def _run_epoch(epoch, mode):
model.train(True)
if mode == "train":
description = 'Train'
dataset = trainData
shuffle = True
else:
description = 'Valid'
dataset = validData
shuffle = False
dataloader = DataLoader(dataset=dataset,
batch_size=batch_size,
shuffle=shuffle,
collate_fn=dataset.collate_fn,
num_workers=8)
trange = tqdm(enumerate(dataloader),
total=len(dataloader),
desc=description)
loss = 0
f1_score = F1()
# from class AbstractDataset(Dataset)
for i, (
x, y, sent_len
) in trange: # x = torch.LongTensor(batch_abstract), y = torch.FloatTensor(batch_label), sent_len = sent_len
# Butters
#print('In _run_epoch, i=', str(i), ' ', 'shape of x', x.shape, ' ', 'shape of y', y.shape, ' ', 'len of sent_len', len(sent_len), '\n')
if (x.size()[0] is not batch_size):
continue
o_labels, batch_loss = _run_iter(x, y)
if mode == "train":
opt.zero_grad()
batch_loss.backward()
opt.step()
loss += batch_loss.item()
f1_score.update(o_labels.cpu(), y)
trange.set_postfix(loss=loss / (i + 1), f1=f1_score.print_score())
if mode == "train":
history['train'].append({
'f1': f1_score.get_score(),
'loss': loss / len(trange)
})
writer.add_scalar('Loss/train', loss / len(trange), epoch)
writer.add_scalar('F1_score/train', f1_score.get_score(), epoch)
else:
history['valid'].append({
'f1': f1_score.get_score(),
'loss': loss / len(trange)
})
writer.add_scalar('Loss/valid', loss / len(trange), epoch)
writer.add_scalar('F1_score/valid', f1_score.get_score(), epoch)
trange.close()
def run_epoch(self, epoch, training):
self.model.train(training)
if training:
description = 'Train'
dataset = self.trainData
shuffle = True
else:
description = 'Valid'
dataset = self.validData
shuffle = False
dataloader = DataLoader(dataset=dataset,
batch_size=self.batch_size,
shuffle=shuffle,
collate_fn=dataset.collate_fn,
num_workers=4)
trange = tqdm(enumerate(dataloader),
total=len(dataloader),
desc=description)
loss = 0
acc = accuracy()
for i, (x, _, y) in trange:
o_labels, batch_loss = self.run_iter(x, y)
if training:
self.opt.zero_grad()
batch_loss.backward()
self.opt.step()
loss += batch_loss.item()
acc.update(o_labels.cpu(), y)
trange.set_postfix(loss=loss / (i + 1), acc=acc.print_score())
if training:
self.history['train'].append({
'acc': acc.get_score(),
'loss': loss / len(trange)
})
self.save_hist()
else:
self.history['valid'].append({
'acc': acc.get_score(),
'loss': loss / len(trange)
})
self.save_hist()
if acc.get_score() > self.best_val:
self.best_val = acc.get_score()
self.save_best(epoch)
def _run_epoch(epoch, training, fine_tune):
model.train(training)
if training:
description = 'Train'
dataset = trainData
shuffle = True
else:
description = 'Valid'
dataset = validData
shuffle = False
if fine_tune:
model.unfreeze()
else:
model.freeze()
dataloader = DataLoader(dataset=dataset,
batch_size=12,
shuffle=shuffle,
collate_fn=dataset.collate_fn,
num_workers=4)
trange = tqdm(enumerate(dataloader),
total=len(dataloader),
desc=description)
loss = 0
f1_score = F1()
for i, (x, y) in trange:
batch_loss, o_labels = _run_iter(x, y)
if training:
opt.zero_grad()
batch_loss.backward()
opt.step()
loss += batch_loss.item()
f1_score.update(o_labels.cpu(), y)
trange.set_postfix(loss=loss / (i + 1), f1=f1_score.print_score())
if training:
history['train'].append({
'f1': f1_score.get_score(),
'loss': loss / len(trange)
})
else:
history['valid'].append({
'f1': f1_score.get_score(),
'loss': loss / len(trange)
})
def _run_epoch(epoch, mode):
model.train(True)
if mode=="train":
description = 'Train'
dataset = trainData
shuffle = True
else:
description = 'Valid'
dataset = validData
shuffle = False
dataloader = DataLoader(dataset=dataset,
batch_size=batch_size,
shuffle=shuffle,
collate_fn=dataset.collate_fn,
num_workers=8)
trange = tqdm(enumerate(dataloader), total=len(dataloader), desc=description)
loss = 0
f1_score = F1()
for i, (x, y, sent_len) in trange:
o_labels, batch_loss = _run_iter(x,y)
if mode=="train":
opt.zero_grad()
batch_loss.backward()
opt.step()
loss += batch_loss.item()
f1_score.update(o_labels.cpu(), y)
