def run(self, epoch=None):
'''
Run a training loop.
'''
if epoch is None:
epoch = self.args.n_epoch
print('mbsize: {}, hidden size: {}, layer: {}, dropout: {}'.format(
self.mb, self.hidden, self.layer, self.dropout),
file=self.file)
# Init the model, the optimizer and some structures for logging
self.reset()
model = MLP(self.hidden, self.layer, self.dropout)
#print(model)
model.reset_parameters()
#model.cuda()
model.to(self.args.device)
#model = torch.nn.DataParallel(model)
opt = optim.Adam(model.parameters())
acc = 0 # best dev. acc.
accc = 0 # test acc. at the time of best dev. acc.
e = -1 # best dev iteration/epoch
times = []
losses = []
ftime = []
btime = []
utime = []
print('Initial evaluation on dev set:')
self._evaluate(model, self.devloader, 'dev')
start = time.time()
# training loop
for t in range(epoch):
print('{}：'.format(t), end='', file=self.file, flush=True)
# train
#start = torch.cuda.Event(True)
#end = torch.cuda.Event(True)
#start.record()
loss, ft, bt, ut = self._train(model, opt)
#end.record()
#end.synchronize()
#ttime = start.elapsed_time(end)
print("(wall time: {:.1f} sec) ".format(time.time() - start),
end='')
#times.append(ttime)
losses.append(loss)
ftime.append(ft)
btime.append(bt)
utime.append(ut)
# predict
curacc = self._evaluate(model, self.devloader, 'dev')
#if curacc > acc:
# e = t
# acc = curacc
# accc = self._evaluate(model, self.testloader, ' test')
etime = [sum(t) for t in zip(ftime, btime, utime)]
print('test acc: {:.2f}'.format(
self._evaluate(model, self.testloader, ' test')))
print('best on val set - ${:.2f}|{:.2f} at {}'.format(acc, accc, e),
file=self.file,
flush=True)
print('', file=self.file)
y_pred, y_true = [], []
for step,(x,y) in enumerate(loader):
x = x.cuda(device)
out = model(x)
y_pred.append(torch.argmax(out, dim=1, keepdim=True).cpu())
y_true.append(y)
return evaluator.eval({
"y_true": torch.cat(y_true, dim=0),
"y_pred": torch.cat(y_pred, dim=0),
})['acc']
bad_counter = 0
best = 0
best_epoch = 0
train_time = 0
model.reset_parameters()
print("--------------------------")
print("Training...")
for epoch in range(args.epochs):
loss_tra,train_ep = train(model,args.dev,train_loader,optimizer)
f1_val = validate(model, args.dev, valid_loader, evaluator)
train_time+=train_ep
if(epoch+1)%5 == 0:
print(f'Epoch:{epoch+1:02d},'
f'Train_loss:{loss_tra:.3f}',
f'Valid_acc:{100*f1_val:.3f}%',
f'Time_cost{train_time:.3f}')
if f1_val > best:
best = f1_val
best_epoch = epoch
torch.save(model.state_dict(), checkpt_file)
