def evaluate(self, save_name=None):
if save_name is None:
model = self.model.cuda()
else:
save_path = "%s%s.mdl" % (PathConfig.MODEL_PATH, save_name)
print("Loading model from %s..." % save_path)
model = torch.load(save_path).cuda()
print("Loaded.\n")
total = ceil(len(self.data["test"][0]) /
NetworkConfig.BATCH_SIZE) * NetworkConfig.BATCH_SIZE
correct_count = 0
bar = ProgressBar(total, "Testing model...")
progress = 0
for batch in Data.batches(self.data["test"]):
X, Y = batch[0], batch[1]
output = model(X)
output = numpy.round(output.cpu().detach().numpy() * 5)
Y = numpy.round(numpy.array(Y) * 5)
correct_count += numpy.sum(output == Y)
progress += NetworkConfig.BATCH_SIZE
bar.refresh(progress)
bar.finish("Accuracy: %f" % (correct_count / total))
def optimize(self, save_name=None):
best_model = self.model
min_loss = float("inf")
total = 0
for phase in Network.Phase:
total += NetworkConfig.EPOCH * \
ceil(len(self.data[str(phase)][0]) / NetworkConfig.BATCH_SIZE) * NetworkConfig.BATCH_SIZE
bar = ProgressBar(total, "Training model...")
progress = 0
for epoch in range(NetworkConfig.EPOCH):
for phase in Network.Phase:
if phase is Network.Phase.TRAIN:
if self.lr_scheduler:
self.model = self.lr_scheduler(self.model,
NetworkConfig.EPOCH)
self.model.train()
else:
self.model.eval()
running_loss = 0.0
for batch in Data.batches(self.data[str(phase)]):
self.optimizer.zero_grad()
X, Y = batch[0], torch.Tensor(batch[1]).cuda()
output = self.model(X)
loss = self.loss_function(output, Y)
if phase is Network.Phase.TRAIN:
loss.backward()
self.optimizer.step()
progress += NetworkConfig.BATCH_SIZE
bar.refresh(progress)
running_loss += loss.data.cpu().numpy()
epoch_loss = running_loss / len(self.data[str(phase)])
if phase is Network.Phase.TRIAL and epoch_loss < min_loss:
best_model = copy.deepcopy(self.model)
min_loss = epoch_loss
bar.finish(
"Model has been optimized in %d epochs with %f as its minimal loss."
% (NetworkConfig.EPOCH, min_loss))
if save_name is None:
save_name = time.strftime("%Y%m%d_%H%M%S")
save_path = "%s%s.mdl" % (PathConfig.MODEL_PATH, save_name)
print("Saving model into %s..." % save_path)
torch.save(best_model, save_path)
print("Saved.\n")
self.model = best_model
