# optim scheduler
scheduler = None
if CONFIG['type'] == "transformer":
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.99)
# record best result
BEST_RESULT = {
"h_acc": 0,
"epoch": 0
}
for epoch in range(CONFIG['start_epoch'], CONFIG['end_epoch']):
# train
train_metrics = model_epoch(loss_name="trainval", epoch=epoch, model=model, neg_sample=CONFIG['neg_sample'],
data_loader=train_loader, concepts=concepts,
optimizer=optimizer, writer=writer, debug=CONFIG['debug'])
for g in [False, True]:
record_name = 'train_g' if g else 'train'
train_class, train_acc = utils.cal_acc(train_metrics, g)
writer.add_scalar(record_name + '_acc', train_acc * 100, epoch)
if CONFIG['skewness']:
train_skew = utils.skewness(train_metrics, g)
writer.add_scalar(record_name + '_skewness', train_skew, epoch)
######################################################################################
# test
record = {tn: {'acc': 0.0, 'class': None} for tn in STATE['split_list'][1:]}
record.update({tn + '_g': {'acc': 0.0, 'class': None} for tn in STATE['split_list'][1:]})
scheduler = None
if CONFIG['type'] == "transformer":
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.99)
# record best result
BEST_RESULT = {"h_acc": 0, "epoch": 0}
for epoch in range(CONFIG['start_epoch'], CONFIG['end_epoch']):
# train
train_metrics = model_epoch(loss_name="trainval",
epoch=epoch,
model=model,
type=CONFIG['type'],
neg_sample=CONFIG['neg_sample'],
data_loader=train_loader,
concepts=concepts,
use_smooth=False,
margin=0.1,
optimizer=optimizer,
writer=writer,
debug=CONFIG['debug'])
for g in [False, True]:
record_name = 'train_g' if g else 'train'
train_class, train_acc = utils.cal_acc(train_metrics, g)
writer.add_scalar(record_name + '_acc', train_acc * 100, epoch)
if CONFIG['skewness']:
train_skew = utils.skewness(train_metrics, g)
writer.add_scalar(record_name + '_skewness', train_skew,
epoch)
# test
# record = {i: {'acc': 0.0, 'class': None} for i in ['conv', 'general']}
test_metrics = {
'total': deque(),
'correct': deque(),
'total_g': deque(),
'correct_g': deque()
}
for tn in STATE['split_list'][1:]:
test_metric = model_epoch(loss_name=tn,
epoch=epoch,
model=model,
type=CONFIG['type'],
data_loader=test_loaders[tn],
concepts=concepts,
optimizer=None,
writer=None,
debug=CONFIG['debug'])
for k in test_metrics.keys():
v = np.asarray(test_metric[k]).argmax(axis=1)
test_metrics[k].extend(v)
print(DATASET + ' skewness:')
for g in [False, True]:
test_skew = utils.skewness(test_metrics, g)
print(test_skew)
optimizer = optim.SGD(params, L_RATE, momentum=CONFIG['momentum'])
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
# scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.5)
# scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[1, 8], gamma=0.1)
for epoch in range(CONFIG['start_epoch'], CONFIG['end_epoch']):
scheduler.step()
# training
train_metrics = model_epoch(loss_name="trainval",
mode="train",
epoch=epoch,
model=model,
k=CONFIG['k'],
d=CONFIG['d'],
data_loader=train_loader,
concepts=concepts,
optimizer=optimizer,
writer=writer)
torch.save(model.state_dict(),
PJ(SAVE_PATH, 'epoch' + str(epoch) + '.pkl'))
train_class, train_acc = cal_acc(train_metrics)
writer.add_scalar('trainval_acc', train_acc * 100, epoch)
######################################################################################
# test
record = {
'params': model.transform.parameters() if CONFIG['freeze'] else model.parameters()
}]
if CONFIG['optim'] == 'SGD':
optimizer = optim.SGD(params, np.float64(CONFIG['l_rate']), momentum=CONFIG['momentum'])
elif CONFIG['optim'] == 'Adam':
optimizer = optim.Adam(params, np.float64(CONFIG['l_rate']))
for epoch in range(1, CONFIG['end_epoch']):
writer = SummaryWriter(PJ(SAVE_PATH, 'val' + str(val_times)))
# training
train_metrics = model_epoch(loss_name="train", epoch=epoch, model=model,
data_loader=train_loader, concepts=concepts,
optimizer=optimizer, writer=writer)
for g in [False, True]:
record_name = 'train_g' if g else 'train'
train_class, train_acc = utils.cal_acc(train_metrics, g)
writer.add_scalar(record_name + '_acc', train_acc * 100, epoch)
######################################################################################
# val
record = {'val': {'acc': 0.0, 'class': None}}
record.update({'val_g': {'acc': 0.0, 'class': None}})
val_metric = model_epoch(loss_name="val", epoch=epoch, model=model,
data_loader=val_loader, concepts=concepts,