '../data/output/voc_final.pkl'
device = torch.device('cuda:{}'.format(args.cuda))
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
split_point = int(len(data) * 2 / 3)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
END_TOKEN = voc_size[2] + 1
model = Leap(voc_size, device=device)
# model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
if args.Test:
model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
model.to(device=device)
tic = time.time()
result = []
for _ in range(10):
test_sample = np.random.choice(data_test,
round(len(data_test) * 0.8),
replace=True)
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, avg_med = eval(
model, test_sample, voc_size, 0)
result.append([ddi_rate, ja, avg_f1, prauc, avg_med])
result = np.array(result)
mean = result.mean(axis=0)
std = result.std(axis=0)
outstring = ""
for m, s in zip(mean, std):
outstring += "{:.4f} $\pm$ {:.4f} & ".format(m, s)
print(outstring)
print('test time: {}'.format(time.time() - tic))
return
model.to(device=device)
print('parameters', get_n_params(model))
optimizer = Adam(model.parameters(), lr=args.lr)
history = defaultdict(list)
best_epoch, best_ja = 0, 0
EPOCH = 50
for epoch in range(EPOCH):
tic = time.time()
print('\nepoch {} --------------------------'.format(epoch + 1))
model.train()
for step, input in enumerate(data_train):
for adm in input:
loss_target = adm[2] + [END_TOKEN]
output_logits = model(adm)
loss = F.cross_entropy(
output_logits,
torch.LongTensor(loss_target).to(device))
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
llprint('\rtraining step: {} / {}'.format(step, len(data_train)))
print()
tic2 = time.time()
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, avg_med = eval(
model, data_eval, voc_size, epoch)
print('training time: {}, test time: {}'.format(
time.time() - tic,
time.time() - tic2))
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['prauc'].append(prauc)
history['med'].append(avg_med)
if epoch >= 5:
print('ddi: {}, Med: {}, Ja: {}, F1: {}, PRAUC: {}'.format(
np.mean(history['ddi_rate'][-5:]),
np.mean(history['med'][-5:]), np.mean(history['ja'][-5:]),
np.mean(history['avg_f1'][-5:]),
np.mean(history['prauc'][-5:])))
torch.save(model.state_dict(), open(os.path.join('saved', args.model_name, \
'Epoch_{}_JA_{:.4}_DDI_{:.4}.model'.format(epoch, ja, ddi_rate)), 'wb'))
if epoch != 0 and best_ja < ja:
best_epoch = epoch
best_ja = ja
print('best_epoch: {}'.format(best_epoch))
dill.dump(
history,
open(
os.path.join('saved', args.model_name,
'history_{}.pkl'.format(args.model_name)), 'wb'))
def main():
if not os.path.exists(os.path.join("saved", model_name)):
os.makedirs(os.path.join("saved", model_name))
data_path = '../data/records_final.pkl'
voc_path = '../data/voc_final.pkl'
ehr_adj_path = '../data/ehr_adj_final.pkl'
ddi_adj_path = '../data/ddi_A_final.pkl'
device = torch.device('cuda:0')
ehr_adj = dill.load(open(ehr_adj_path, 'rb'))
ddi_adj = dill.load(open(ddi_adj_path, 'rb'))
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
split_point = int(len(data) * 2 / 3)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
EPOCH = 40
LR = 0.0002
TEST = args.eval
Neg_Loss = args.ddi
DDI_IN_MEM = args.ddi
TARGET_DDI = 0.05
T = 0.5
decay_weight = 0.85
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
model = GAMENet(voc_size,
ehr_adj,
ddi_adj,
emb_dim=64,
device=device,
ddi_in_memory=DDI_IN_MEM)
if TEST:
model.load_state_dict(torch.load(open(resume_name, 'rb')))
