def unseen_phase(self):
self.model.eval()
losses = []
preds = torch.empty(0, 9).long().to(self.device)
acts = torch.empty(0, 9).long().to(self.device)
for i, batch in enumerate(dl):
with torch.no_grad():
inputs, labels = batch['image'], batch['labels'].long().view(
-1, 9)
inputs, labels = inputs.to(device), labels.to(device)
pred = nn.Softmax(dim=1)
outputs = self.model(inputs)
outputs = pred(outputs)
preds = torch.cat((preds, outputs), 0)
acts = torch.cat((acts, labels.long()), 0)
for j in range(0, batch['labels'].size()[0]):
if torch.max(outputs, 1)[1][j] == torch.max(labels,
1)[1][j]:
losses.append(1)
else:
losses.append(0)
del inputs
del labels
del batch
torch.cuda.empty_cache()
epoch_loss = (sum(losses) * 100.00) / len(losses)
return epoch_loss, preds, acts
def train_phase(self, tr='train'):
if tr == 'train':
self.model.train()
dl = self.train_dl
if tr == 'val':
self.model.eval()
dl = self.val_dl
losses = []
for i, batch in enumerate(dl):
inputs, labels = batch['image'], batch['labels'].long().view(-1, 9)
inputs, labels = inputs.to(self.device), labels.to(self.device)
labels = torch.argmax(labels, 1)
self.optimizer.zero_grad()
if self.loss in ['normsoftmax', 'panc', 'pnca']:
self.loss_optimizer.zero_grad()
embeddings = self.model(inputs)
indices_tuple = self.mining_func(embeddings, labels)
eloss = self.loss_function(embeddings, labels, indices_tuple)
losses.append(eloss.item())
eloss.backward()
if self.loss in ['normsoftmax', 'panc', 'pnca']:
self.loss_optimizer.step()
self.optimizer.step()
del inputs
del labels
del embeddings
torch.cuda.empty_cache()
epoch_loss = sum(losses) / len(losses)
return epoch_loss
def train_phase(self, tr='train'):
if tr == 'train':
self.model.train()
dl = self.train_dl
if tr == 'val':
self.model.eval()
dl = self.val_dl
losses = []
for i, batch in enumerate(dl):
inputs, labels = batch['image'], batch['labels'].long().view(-1,self.hubble_classes)
inputs, labels = inputs.to(self.device), labels.to(self.device)
self.optimizer.zero_grad()
outputs = self.model(inputs)
loss = self.criterion(outputs, torch.max(labels, 1)[1])
losses.append(loss.item())
loss = torch.sqrt(loss)
loss.backward()
self.optimizer.step()
del batch
del inputs
del labels
torch.cuda.empty_cache()
epoch_loss = sum(losses) / len(losses)
return epoch_loss
def train_phase(self, tr='train'):
if tr == 'train':
self.model.train()
dl = self.train_dl
if tr == 'val':
self.model.eval()
dl = self.val_dl
losses = []
for i, batch in enumerate(dl):
inputs, labels = batch['image'], batch['labels'].long().view(-1, 9)
inputs, labels = inputs.to(self.device), labels.to(self.device)
self.optimizer.zero_grad() # 1. Zero the parameter gradients
pred = nn.Softmax(dim=1)
outputs = self.model(inputs)
loss = self.criterion(outputs, torch.max(labels, 1)[1])
losses.append(loss.item())
loss.backward()
self.optimizer.step()
del inputs
del labels
del outputs
torch.cuda.empty_cache()
epoch_loss = sum(losses) / len(losses)
return epoch_loss