def validate(val_loader, model, criterion, device, model_h=None, mode=None):
losses = AverageMeter('Loss', ':.4e')
top1 = AverageMeter('Acc@1', ':6.2f')
# keep predicted results and gts for calculate F1 Score
gts = []
preds = []
# switch to evaluate mode
model.eval()
with torch.no_grad():
for i, sample in enumerate(val_loader):
x = sample[0]
if mode == "d":
t = sample[2]
else:
t = sample[1]
x = x.to(device)
t = t.to(device)
batch_size = x.shape[0]
# compute output and loss
if model_h is not None:
model_h.eval()
output = model_h(model(x))
else:
output = model(x)
loss = criterion(output, t)
# measure accuracy and record loss
acc1 = accuracy(output, t, topk=(1,))
losses.update(math.sqrt(loss.item())*100, batch_size)
top1.update(acc1[0].item(), batch_size)
# keep predicted results and gts for calculate F1 Score
_, pred = output.max(dim=1)
gts += list(t.to("cpu").numpy())
preds += list(pred.to("cpu").numpy())
f1s = f1_score(gts, preds, average="macro")
return losses.avg, top1.avg, f1s
def train(train_loader, model, criterion, optimizer, epoch, device):
# 平均を計算してくれるクラス
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
losses = AverageMeter('Loss', ':.4e')
top1 = AverageMeter('Acc@1', ':6.2f')
# 進捗状況を表示してくれるクラス
progress = ProgressMeter(len(train_loader),
[batch_time, data_time, losses, top1],
prefix="Epoch: [{}]".format(epoch))
# keep predicted results and gts for calculate F1 Score
gts = []
preds = []
# switch to train mode
model.train()
end = time.time()
for i, sample in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
x = sample[0]
t = sample[1]
x = x.to(device)
t = t.to(device)
batch_size = x.shape[0]
# compute output and loss
output = model(x)
loss = criterion(output, t)
# measure accuracy and record loss
acc1 = accuracy(output, t, topk=(1, ))
losses.update(loss.item(), batch_size)
top1.update(acc1[0].item(), batch_size)
# keep predicted results and gts for calculate F1 Score
_, pred = output.max(dim=1)
gts += list(t.to("cpu").numpy())
preds += list(pred.to("cpu").numpy())
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# show progress bar per 50 iteration
if i != 0 and i % 50 == 0:
progress.display(i)
# calculate F1 Score
f1s = f1_score(gts, preds, average="macro")
return losses.avg, top1.avg, f1s
def train_adv(train_loader, model_g, model_h, model_d, criterion, optimizer_gh, optimizer_d, epoch, device, beta=1):
model_g.to(device)
model_h.to(device)
model_d.to(device)
# average meter
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
losses_g = AverageMeter('Loss_g', ':.4e')
top1_g = AverageMeter('Acc@1_g', ':6.2f')
losses_d = AverageMeter('Loss_d', ':.4e')
top1_d = AverageMeter('Acc@1_d', ':6.2f')
# progress meter
progress = ProgressMeter(
len(train_loader),
[batch_time, data_time, losses_g, top1_g, losses_d, top1_d],
prefix="Epoch: [{}]".format(epoch)
)
# keep predicted results and gts for calculate F1 Score
gts = []
preds = []
end = time.time()
for i, sample in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
x = sample[0]
t = sample[1]
g = sample[2]
x = x.to(device)
t = t.to(device)
g = g.to(device)
batch_size = x.shape[0]
# train discriminator
model_g.eval()
model_d.train()
feat = model_g(x)
output = model_d(feat)
loss_d = criterion(output, g)
optimizer_d.zero_grad()
loss_d.backward()
optimizer_d.step()
acc1 = accuracy(output, g, topk=(1,))
losses_d.update(loss_d.item(), batch_size)
top1_d.update(acc1[0].item(), batch_size)
# train generator
model_g.train()
model_h.train()
model_d.eval()
feat = model_g(x)
output = model_h(feat)
loss_g = criterion(output, t)
# decieve discriminator
output_d = model_d(model_g(x))
adv_g = torch.LongTensor([0 if i == 1 else 1 for i in g]).to(device)
loss_adv_g = criterion(output_d, adv_g)
loss_g = loss_g + beta*loss_adv_g
optimizer_gh.zero_grad()
loss_g.backward()
optimizer_gh.step()
acc1 = accuracy(output, t, topk=(1,))
losses_g.update(loss_g.item(), batch_size)
top1_g.update(acc1[0].item(), batch_size)
# keep predicted results and gts for calculate F1 Score
_, pred = output.max(dim=1)
gts += list(t.to("cpu").numpy())
preds += list(pred.to("cpu").numpy())
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# show progress bar per 50 iteration
if i != 0 and i % 50 == 0:
progress.display(i)
# calculate F1 Score
f1s = f1_score(gts, preds, average="macro")
return losses_g.avg, losses_d.avg, top1_g.avg, top1_d.avg, f1s