model.to(device=device)
print('parameters', get_n_params(model))
optimizer = Adam(list(model.parameters()), lr=LR)
if TEST:
eval(model, data_test, voc_size, 0)
else:
history = defaultdict(list)
best_epoch = 0
best_ja = 0
for epoch in range(EPOCH):
loss_record1 = []
start_time = time.time()
model.train()
prediction_loss_cnt = 0
neg_loss_cnt = 0
for step, input in enumerate(data_train):
for idx, adm in enumerate(input):
seq_input = input[:idx + 1]
loss1_target = np.zeros((1, voc_size[2]))
loss1_target[:, adm[2]] = 1
loss3_target = np.full((1, voc_size[2]), -1)
for idx, item in enumerate(adm[2]):
loss3_target[0][idx] = item
target_output1, batch_neg_loss = model(seq_input)
loss1 = F.binary_cross_entropy_with_logits(
target_output1,
torch.FloatTensor(loss1_target).to(device))
loss3 = F.multilabel_margin_loss(
F.sigmoid(target_output1),
torch.LongTensor(loss3_target).to(device))
if Neg_Loss:
target_output1 = F.sigmoid(
target_output1).detach().cpu().numpy()[0]
target_output1[target_output1 >= 0.5] = 1
target_output1[target_output1 < 0.5] = 0
y_label = np.where(target_output1 == 1)[0]
current_ddi_rate = ddi_rate_score([[y_label]])
if current_ddi_rate <= TARGET_DDI:
loss = 0.9 * loss1 + 0.01 * loss3
prediction_loss_cnt += 1
else:
rnd = np.exp((TARGET_DDI - current_ddi_rate) / T)
if np.random.rand(1) < rnd:
loss = batch_neg_loss
neg_loss_cnt += 1
else:
loss = 0.9 * loss1 + 0.01 * loss3
prediction_loss_cnt += 1
else:
loss = 0.9 * loss1 + 0.01 * loss3
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
loss_record1.append(loss.item())
llprint(
'\rTrain--Epoch: %d, Step: %d/%d, L_p cnt: %d, L_neg cnt: %d'
% (epoch, step, len(data_train), prediction_loss_cnt,
neg_loss_cnt))
# annealing
T *= decay_weight
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1 = eval(
model, data_eval, voc_size, epoch)
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['prauc'].append(prauc)
end_time = time.time()
elapsed_time = (end_time - start_time) / 60
llprint(
'\tEpoch: %d, Loss: %.4f, One Epoch Time: %.2fm, Appro Left Time: %.2fh\n'
% (epoch, np.mean(loss_record1), elapsed_time, elapsed_time *
(EPOCH - epoch - 1) / 60))
torch.save(
model.state_dict(),
open(
os.path.join(
'saved', model_name,
'Epoch_%d_JA_%.4f_DDI_%.4f.model' %
(epoch, ja, ddi_rate)), 'wb'))
print('')
if epoch != 0 and best_ja < ja:
best_epoch = epoch
best_ja = ja
dill.dump(history,
open(os.path.join('saved', model_name, 'history.pkl'), 'wb'))
# test
torch.save(
model.state_dict(),
open(os.path.join('saved', model_name, 'final.model'), 'wb'))
print('best_epoch:', best_epoch)
def main():
data_path = '../data/output/records_final.pkl'
voc_path = '../data/output/voc_final.pkl'
ehr_adj_path = '../data/output/ehr_adj_final.pkl'
ddi_adj_path = '../data/output/ddi_A_final.pkl'
device = torch.device('cuda:{}'.format(args.cuda))
ehr_adj = dill.load(open(ehr_adj_path, 'rb'))
ddi_adj = dill.load(open(ddi_adj_path, 'rb'))
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
# np.random.seed(2048)
# np.random.shuffle(data)
split_point = int(len(data) * 2 / 3)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
model = GAMENet(voc_size,
ehr_adj,
ddi_adj,
emb_dim=args.dim,
device=device,
ddi_in_memory=args.ddi)
# model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
if args.Test:
model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
model.to(device=device)
tic = time.time()
result = []
for _ in range(10):
