def submit():
load_model_path = "checkpoints/LSTM_0615_22:42:44.pth"
model = Sequence(14, 128, 1).cuda()
model.load(load_model_path)
user_list = []
probability = []
submit_data = AppData("data/submit_data_20d.json", iflabel=False)
submit_dataloader = DataLoader(submit_data, 512, shuffle=False)
for ii, (input, property, user_id) in tqdm(enumerate(submit_dataloader)):
val_input = Variable(input, volatile=True).cuda()
val_input2 = Variable(property).cuda()
score = model(val_input, val_input2).cpu()
probability.extend(t.nn.functional.softmax(score)[:, 1].data.tolist())
user_list.extend(user_id.tolist())
index = np.argmax(probability)
print(user_list[index], probability[index])
index2 = np.argmin(probability)
print(user_list[index2], probability[index2])
all = zip(probability, user_list)
all = sorted(all, key=lambda x: x[0], reverse=True)
with open('submission/submission_06_15_22_45_21785.txt', 'w') as f:
for i in all[:21785]:
f.writelines(str(i[1]) + '\n')
print(all[21785], all[23800])
print(len(all))
def test_offline():
test_data = AppData("data/train_20d_3p.json", iflabel=True)
test_dataloader = DataLoader(test_data, 128, shuffle=False, num_workers=2)
load_model_path = "checkpoints/LSTM_0616_08:30:43.pth"
model = Sequence(14, 128, 1).cuda()
model.load(load_model_path)
test_cm, test_f1 = val(model, test_dataloader)
print(test_f1)
print(test_cm)
def submit():
model_time = "0628_10:39:34"
load_model_path = "checkpoints/06281033/LSTM_%s.pth" % model_time
model = Sequence(31, 128, 1).cuda()
model.load(load_model_path)
user_list = []
probability = []
submit_data = AppData("../kesci/data/data_v3_23d/submit2.json",
iflabel=False)
submit_dataloader = DataLoader(submit_data, 512, shuffle=False)
for ii, (input, property, user_id) in tqdm(enumerate(submit_dataloader)):
val_input = Variable(input, volatile=True).cuda()
val_input2 = Variable(property).cuda()
score = model(val_input, val_input2).cpu()
probability.extend(F.sigmoid(score)[:, 0].data.tolist())
user_list.extend(user_id.tolist())
index = np.argmax(probability)
print(user_list[index], probability[index])
index2 = np.argmin(probability)
print(user_list[index2], probability[index2])
all = zip(probability, user_list)
all = sorted(all, key=lambda x: x[0], reverse=True)
getnum = 24000
with open(
'../kesci/submission/submission_%s_%d.txt' % (model_time, getnum),
'w') as f:
num = 0
for i in all[:getnum]:
#if i[0] > 0.5:
f.writelines(str(i[1]) + '\n')
num += 1
print(num)
with open('../kesci/submission/score_%s.txt' % model_time, 'w') as f:
for i in all:
f.writelines("%s\t%s\n" % (str(i[1]), str(i[0])))
def train():
vis = Visualizer("Kesci" + time.strftime('%m%d%H%M'))
train_data = AppData("../kesci/data/data_v3_23d/train_ab.json",
iflabel=True)
val_data = AppData("../kesci/data/data_v3_23d/val_ab.json", iflabel=True)
train_dataloader = DataLoader(train_data, 256, shuffle=True, num_workers=4)
val_dataloader = DataLoader(val_data, 512, shuffle=False, num_workers=2)
test_data = AppData("../kesci/data/data_v3_23d/test_ab.json", iflabel=True)
test_dataloader = DataLoader(test_data, 512, shuffle=False, num_workers=2)
criterion = t.nn.BCEWithLogitsLoss().cuda()
learning_rate = 0.002
weight_decay = 0.0003
model = DoubleSequence(31, 128, 1).cuda()
optimizer = t.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
loss_meter = meter.AverageValueMeter()
confusion_matrix = meter.ConfusionMeter(2)
previous_loss = 1e100
for epoch in range(400):
loss_meter.reset()
confusion_matrix.reset()
for ii, (data, property, target) in tqdm(enumerate(train_dataloader)):
input = Variable(data).cuda()
input2 = Variable(property).cuda()
target = Variable(target).cuda()
output = model(input, input2)
optimizer.zero_grad()
loss = criterion(output, target)
loss.backward()
optimizer.step()
loss_meter.add(loss.data[0])
if ii % 100 == 99:
vis.plot('loss', loss_meter.value()[0])
if epoch % 3 == 2:
train_cm, train_f1 = val(model, train_dataloader)
