def train_loop(self, train_w_loader, train_thetas_loader, test_loader,
model):
best_top1 = 0.0
# firstly, train weights only
for epoch in range(self.train_thetas_from_the_epoch):
self.writer.add_scalar('learning_rate/weights',
self.w_optimizer.param_groups[0]['lr'],
epoch)
self.logger.info("Firstly, start to train weights for epoch %d" %
(epoch))
self._training_step(model,
train_w_loader,
self.w_optimizer,
epoch,
info_for_logger="_w_step_")
self.w_scheduler.step()
for epoch in range(self.train_thetas_from_the_epoch, self.cnt_epochs):
self.writer.add_scalar('learning_rate/weights',
self.w_optimizer.param_groups[0]['lr'],
epoch)
self.writer.add_scalar('learning_rate/theta',
self.theta_optimizer.param_groups[0]['lr'],
epoch)
self.logger.info("Start to train weights for epoch %d" % (epoch))
self._training_step(model,
train_w_loader,
self.w_optimizer,
epoch,
info_for_logger="_w_step_")
self.w_scheduler.step()
self.logger.info("Start to train theta for epoch %d" % (epoch))
self._training_step(model,
train_thetas_loader,
self.theta_optimizer,
epoch,
info_for_logger="_theta_step_")
top1_avg = self._validate(model, test_loader, epoch)
if best_top1 < top1_avg:
best_top1 = top1_avg
self.logger.info("Best top1 acc by now. Save model")
save(model, self.path_to_save_model)
self.temperature = self.temperature * self.exp_anneal_rate
def train_loop(self, train_w_loader, train_thetas_loader, test_loader,
model):
best_acc = 9999.9
best_lat = 9999.9
score = 99999.9
# firstly, train weights only
for epoch in range(self.train_thetas_from_the_epoch):
# self.writer.add_scalar('learning_rate/weights', self.w_optimizer.param_groups[0]['lr'], epoch)
self.logger.info("Prepare train %d" % (epoch))
self._training_step(model, train_w_loader, self.w_optimizer)
self.w_scheduler.step()
for epoch in range(self.train_thetas_from_the_epoch, self.cnt_epochs):
# self.writer.add_scalar('learning_rate/weights', self.w_optimizer.param_groups[0]['lr'], epoch)
# self.writer.add_scalar('learning_rate/theta', self.theta_optimizer.param_groups[0]['lr'], epoch)
self.logger.info("Epoch %d" % (epoch))
self.logger.info("Train Weights for epoch %d" % (epoch))
self._training_step(model, train_w_loader, self.w_optimizer)
self.w_scheduler.step()
self.logger.info("Train Thetas for epoch %d" % (epoch))
self._training_step(model, train_thetas_loader,
self.theta_optimizer)
acc, lat = self._validate(model, test_loader)
# import pdb
# pdb.set_trace()
current_score = acc + np.abs(lat - self.target_latency)
if acc < best_acc or lat < best_lat or current_score < score:
if acc < best_acc:
best_acc = acc
if lat < best_lat:
best_lat = lat
if current_score < score:
score = current_score
self.logger.info("Best top1 score by now. Save model")
save(
model, self.path_to_save_model +
"score%.3f_acc%.8f_lat%.3f.pth" % (score, acc, lat))
self.temperature = self.temperature * self.exp_anneal_rate
def train_loop(self, train_loader, valid_loader, model):
best_top1 = 0.0
for epoch in range(self.cnt_epochs):
self.writer.add_scalar('learning_rate', self.optimizer.param_groups[0]['lr'], epoch)
#if epoch and epoch % self.lr_decay_period == 0:
# self.optimizer.param_groups[0]['lr'] *= self.lr_decay
# training
self._train(train_loader, model, epoch)
# validation
top1_avg = self._validate(valid_loader, model, epoch)
if best_top1 < top1_avg:
best_top1 = top1_avg
self.logger.info("Best top1 accuracy by now. Save model")
save(model, self.path_to_save_model)
self.scheduler.step()
def train_loop(self, train_w_loader, train_thetas_loader, test_loader,
model):
global n
best_top1 = 0.0
best_lat = 10000000
best_energy = 10000000
n = 1
# firstly, train weights only
for epoch in range(self.train_thetas_from_the_epoch):
self.writer.add_scalar('learning_rate/weights',
self.w_optimizer.param_groups[0]['lr'],
epoch)
self.logger.info("Firstly, start to train weights for epoch %d" %
(epoch))
self._training_step(model,
train_w_loader,
self.w_optimizer,
epoch,
info_for_logger="_w_step_")
self.w_scheduler.step()
for epoch in range(self.train_thetas_from_the_epoch, self.cnt_epochs):
self.writer.add_scalar('learning_rate/weights',
self.w_optimizer.param_groups[0]['lr'],
epoch)
self.writer.add_scalar('learning_rate/theta',
self.theta_optimizer.param_groups[0]['lr'],
epoch)
self.logger.info("Start to train weights for epoch %d" % (epoch))
self._training_step(model,
train_w_loader,
self.w_optimizer,
epoch,
info_for_logger="_w_step_")
self.w_scheduler.step()
self.logger.info("Start to train theta for epoch %d" % (epoch))
self._training_step(model,
train_thetas_loader,
self.theta_optimizer,
epoch,
info_for_logger="_theta_step_")
top1_avg, top3_avg, lat_avg, energy_avg = self._validate(
model, test_loader, epoch)
#if best_top1 < top1_avg and lat_avg < best_lat:
#if best_top1 < top1_avg: #original
if (best_top1 < top1_avg and lat_avg < best_lat
and energy_avg < best_energy) or best_top1 < top1_avg:
if best_top1 < top1_avg:
best_top1 = top1_avg
print("Best Acc!!")
if lat_avg < best_lat:
best_lat = lat_avg
print("Best Speed!!")
if energy_avg < best_energy:
best_energy = energy_avg
print("Best Energy!!")
self.logger.info("Best top1 acc by now. Save model")
#print("Over Acc: 0.70")
#print("Model Number = " + str(n))
save(model, self.path_to_save_model + str(n) + '.pth')
#n += 1
'''
if (top1_avg >= 0.75) or (top1_avg >= 0.75 and lat_avg < best_lat) or (top1_avg >= 0.75 and energy_avg < best_energy) :
if lat_avg < best_lat:
best_lat = lat_avg
print("Best Latency!!")
if energy_avg < best_energy:
best_energy = energy_avg
print("Best Energy!!")
self.logger.info("Best top1 acc by now. Save model")
print("Over Acc: 0.75")
#print("Model Number = " + str(n))
save(model, self.path_to_save_model + str(n) + '.pth')
#n+=1
'''
self.temperature = self.temperature * self.exp_anneal_rate