trange.set_postfix(
loss=loss / (i + 1), f1=f1_score.print_score())
if mode=="train":
history['train'].append({'f1':f1_score.get_score(), 'loss':loss/ len(trange)})
writer.add_scalar('Loss/train', loss/ len(trange), epoch)
writer.add_scalar('F1_score/train', f1_score.get_score(), epoch)
else:
history['valid'].append({'f1':f1_score.get_score(), 'loss':loss/ len(trange)})
writer.add_scalar('Loss/valid', loss/ len(trange), epoch)
writer.add_scalar('F1_score/valid', f1_score.get_score(), epoch)
trange.close()
def run_epoch(self, epoch, training):
self.classficationA.train(training)
self.classficationB.train(training)
if training:
description = 'Train'
dataset = self.trainData
shuffle = True
else:
description = 'Valid'
dataset = self.validData
shuffle = False
dataloader = DataLoader(dataset=dataset,
batch_size=256,
shuffle=shuffle,
collate_fn=dataset.collate_fn,
num_workers=4)
trange = tqdm(enumerate(dataloader),
total=len(dataloader),
desc=description)
mse_loss = 0
lossA = 0
lossB = 0
accA = accuracy()
accB = accuracy()
for i, (ft, missing_ft, labels) in trange:
ft = ft.to(self.device)
missing_ft = missing_ft.to(self.device)
all_ft = torch.cat([ft, missing_ft], dim=1)
labels = labels.to(self.device)
# ------------------
# Train ClassifierA
# ------------------
missing_out, missing_hidden_out = self.classficationA(ft)
all_out, all_hidden_out = self.classficationB(all_ft)
batch_loss = self.criterion(missing_out, labels)
batch_mse_loss = 0
for missing_hidden, all_hidden in zip(missing_hidden_out,
all_hidden_out):
batch_mse_loss += self.mse_loss(missing_hidden, all_hidden)
mse_loss += batch_mse_loss.item()
if training:
self.opt_C_A.zero_grad()
(batch_mse_loss + batch_loss).backward()
self.opt_C_A.step()
lossA += batch_loss.item()
accA.update(missing_out, labels)
# ------------------
# Train ClassifierB
# ------------------
all_out, _ = self.classficationB(all_ft)
batch_loss = self.criterion(all_out, labels)
if training:
self.opt_C_B.zero_grad()
batch_loss.backward()
self.opt_C_B.step()
lossB += batch_loss.item()
accB.update(all_out, labels)
trange.set_postfix(accA=accA.print_score(),
accB=accB.print_score(),
lossA=lossA / (i + 1),
lossB=lossB / (i + 1),
mseLoss=mse_loss / (i + 1))
if training:
self.history['train'].append({
'accA': accA.get_score(),
'accB': accB.get_score(),
'lossA': lossA / len(trange),
'lossB': lossB / len(trange),
'mseLoss': mse_loss / len(trange)
})
self.save_hist()
else:
self.history['valid'].append({
'accA': accA.get_score(),
'accB': accB.get_score(),
'lossA': lossA / len(trange),
'lossB': lossB / len(trange),
'mseLoss': mse_loss / len(trange)
})
self.save_hist()
if self.best_val < accA.get_score():
self.best_val = accA.get_score()
self.save_best(epoch)
def run_epoch(self, epoch, training):
self.fadding_model.train(training)
self.fixed_model.train(False)
if training:
description = 'Train'
dataset = self.trainData
shuffle = True
else:
description = 'Valid'
dataset = self.validData
shuffle = False
dataloader = DataLoader(dataset=dataset,
batch_size=self.batch_size,
shuffle=shuffle,
collate_fn=dataset.collate_fn,
num_workers=4)
trange = tqdm(enumerate(dataloader),
total=len(dataloader),
desc=description)
loss = 0
acc_fadding = accuracy()
acc_fixed = accuracy()
for i, (ft, missing_ft, labels) in trange:
ft = ft.to(self.device)
missing_ft = missing_ft.to(self.device)
labels = labels.to(self.device)
missing_fadding_ft = missing_ft * (0.9**((epoch * 100)**(1 / 2)))
missing_0_ft = missing_ft * 0
fadding_ft = torch.cat([missing_fadding_ft, ft], dim=1)
zero_ft = torch.cat([missing_0_ft, ft], dim=1)
raw_ft = torch.cat([missing_ft, ft], dim=1)
fadding_out, fadding_hiddens = self.fadding_model(fadding_ft)
zero_out, _ = self.fadding_model(zero_ft)
raw_out, raw_hiddens = self.fixed_model(raw_ft)
batch_loss = 0
for raw_hidden, fadding_hidden in zip(raw_hiddens,
fadding_hiddens):
batch_loss += self.criteria(raw_hidden, fadding_hidden)
batch_loss += self.criteria(raw_out, fadding_out)
if training:
self.opt.zero_grad()
batch_loss.backward()
self.opt.step()
