def train_model_multiclass(directory, f, n_classes, opt):
filepath = directory + f
print('Loading data...')
train_ldrs, test_ldrs = load_data(filepath + '.csv')
tr_loss = []
tr_acc = []
vl_loss = []
vl_acc = []
for train_ldr, test_ldr in zip(train_ldrs, test_ldrs):
net = utils.SmallNet(n_classes).to(device)
net.size = 25
criterion = nn.CrossEntropyLoss()
if opt == 'SGD':
optimizer = optim.SGD(net.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=0.0005,
nesterov=True)
elif opt == 'Adam':
optimizer = optim.Adam(net.parameters(),
lr=0.001,
betas=(0.9, 0.999),
weight_decay=0.0005,
amsgrad=False)
elif opt == 'RMSprop':
optimizer = optim.RMSprop(net.parameters(),
lr=0.01,
weight_decay=0.0005,
momentum=0.9)
else:
raise ValueError('Invalid optimizer selected. Choose \'SGD\' or '
'\'Adam\'.')
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=n_schedule,
gamma=0.1)
print('Training...')
print('Filters per layer:', net.n_filters)
print('Criterion:', criterion)
print(optimizer)
losses = [[], [100]]
accs = [[], []]
early_stopping = 0
for epoch in range(n_epochs):
# Training
net.training = True
train_correct = 0
train_total = 0
train_loss = 0.0
for local_batch, local_labels in train_ldr:
# Transfer to GPU
local_batch = local_batch.to(device, dtype=torch.float)
local_labels = local_labels.view(-1).to(device,
dtype=torch.long)
# Train
optimizer.zero_grad()
# Forward + backward + optimize
logits = net(local_batch).view(-1, n_classes)
loss = criterion(logits, local_labels)
loss.backward()
optimizer.step()
# Tracking
train_loss += loss.item()
predicted = torch.argmax(logits, dim=1)
train_total += local_labels.size(0)
train_correct += (predicted == local_labels).sum().item()
train_acc = train_correct / train_total
scheduler.step()
# Validation
net.training = False
val_correct = 0
val_total = 0
val_loss = 0
with torch.no_grad():
for local_batch, local_labels in test_ldr:
# Transfer to GPU
local_batch = local_batch.to(device, dtype=torch.float)
local_labels = local_labels.to(device)
# Test
logits = net(local_batch).view(-1, n_classes)
loss = criterion(logits, local_labels)
# Tracking
val_loss += loss.item()
predicted = torch.argmax(logits, dim=1)
val_total += local_labels.size(0)
val_correct += (predicted == local_labels).sum().item()
val_acc = val_correct / val_total
losses[0].append(train_loss)
losses[1].append(val_loss)
accs[0].append(train_acc)
accs[1].append(val_acc)
if val_loss >= losses[1][-2]:
early_stopping += 1
elif early_stopping > 0:
early_stopping -= 1
early = False
if early_stopping == n_early:
early = True
if epoch % 10 == 9 or early:
print('Epoch:', epoch + 1,
'| Train Acc:', round(train_acc, 8), '| Train Loss:',
round(train_loss,
8), '| Val Acc:', round(val_acc, 8), '| Val Loss:',
round(val_loss, 8), '| Early:', early_stopping)
if early:
print('Early stopping.')
break
losses[1] = losses[1][1:]
tr_loss.append(losses[0])
tr_acc.append(accs[0])
vl_loss.append(losses[1])
vl_acc.append(accs[1])
best = [mean(heapq.nlargest(10, a)) for a in vl_acc]
if plot_:
# Plot loss and accuracy
savedir_ = savedir + '\cnn-2d\\' + f[1:] + '\\'
plot(savedir_, f, tr_loss, tr_acc, vl_loss, vl_acc, best)
return best
def train_model_2_class(directory, f, opt, d):
print('Loading data...')
