def validation_step(self, batch, batch_idx):
if self.dataset_name == 'stl10':
labeled_batch = batch[1]
unlabeled_batch = batch[0]
batch = unlabeled_batch
(img_1, img_2), y = batch
h1, z1 = self.forward(img_1)
h2, z2 = self.forward(img_2)
loss = self.loss_func(z1, z2, self.temp)
result = {'val_loss': loss}
if self.online_evaluator:
if self.dataset_name == 'stl10':
(img_1, img_2), y = labeled_batch
h1, z1 = self.forward(img_1)
z_in = z1.reshape(z1.size(0), -1)
mlp_preds = self.non_linear_evaluator(z_in)
mlp_loss = F.cross_entropy(mlp_preds, y)
acc = metrics.accuracy(mlp_preds, y)
result['mlp_acc'] = acc
result['mlp_loss'] = mlp_loss
return result
def validation_step(self, batch, batch_nb):
# in STL10 we pass in both lab+unl for online ft
if self.hparams.dataset == 'stl10':
labeled_batch = batch[1]
unlabeled_batch = batch[0]
batch = unlabeled_batch
img_1, y = batch
# generate features
# Latent features
Z = self(img_1)
# infoNCE loss
nce_loss = self.info_nce(Z)
result = {'val_nce': nce_loss}
if self.online_evaluator:
if self.hparams.dataset == 'stl10':
img_1, y = labeled_batch
Z = self(img_1)
z_in = Z.reshape(Z.size(0), -1)
mlp_preds = self.non_linear_evaluator(z_in)
mlp_loss = F.cross_entropy(mlp_preds, y)
acc = metrics.accuracy(mlp_preds, y)
result['mlp_acc'] = acc
result['mlp_loss'] = mlp_loss
return result
def validation_step(self, batch, batch_idx):
if isinstance(self.datamodule, STL10DataModule):
labeled_batch = batch[1]
unlabeled_batch = batch[0]
batch = unlabeled_batch
(img_1, img_2), y = batch
h1, z1 = self.forward(img_1)
h2, z2 = self.forward(img_2)
loss = self.loss_func(z1, z2, self.hparams.loss_temperature)
result = {'val_loss': loss}
if self.online_evaluator:
if isinstance(self.datamodule, STL10DataModule):
(img_1, img_2), y = labeled_batch
h1, z1 = self.forward(img_1)
z_in = z1.reshape(z1.size(0), -1)
mlp_preds = self.non_linear_evaluator(z_in)
mlp_loss = F.cross_entropy(mlp_preds, y)
acc = metrics.accuracy(mlp_preds, y)
result['mlp_acc'] = acc
result['mlp_loss'] = mlp_loss
return result
def test_step(self, batch, batch_idx):
img_1, y = batch
Z = self(img_1)
z_in = Z.reshape(Z.size(0), -1)
mlp_preds = self.non_linear_evaluator(z_in)
mlp_loss = F.cross_entropy(mlp_preds, y)
acc = metrics.accuracy(mlp_preds, y)
return {'acc': acc, 'loss': mlp_loss}