loss += batch_loss.item()
acc_fadding.update(fadding_out, labels)
acc_fixed.update(zero_out, labels)
trange.set_postfix(loss=loss / (i + 1),
acc_fadding=acc_fadding.print_score(),
acc_fixed=acc_fixed.print_score())
# self.scheduler.step()
if training:
self.history['train'].append({
'acc-fadding': acc_fadding.get_score(),
'acc_fixed': acc_fixed.get_score(),
'loss': loss / len(trange)
})
self.save_hist()
else:
self.history['valid'].append({
'acc-fadding': acc_fadding.get_score(),
'acc_fixed': acc_fixed.get_score(),
'loss': loss / len(trange)
})
self.save_hist()
if acc_fixed.get_score() > self.best_val:
self.best_val = acc_fixed.get_score()
self.save_best(epoch)
all_delta_Ei = []
for i in trange:
qubits_mat, beta, E_off, flip_count, delta_Ei_list = TSPsolver(city_size,
qubits_mat,
distance_mat,
beta,
beta_increment,
E_off,
flip_count,
E_off_increment)
a_term,b_term = calculate_E(city_size, qubits_mat, distance_mat, A,B)
trange.set_postfix(Energy = a_term+b_term,
one_bits = np.sum(qubits_mat),
beta = beta,
E_off = E_off,
Aterm = a_term,
Bterm = b_term,
flip = flip_count)
total_E.append(a_term+b_term)
#%%
# make plot
# find city order
start_point, end_point = [],[]
for i in range(city_size-1):
first = np.where(qubits_mat[:,i]==1)[0]
second = np.where(qubits_mat[:,i+1]==1)[0]
first, second = np.repeat(first, len(second)), np.tile(second, len(first))
start_point.extend(first)
end_point.extend(second)
def run_epoch(self, epoch, training):
self.generator.train(training)
self.discriminator.train(training)
self.classfication.train(training)
if training:
description = 'Train'
dataset = self.trainData
shuffle = True
else:
description = 'Valid'
dataset = self.validData
shuffle = False
dataloader = DataLoader(dataset=dataset,
batch_size=256,
shuffle=shuffle,
collate_fn=dataset.collate_fn,
num_workers=4)
trange = tqdm(enumerate(dataloader),
total=len(dataloader),
desc=description)
g_loss = 0
d_loss = 0
loss = 0
acc = accuracy()
for i, (ft, missing_ft, labels) in trange:
ft = ft.to(self.device)
missing_ft = missing_ft.to(self.device)
labels = labels.to(self.device)
batch_size = ft.shape[0]
true = Variable(torch.FloatTensor(batch_size, 1).fill_(1.0),
requires_grad=False).to(self.device) # (batch, 1)
fake = Variable(torch.FloatTensor(batch_size, 1).fill_(0.0),
requires_grad=False).to(self.device) # (batch, 1)
# -----------------
# Train Generator
# -----------------
gen_missing = self.generator(ft.detach())
validity = self.discriminator(gen_missing)
batch_g_loss = self.adversarial_loss(validity, true)
if training:
self.opt_G.zero_grad()
batch_g_loss.backward()
self.opt_G.step()
g_loss += batch_g_loss.item()
# ---------------------
# Train Discriminator
# ---------------------
real_pred = self.discriminator(missing_ft)
d_real_loss = self.adversarial_loss(real_pred, true)
fake_missing = self.generator(ft.detach())
fake_pred = self.discriminator(fake_missing)
d_fake_loss = self.adversarial_loss(fake_pred, fake)
batch_d_loss = (d_real_loss + d_fake_loss) / 2
if training:
self.opt_D.zero_grad()
batch_d_loss.backward()
self.opt_D.step()
d_loss += batch_d_loss.item()
# ------------------
# Train Classifier
# ------------------
gen_missing = self.generator(ft.detach())
all_features = torch.cat((ft, gen_missing), dim=1)
o_labels = self.classfication(all_features)
batch_loss = self.criterion(o_labels, labels)
if training:
self.opt_C.zero_grad()
batch_loss.backward()
self.opt_C.step()
loss += batch_loss.item()
acc.update(o_labels, labels)
trange.set_postfix(acc=acc.print_score(),
g_loss=g_loss / (i + 1),
d_loss=d_loss / (i + 1),
loss=loss / (i + 1))
if training:
self.history['train'].append({
'acc': acc.get_score(),
'g_loss': g_loss / len(trange),
'd_loss': d_loss / len(trange),
'loss': loss / len(trange)
})
self.save_hist()
else:
self.history['valid'].append({
'acc': acc.get_score(),
'g_loss': g_loss / len(trange),
'd_loss': d_loss / len(trange),
'loss': loss / len(trange)
})
self.save_hist()
if self.best_val < acc.get_score():
self.best_val = acc.get_score()
self.save_best(epoch)