train_ldrs, test_ldrs = load_data(directory, f, d)
tr_loss = []
tr_acc = []
vl_loss = []
vl_acc = []
train_sizes = []
test_sizes = []
for train_ldr, test_ldr in zip(train_ldrs, test_ldrs):
train_sizes.append(len(train_ldr.dataset))
test_sizes.append(len(test_ldr.dataset))
net = utils.SmallNet(2, n_features).to(device)
net.size = net_size
net.n_filters = net.size
criterion = nn.BCEWithLogitsLoss()
if opt == 'SGD':
optimizer = optim.SGD(net.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=0.0005,
nesterov=True)
elif opt == 'Adam':
optimizer = optim.Adam(net.parameters(),
lr=0.001,
betas=(0.9, 0.999),
weight_decay=0.0005,
amsgrad=False)
elif opt == 'RMSprop':
optimizer = optim.RMSprop(net.parameters(),
lr=0.01,
weight_decay=0.0005,
momentum=0.9)
else:
raise ValueError('Invalid optimizer selected. Choose \'SGD\' or '
'\'Adam\'.')
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=n_schedule,
gamma=0.1)
print('Training...')
print('Filters per layer:', net.n_filters)
print('Criterion:', criterion)
print('Optimizer:', opt)
losses = [[], [100]]
accs = [[], []]
early_stopping = 0
for epoch in range(n_epochs):
# Training
net.training = True
train_correct = 0
train_total = 0
train_loss = 0.0
for local_batch, local_labels in train_ldr:
# Transfer to GPU
local_batch = local_batch.to(device, dtype=torch.float)
local_labels = local_labels.view(-1, 1).to(device,
dtype=torch.float)
# Train
optimizer.zero_grad()
# Forward + backward + optimize
logits = net(local_batch).view(-1, 1)
loss = criterion(logits, local_labels)
loss.backward()
optimizer.step()
# Tracking
train_loss += loss.item()
outputs = torch.sigmoid(logits)
predicted = (outputs >= 0.5).view(-1).to(device,
dtype=torch.long)
local_labels = local_labels.view(-1).to(device,
dtype=torch.long)
train_total += local_labels.size(0)
train_correct += (predicted == local_labels).sum().item()
train_acc = train_correct / train_total
scheduler.step()
# Validation
net.training = False
val_correct = 0
val_total = 0
val_loss = 0
with torch.no_grad():
for local_batch, local_labels in test_ldr:
# Transfer to GPU
local_batch = local_batch.to(device, dtype=torch.float)
local_labels = local_labels.view(-1,
1).to(device,
dtype=torch.float)
# Test
logits = net(local_batch).view(-1, 1)
loss = criterion(logits, local_labels)
# Tracking
val_loss += loss.item()
outputs = torch.sigmoid(logits)
predicted = (outputs >= 0.5).view(-1).to(device,
dtype=torch.long)
local_labels = local_labels.view(-1).to(device,
dtype=torch.long)
val_total += local_labels.size(0)
val_correct += (predicted == local_labels).sum().item()
val_acc = val_correct / val_total
losses[0].append(train_loss)
losses[1].append(val_loss)
accs[0].append(train_acc)
accs[1].append(val_acc)
if val_loss <= losses[1][-2]:
early_stopping += 1
elif early_stopping > 0:
early_stopping -= 1
early = False
if early_stopping >= n_early and epoch > min_epochs:
early = True
if epoch % 10 == 9 or early:
print('Epoch:', epoch + 1,
'| Train Acc:', round(train_acc, 8), '| Train Loss:',
round(train_loss,
8), '| Val Acc:', round(val_acc, 8), '| Val Loss:',
round(val_loss, 8), '| Early:', early_stopping)
if early:
print('Early stopping.')
break
losses[1] = losses[1][1:]
tr_loss.append(losses[0])
tr_acc.append(accs[0])
vl_loss.append(losses[1])
vl_acc.append(accs[1])
best = [max(a) for a in vl_acc]
if plot_:
# Plot loss and accuracy
savedir_ = savedir + '\mlp-mrmr-' + str(n_features) + '-features\\' + \
d[1:] + '\\'
plot(savedir_, d, tr_loss, tr_acc, vl_loss, vl_acc, best)
return best, train_sizes, test_sizes, vl_acc