test_sample = np.random.choice(data_test,
round(len(data_test) * 0.8),
replace=True)
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, avg_med = eval(
model, test_sample, voc_size, 0)
result.append([ddi_rate, ja, avg_f1, prauc, avg_med])
result = np.array(result)
mean = result.mean(axis=0)
std = result.std(axis=0)
outstring = ""
for m, s in zip(mean, std):
outstring += "{:.4f} $\pm$ {:.4f} & ".format(m, s)
print(outstring)
print('test time: {}'.format(time.time() - tic))
return
model.to(device=device)
print('parameters', get_n_params(model))
optimizer = Adam(list(model.parameters()), lr=args.lr)
history = defaultdict(list)
best_epoch, best_ja = 0, 0
EPOCH = 50
for epoch in range(EPOCH):
tic = time.time()
print('\nepoch {} --------------------------'.format(epoch + 1))
prediction_loss_cnt, neg_loss_cnt = 0, 0
model.train()
for step, input in enumerate(data_train):
for idx, adm in enumerate(input):
seq_input = input[:idx + 1]
loss_bce_target = np.zeros((1, voc_size[2]))
loss_bce_target[:, adm[2]] = 1
loss_multi_target = np.full((1, voc_size[2]), -1)
for idx, item in enumerate(adm[2]):
loss_multi_target[0][idx] = item
target_output1, loss_ddi = model(seq_input)
loss_bce = F.binary_cross_entropy_with_logits(
target_output1,
torch.FloatTensor(loss_bce_target).to(device))
loss_multi = F.multilabel_margin_loss(
F.sigmoid(target_output1),
torch.LongTensor(loss_multi_target).to(device))
if args.ddi:
target_output1 = F.sigmoid(
target_output1).detach().cpu().numpy()[0]
target_output1[target_output1 >= 0.5] = 1
target_output1[target_output1 < 0.5] = 0
y_label = np.where(target_output1 == 1)[0]
current_ddi_rate = ddi_rate_score(
[[y_label]], path='../data/output/ddi_A_final.pkl')
if current_ddi_rate <= args.target_ddi:
loss = 0.9 * loss_bce + 0.1 * loss_multi
prediction_loss_cnt += 1
else:
rnd = np.exp(
(args.target_ddi - current_ddi_rate) / args.T)
if np.random.rand(1) < rnd:
loss = loss_ddi
neg_loss_cnt += 1
else:
loss = 0.9 * loss_bce + 0.1 * loss_multi
prediction_loss_cnt += 1
else:
loss = 0.9 * loss_bce + 0.1 * loss_multi
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
llprint('\rtraining step: {} / {}'.format(step, len(data_train)))
args.T *= args.decay_weight
print()
tic2 = time.time()
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, avg_med = eval(
model, data_eval, voc_size, epoch)
print('training time: {}, test time: {}'.format(
time.time() - tic,
time.time() - tic2))
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['prauc'].append(prauc)
history['med'].append(avg_med)
if epoch >= 5:
print('ddi: {}, Med: {}, Ja: {}, F1: {}, PRAUC: {}'.format(
np.mean(history['ddi_rate'][-5:]),
np.mean(history['med'][-5:]), np.mean(history['ja'][-5:]),
np.mean(history['avg_f1'][-5:]),
np.mean(history['prauc'][-5:])))
torch.save(model.state_dict(), open(os.path.join('saved', args.model_name, \
'Epoch_{}_JA_{:.4}_DDI_{:.4}.model'.format(epoch, ja, ddi_rate)), 'wb'))
if epoch != 0 and best_ja < ja:
best_epoch = epoch
best_ja = ja
print('best_epoch: {}'.format(best_epoch))
dill.dump(
history,
open(
os.path.join('saved', args.model_name,
'history_{}.pkl'.format(args.model_name)), 'wb'))
def main():
# load data
data_path = '../data/output/records_final.pkl'
voc_path = '../data/output/voc_final.pkl'
ddi_adj_path = '../data/output/ddi_A_final.pkl'
device = torch.device('cuda:{}'.format(args.cuda))
ddi_adj = dill.load(open(ddi_adj_path, 'rb'))
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