vis.plot('train_f1', train_f1)
val_cm, val_f1 = val(model, val_dataloader)
vis.plot_many({'val_f1': val_f1, 'learning_rate': learning_rate})
if loss_meter.value()[0] > previous_loss:
learning_rate = learning_rate * 0.9
# 第二种降低学习率的方法:不会有moment等信息的丢失
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
previous_loss = loss_meter.value()[0]
if epoch % 3 == 2:
model.save()
test_cm, test_f1 = val(model, test_dataloader)
vis.plot('test_f1', test_f1)
vis.log(
"训练集：{train_f1:%}, {train_pre:%}, {train_rec:%} | 验证集：{val_f1:%}, {val_pre:%}, {val_rec:%} | \
测试集：{test_f1:%}, {test_pre:%}, {test_rec:%} | {train_true_num:%}, {val_true_num:%}, {test_true_num:%}"
.format(
train_f1=train_f1,
val_f1=val_f1,
test_f1=test_f1,
train_true_num=train_cm.value()[:, 0].sum() /
len(train_data),
val_true_num=val_cm.value()[:, 0].sum() / len(val_data),
test_true_num=test_cm.value()[:, 0].sum() / len(test_data),
train_pre=train_cm.value()[0][0] /
train_cm.value()[0].sum(),
train_rec=train_cm.value()[0][0] /
train_cm.value()[:, 0].sum(),
val_pre=val_cm.value()[0][0] / val_cm.value()[0].sum(),
val_rec=val_cm.value()[0][0] / val_cm.value()[:, 0].sum(),
test_pre=test_cm.value()[0][0] / test_cm.value()[0].sum(),
test_rec=test_cm.value()[0][0] /
test_cm.value()[:, 0].sum()))
def train():
vis = Visualizer("Kesci")
train_data = AppData("data/data_16d_target/train.json", iflabel=True)
val_data = AppData("data/data_16d_target/val.json", iflabel=True)
train_dataloader = DataLoader(train_data, 32, shuffle=True, num_workers=4)
val_dataloader = DataLoader(val_data, 256, shuffle=False, num_workers=2)
test_data = AppData("data/data_16d_target/test.json", iflabel=True)
test_dataloader = DataLoader(test_data, 256, shuffle=False, num_workers=2)
criterion = t.nn.CrossEntropyLoss().cuda()
learning_rate = 0.003
weight_decay = 0.0002
model = Sequence(15, 128, 1).cuda()
optimizer = t.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
loss_meter = meter.AverageValueMeter()
confusion_matrix = meter.ConfusionMeter(2)
previous_loss = 1e100
for epoch in range(500):
loss_meter.reset()
confusion_matrix.reset()
for ii, (data, property, label) in tqdm(enumerate(train_dataloader)):
input = Variable(data).cuda()
input2 = Variable(property).cuda()
target = Variable(label).cuda().view(-1)
output = model(input, input2)
optimizer.zero_grad()
loss = criterion(output, target)
loss.backward()
optimizer.step()
loss_meter.add(loss.data[0])
confusion_matrix.add(output.data, target.data)
if ii % 100 == 99:
vis.plot('loss', loss_meter.value()[0])
if epoch % 3 == 2:
train_cm, train_f1 = val(model, train_dataloader)
vis.plot('train_f1', train_f1)
val_cm, val_f1 = val(model, val_dataloader)
vis.plot_many({'val_f1': val_f1, 'learning_rate': learning_rate})
# vis.log("epoch:{epoch},lr:{lr},loss:{loss},train_cm:{train_cm},val_cm:{val_cm}".format(
# epoch=epoch, loss=loss_meter.value()[0], val_cm=str(val_cm.value()),
# train_cm=str(confusion_matrix.value()), lr=learning_rate))
if loss_meter.value()[0] > previous_loss:
learning_rate = learning_rate * 0.95
# 第二种降低学习率的方法:不会有moment等信息的丢失
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
previous_loss = loss_meter.value()[0]
if epoch % 10 == 9:
model.save()
test_cm, test_f1 = val(model, test_dataloader)
vis.plot('test_f1', test_f1)
vis.log(
"model:{model} | {train_f1}, {train_pre}, {train_rec} | {val_f1}, {val_pre}, {val_rec} | {test_f1}, {test_pre}, {test_rec}"
.format(train_f1=train_f1,
val_f1=val_f1,
test_f1=test_f1,
model=time.strftime('%m%d %H:%M:%S'),
train_pre=str(train_cm.value()[0][0] /
train_cm.value()[:, 0].sum()),
train_rec=str(train_cm.value()[0][0] /
train_cm.value()[0].sum()),
val_pre=str(val_cm.value()[0][0] /
val_cm.value()[:, 0].sum()),
val_rec=str(val_cm.value()[0][0] /
val_cm.value()[0].sum()),
test_pre=str(test_cm.value()[0][0] /
test_cm.value()[:, 0].sum()),
test_rec=str(test_cm.value()[0][0] /
test_cm.value()[0].sum())))