def train(self, epochs, optimizer, save=True, savename='', best_accuracy=0, soft_labels=False):
"""Performs training."""
best_accuracy = best_accuracy
error_history = []
if soft_labels:
model_copy = copy.deepcopy(self.model)
self.model_copy = model_copy.cuda()
# Check the sucess here
self.model_copy.eval()
else:
self.model_copy = None
if self.args.cuda:
self.model = self.model.cuda()
for idx in range(epochs):
epoch_idx = idx + 1
# print('Epoch: %d' % (epoch_idx))
# optimizer = utils.step_lr_CIFAR(epoch_idx, self.args.lr, self.args.lr_decay_every,
# self.args.lr_decay_factor, optimizer)
optimizer = utils.step_lr(epoch_idx, self.args.lr, self.args.lr_decay_every,
self.args.lr_decay_factor, optimizer)
if self.args.train_bn:
self.model.train()
else:
self.model.train_nobn()
self.do_epoch(epoch_idx, optimizer)
errors = self.eval(self.pruner.current_dataset_idx)
error_history.append(errors)
accuracy = 100 - errors[0] # Top-1 accuracy.
# Save performance history and stats.
with open(savename + '.json', 'w') as fout:
json.dump({
'error_history': error_history,
'args': vars(self.args),
}, fout)
# Save best model, if required.
if save and accuracy > best_accuracy:
print('Best model so far, Accuracy: %0.2f%% -> %0.2f%%' %
(best_accuracy, accuracy))
best_accuracy = accuracy
self.save_model(epoch_idx, best_accuracy, errors, savename)
print('Finished finetuning...')
print('Best error/accuracy: %0.2f%%, %0.2f%%' %
(100 - best_accuracy, best_accuracy))
print('-' * 16)
def train(self,
epochs,
optimizer,
save=True,
savename='',
best_accuracy=0):
"""Performs training."""
best_accuracy = best_accuracy
error_history = []
if self.args.cuda:
self.model = self.model.cuda()
for idx in range(epochs):
epoch_idx = idx + 1
print('Epoch: %d' % (epoch_idx))
optimizer = utils.step_lr(epoch_idx, self.args.lr,
self.args.lr_decay_every,
self.args.lr_decay_factor, optimizer)
self.model.train()
self.do_epoch(epoch_idx, optimizer)
errors = self.eval()
error_history.append(errors)
accuracy = 100 - errors[0] # Top-1 accuracy.
# Save performance history and stats.
with open(savename + '.json', 'w') as fout:
json.dump(
{
'error_history': error_history,
'args': vars(self.args),
}, fout)
# Save best model, if required.
if save and accuracy > best_accuracy:
print('Best model so far, Accuracy: %0.2f%% -> %0.2f%%' %
(best_accuracy, accuracy))
best_accuracy = accuracy
self.save_model(epoch_idx, best_accuracy, errors, savename)
if isinstance(self.model, nn.DataParallel):
self.model = self.model.module
print('Finished finetuning...')
print('Best error/accuracy: %0.2f%%, %0.2f%%' %
(100 - best_accuracy, best_accuracy))
print('-' * 16)
def update_lr(self, epoch_idx):
"""Update learning rate of every optimizer."""
for optimizer, init_lr, decay_every in zip(self.optimizers, self.lrs,
self.decay_every):
optimizer = utils.step_lr(epoch_idx, init_lr, decay_every,
self.args.lr_decay_factor, optimizer)
def train(self,
epochs,
optimizer,
save=True,
savename='',
best_accuracy=0):
"""Performs training."""
best_accuracy = best_accuracy
error_history = []
if self.args.cuda:
self.model = self.model.cuda()
self.md = ModelData(self.args.train_path, self.train_data_loader,
self.test_data_loader)
bptt, em_sz, nh, nl = 70, 400, 1150, 3
opt_fn = partial(optim.Adam, betas=(0.8, 0.99))
dps = np.array([0.4, 0.5, 0.05, 0.3, 0.4]) * 1.0
self.modell = TextModel(to_gpu(self.model.model))
self.learn = RNN_Learner(self.md,
TextModel(to_gpu(self.model.model)),
opt_fn=opt_fn)
self.learn.reg_fn = partial(seq2seq_reg, alpha=2, beta=1)
self.learn.clip = 25.
self.learn.metrics = [metrics.accuracy]
lr = 0.01
lrm = 2.6
self.lrs = np.array(
[lr / (lrm**4), lr / (lrm**3), lr / (lrm**2), lr / lrm, lr])
self.wd = 1e-6
self.learn.unfreeze()
#self.learn.fit(self.lrs, 3, wds=self.wd)
self.model.shared = self.model.model[0]
self.model.classifier = self.model.model[1]
optimizer = self.learn.get_layer_opt(self.lrs, self.wd).opt
self.loptimizer = self.learn.get_layer_opt(self.lrs, self.wd)
set_trainable(children(self.model.shared), True)
set_trainable(children(self.model.classifier), True)
for idx in range(epochs):
epoch_idx = idx + 1
print('Epoch: %d' % (epoch_idx))
optimizer = utils.step_lr(epoch_idx, self.args.lr,
self.args.lr_decay_every,
self.args.lr_decay_factor, optimizer)
self.model.train()
self.do_epoch(epoch_idx, optimizer)
errors, _, _, _ = self.eval()
self.eval_all()
error_history.append(errors)
accuracy = 100 - errors[0] # Top-1 accuracy.
# Save performance history and stats.
with open(savename + '.json', 'w') as fout:
json.dump(
{
'error_history': error_history,
'args': vars(self.args),
}, fout)
# Save best model, if required.
if save and accuracy > best_accuracy:
print('Best model so far, Accuracy: %0.2f%% -> %0.2f%%' %
(best_accuracy, accuracy))
best_accuracy = accuracy
self.save_model(epoch_idx, best_accuracy, errors, savename)
if isinstance(self.model, nn.DataParallel):
self.model = self.model.module
print('Finished finetuning...')
print('Best error/accuracy: %0.2f%%, %0.2f%%' %
(100 - best_accuracy, best_accuracy))
print('-' * 16)