np.random.seed(1203)
np.random.shuffle(data)
split_point = int(len(data) * 3 / 5)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
model = MICRON(voc_size, ddi_adj, emb_dim=args.dim, device=device)
# model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
if args.Test:
model.load_state_dict(torch.load(open(args.resume_path, 'rb')))
model.to(device=device)
tic = time.time()
label_list, prob_list = eval(model, data_eval, voc_size, 0, 1)
threshold1, threshold2 = [], []
for i in range(label_list.shape[1]):
_, _, boundary = roc_curve(label_list[:, i],
prob_list[:, i],
pos_label=1)
# boundary1 should be in [0.5, 0.9], boundary2 should be in [0.1, 0.5]
threshold1.append(
min(
0.9,
max(0.5, boundary[max(0,
round(len(boundary) * 0.05) - 1)])))
threshold2.append(
max(
0.1,
min(
0.5, boundary[min(round(len(boundary) * 0.95),
len(boundary) - 1)])))
print(np.mean(threshold1), np.mean(threshold2))
threshold1 = np.ones(voc_size[2]) * np.mean(threshold1)
threshold2 = np.ones(voc_size[2]) * np.mean(threshold2)
eval(model, data_test, voc_size, 0, 0, threshold1, threshold2)
print('test time: {}'.format(time.time() - tic))
return
model.to(device=device)
print('parameters', get_n_params(model))
# exit()
optimizer = RMSprop(list(model.parameters()),
lr=args.lr,
weight_decay=args.weight_decay)
# start iterations
history = defaultdict(list)
best_epoch, best_ja = 0, 0
weight_list = [[0.25, 0.25, 0.25, 0.25]]
EPOCH = 40
for epoch in range(EPOCH):
t = 0
tic = time.time()
print('\nepoch {} --------------------------'.format(epoch + 1))
sample_counter = 0
mean_loss = np.array([0, 0, 0, 0])
model.train()
for step, input in enumerate(data_train):
loss = 0
if len(input) < 2: continue
for adm_idx, adm in enumerate(input):
if adm_idx == 0: continue
# sample_counter += 1
seq_input = input[:adm_idx + 1]
loss_bce_target = np.zeros((1, voc_size[2]))
loss_bce_target[:, adm[2]] = 1
loss_bce_target_last = np.zeros((1, voc_size[2]))
loss_bce_target_last[:, input[adm_idx - 1][2]] = 1
loss_multi_target = np.full((1, voc_size[2]), -1)
for idx, item in enumerate(adm[2]):
loss_multi_target[0][idx] = item
loss_multi_target_last = np.full((1, voc_size[2]), -1)
for idx, item in enumerate(input[adm_idx - 1][2]):
loss_multi_target_last[0][idx] = item
result, result_last, _, loss_ddi, loss_rec = model(seq_input)
loss_bce = 0.75 * F.binary_cross_entropy_with_logits(result, torch.FloatTensor(loss_bce_target).to(device)) + \
(1 - 0.75) * F.binary_cross_entropy_with_logits(result_last, torch.FloatTensor(loss_bce_target_last).to(device))
loss_multi = 5e-2 * (0.75 * F.multilabel_margin_loss(F.sigmoid(result), torch.LongTensor(loss_multi_target).to(device)) + \
(1 - 0.75) * F.multilabel_margin_loss(F.sigmoid(result_last), torch.LongTensor(loss_multi_target_last).to(device)))
y_pred_tmp = F.sigmoid(result).detach().cpu().numpy()[0]
y_pred_tmp[y_pred_tmp >= 0.5] = 1
y_pred_tmp[y_pred_tmp < 0.5] = 0
y_label = np.where(y_pred_tmp == 1)[0]
current_ddi_rate = ddi_rate_score(
[[y_label]], path='../data/output/ddi_A_final.pkl')
# l2 = 0
# for p in model.parameters():
# l2 = l2 + (p ** 2).sum()
if sample_counter == 0:
lambda1, lambda2, lambda3, lambda4 = weight_list[-1]
else:
current_loss = np.array([
loss_bce.detach().cpu().numpy(),
loss_multi.detach().cpu().numpy(),
loss_ddi.detach().cpu().numpy(),
loss_rec.detach().cpu().numpy()
])
current_ratio = (current_loss -
np.array(mean_loss)) / np.array(mean_loss)
