def main():
net = load_net()
net = net.cuda()
net.eval()
dt = dataset.FashionDataset('validation', transform=dataset.val_transform)
loader = data.DataLoader(dt,
batch_size=32,
shuffle=False,
num_workers=8,
pin_memory=True,
drop_last=False)
prec_meter = AverageMeter()
reca_meter = AverageMeter()
avF1_meter = AverageMeter()
for i, [image, label] in enumerate(loader):
image = image.cuda()
label = label.cuda()
with torch.no_grad():
output = net(image)
prec, reca, avF1 = statistics(output, label)
prec_meter.update(prec)
reca_meter.update(reca)
avF1_meter.update(avF1)
if i % 20 == 0:
print('Iter: %d, precision: %.4f, recall: %.4f, F1: %.4f' %
(i, prec_meter.avg, reca_meter.avg, avF1_meter.avg))
print('precision: %.4f, recall: %.4f, F1: %.4f' %
(prec_meter.avg, reca_meter.avg, avF1_meter.avg))
def main():
base_lr = 1e-1
net, opt = load_net_opt(base_lr=base_lr)
crit = nn.BCEWithLogitsLoss().cuda()
net = net.cuda()
dt = dataset.FashionDataset('train', transform=dataset.train_transform)
loader = data.DataLoader(dt, batch_size=32, shuffle=True, num_workers=8, pin_memory=True)
dt_iter = iter(loader)
writer = SummaryWriter('./tf_record')
loss_meter = AverageMeter()
prec_meter = AverageMeter()
reca_meter = AverageMeter()
avF1_meter = AverageMeter()
for step in range(1, 100001):
adjust_lr(step, base_lr, opt)
try:
image, label = next(dt_iter)
except StopIteration:
dt_iter = iter(loader)
image, label = next(dt_iter)
image = image.cuda()
label = label.cuda()
output = net(image)
loss = crit(output, label)
loss_meter.update(loss)
opt.zero_grad()
loss.backward()
opt.step()
prec, reca, avF1 = statistics(output, label)
prec_meter.update(prec)
reca_meter.update(reca)
avF1_meter.update(avF1)
if step % 200 == 0:
print('Iter: %d, loss: %f, precision: %.4f, recall: %.4f, F1: %.4f' %
(step, loss_meter.avg, prec_meter.avg, reca_meter.avg, avF1_meter.avg))
writer.add_scalar('loss', loss_meter.avg, step)
writer.add_scalar('prec', prec_meter.avg, step)
writer.add_scalar('reca', reca_meter.avg, step)
writer.add_scalar('F1', avF1_meter.avg, step)
if step % 500 == 0:
torch.save(net.state_dict(), 'models/step_%d.pth' % step)
if step % 2000 == 0:
torch.save(net.state_dict(), 'models/latest.pth')
def main():
base_lr = 1e-2
net, opt = load_net_opt(base_lr=base_lr)
net = net.cuda()
dt = dataset.FashionDataset('train', transform=dataset.train_transform)
loader = data.DataLoader(dt, batch_size=32, shuffle=True, num_workers=8, pin_memory=True)
dt_iter = iter(loader)
writer = SummaryWriter('./tf_record')
loss_meters = {i: AverageMeter() for i in [1, 2, 3, 4, 5, 6, 8, 9]}
loss_meter = AverageMeter()
prec_meter = AverageMeter()
reca_meter = AverageMeter()
avF1_meter = AverageMeter()
for step in range(1, 160001):
adjust_lr(step, base_lr, opt)
try:
image, label = next(dt_iter)
except StopIteration:
dt_iter = iter(loader)
image, label = next(dt_iter)
image = image.cuda()
label = label.cuda()
output = net(image)
eps = 1e-6
losses = []
for task_id in [1, 2, 3, 4, 5, 6, 8, 9]:
start_idx = start_idxs[task_id]
end_idx = start_idxs[task_id+1]
if task_id in [2, 4, 6]:
prob = torch.sigmoid(output[:, start_idx:end_idx])
prob = torch.clamp(prob, eps, 1-eps)
log_prob = -torch.log(prob)
logprob_ = -torch.log(1 - prob)
target = label[:, start_idx:end_idx]
loss = (log_prob * target + logprob_ * (1 - target)).mean()
else:
prob = F.softmax(output[:, start_idx:end_idx], dim=1)
prob = torch.clamp(prob, eps, 1-eps)
log_prob = -torch.log(prob)
target = label[:, start_idx:end_idx]
loss = (log_prob * target).sum(dim=1).mean()
losses.append(loss)
loss_meters[task_id].update(loss)
loss = sum(losses)# / 8
loss_meter.update(loss)
opt.zero_grad()
loss.backward()
opt.step()
prec, reca, avF1 = statistics(output, label)
prec_meter.update(prec)
reca_meter.update(reca)
avF1_meter.update(avF1)
if step % 1000 == 0:
print('Iter: %d, loss: %f, precision: %.4f, recall: %.4f, F1: %.4f' %
(step, loss_meter.avg, prec_meter.avg, reca_meter.avg, avF1_meter.avg))
print(', '.join(['loss%d: %.4f' % (task_id, meter.avg) for task_id, meter in loss_meters.items()]))
writer.add_scalar('loss', loss_meter.avg, step)
writer.add_scalar('prec', prec_meter.avg, step)
writer.add_scalar('reca', reca_meter.avg, step)
writer.add_scalar('F1', avF1_meter.avg, step)
for task_id, meter in loss_meters.items():
writer.add_scalar('loss%d', meter.avg, step)
if step % 2000 == 0:
torch.save(net.state_dict(), 'models/latest.pth')
if step % 10000 == 0:
torch.save(net.state_dict(), 'models/step_%d.pth' % step)