instant_weight = np.exp(current_ratio) / sum(
np.exp(current_ratio))
lambda1, lambda2, lambda3, lambda4 = instant_weight * 0.75 + np.array(
weight_list[-1]) * 0.25
# update weight_list
weight_list.append([lambda1, lambda2, lambda3, lambda4])
# update mean_loss
mean_loss = (mean_loss * (sample_counter - 1) + np.array([loss_bce.detach().cpu().numpy(), \
loss_multi.detach().cpu().numpy(), loss_ddi.detach().cpu().numpy(), loss_rec.detach().cpu().numpy()])) / sample_counter
# lambda1, lambda2, lambda3, lambda4 = weight_list[-1]
if current_ddi_rate > 0.08:
loss += lambda1 * loss_bce + lambda2 * loss_multi + \
lambda3 * loss_ddi + lambda4 * loss_rec
else:
loss += lambda1 * loss_bce + lambda2 * loss_multi + \
lambda4 * loss_rec
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
llprint('\rtraining step: {} / {}'.format(step, len(data_train)))
tic2 = time.time()
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, add, delete, avg_med = eval(
model, data_eval, voc_size, epoch)
print('training time: {}, test time: {}'.format(
time.time() - tic,
time.time() - tic2))
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['prauc'].append(prauc)
history['add'].append(add)
history['delete'].append(delete)
history['med'].append(avg_med)
if epoch >= 5:
print(
'ddi: {}, Med: {}, Ja: {}, F1: {}, Add: {}, Delete: {}'.format(
np.mean(history['ddi_rate'][-5:]),
np.mean(history['med'][-5:]), np.mean(history['ja'][-5:]),
np.mean(history['avg_f1'][-5:]),
np.mean(history['add'][-5:]),
np.mean(history['delete'][-5:])))
torch.save(model.state_dict(), open(os.path.join('saved', args.model_name, \
'Epoch_{}_JA_{:.4}_DDI_{:.4}.model'.format(epoch, ja, ddi_rate)), 'wb'))
if epoch != 0 and best_ja < ja:
best_epoch = epoch
best_ja = ja
print('best_epoch: {}'.format(best_epoch))
dill.dump(
history,
open(
os.path.join('saved', args.model_name,
'history_{}.pkl'.format(args.model_name)), 'wb'))
def main():
# load data
data_path = '../data/output/records_final.pkl'
voc_path = '../data/output/voc_final.pkl'
device = torch.device('cuda:{}'.format(args.cuda))
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
np.random.seed(1203)
np.random.shuffle(data)
split_point = int(len(data) * 3 / 5)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
model = Retain(voc_size, device=device)
# model.load_state_dict(torch.load(open(os.path.join("saved", args.model_name, args.resume_path), 'rb')))
if args.Test:
model.load_state_dict(
torch.load(
open(os.path.join("saved", args.model_name, args.resume_path),
'rb')))
model.to(device=device)
tic = time.time()
eval(model, data_test, voc_size, 0)
print('test time: {}'.format(time.time() - tic))
return
print('parameters', get_n_params(model))
optimizer = Adam(model.parameters(), args.lr)
model.to(device=device)
history = defaultdict(list)
best_epoch, best_ja = 0, 0
EPOCH = 40
for epoch in range(EPOCH):
tic = time.time()
print('\nepoch {} --------------------------'.format(epoch + 1))
model.train()
for step, input in enumerate(data_train):
if len(input) < 2: continue
loss = 0
for i in range(1, len(input)):
target = np.zeros((1, voc_size[2]))
target[:, input[i][2]] = 1
output_logits = model(input[:i])
loss += F.binary_cross_entropy_with_logits(
output_logits,
torch.FloatTensor(target).to(device))
optimizer.zero_grad()
loss.backward()
optimizer.step()
llprint('\rtraining step: {} / {}'.format(step, len(data_train)))
print()
tic2 = time.time()
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1, avg_add, avg_delete, avg_med = eval(
model, data_eval, voc_size, epoch)
print('training time: {}, test time: {}'.format(
time.time() - tic,
time.time() - tic2))
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['avg_add'].append(avg_add)
history['avg_delete'].append(avg_delete)
history['prauc'].append(prauc)
history['med'].append(avg_med)
if epoch >= 5:
print(
'ddi: {}, Med: {}, Ja: {}, F1: {}, Add: {}, Delete: {}'.format(
np.mean(history['ddi_rate'][-5:]),
np.mean(history['med'][-5:]), np.mean(history['ja'][-5:]),
np.mean(history['avg_f1'][-5:]),
np.mean(history['avg_add'][-5:]),
np.mean(history['avg_delete'][-5:]),
np.mean(history['prauc'][-5:])))
torch.save(model.state_dict(), open(os.path.join('saved', args.model_name, \
'Epoch_{}_JA_{:.4}_DDI_{:.4}.model'.format(epoch, ja, ddi_rate)), 'wb'))
if epoch != 0 and best_ja < ja:
best_epoch = epoch
best_ja = ja
print('best_epoch: {}'.format(best_epoch))
dill.dump(
history,
open(
os.path.join('saved', args.model_name,
'history_{}.pkl'.format(args.model_name)), 'wb'))
def main():
if not os.path.exists(os.path.join("saved", model_name)):
os.makedirs(os.path.join("saved", model_name))
data_path = '../../data/records_final.pkl'
voc_path = '../../data/voc_final.pkl'
device = torch.device('cuda:0')
data = dill.load(open(data_path, 'rb'))
voc = dill.load(open(voc_path, 'rb'))
diag_voc, pro_voc, med_voc = voc['diag_voc'], voc['pro_voc'], voc[
'med_voc']
split_point = int(len(data) * 2 / 3)
data_train = data[:split_point]
eval_len = int(len(data[split_point:]) / 2)
data_test = data[split_point:split_point + eval_len]
data_eval = data[split_point + eval_len:]
voc_size = (len(diag_voc.idx2word), len(pro_voc.idx2word),
len(med_voc.idx2word))
EPOCH = 30
LR = 0.0002
TEST = False
END_TOKEN = voc_size[2] + 1
model = Leap(voc_size, device=device)
if TEST:
model.load_state_dict(
torch.load(
open(os.path.join("saved", model_name, resume_name), 'rb')))
# pass
model.to(device=device)
print('parameters', get_n_params(model))
optimizer = Adam(model.parameters(), lr=LR)
if TEST:
eval(model, data_test, voc_size, 0)
else:
history = defaultdict(list)
for epoch in range(EPOCH):
loss_record = []
start_time = time.time()
model.train()
for step, input in enumerate(data_train):
for adm in input:
loss_target = adm[2] + [END_TOKEN]
output_logits = model(adm)
loss = F.cross_entropy(
output_logits,
torch.LongTensor(loss_target).to(device))
loss_record.append(loss.item())
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
llprint('\rTrain--Epoch: %d, Step: %d/%d' %
(epoch, step, len(data_train)))
ddi_rate, ja, prauc, avg_p, avg_r, avg_f1 = eval(
model, data_eval, voc_size, epoch)
history['ja'].append(ja)
history['ddi_rate'].append(ddi_rate)
history['avg_p'].append(avg_p)
history['avg_r'].append(avg_r)
history['avg_f1'].append(avg_f1)
history['prauc'].append(prauc)
end_time = time.time()
elapsed_time = (end_time - start_time) / 60
llprint(
'\tEpoch: %d, Loss1: %.4f, One Epoch Time: %.2fm, Appro Left Time: %.2fh\n'
% (epoch, np.mean(loss_record), elapsed_time, elapsed_time *
(EPOCH - epoch - 1) / 60))
torch.save(
model.state_dict(),
open(
os.path.join(
'saved', model_name,
'Epoch_%d_JA_%.4f_DDI_%.4f.model' %
(epoch, ja, ddi_rate)), 'wb'))
print('')
dill.dump(history,
open(os.path.join('saved', model_name, 'history.pkl'), 'wb'))
# test
torch.save(
model.state_dict(),
open(os.path.join('saved', model_name, 'final.model'), 'wb'))
