def predict(self, inp):
"""
Capsuled prediction method.
Only single model usage supported for now.
"""
inp = torch.Tensor(inp).to(DEVICE)
return self.model(inp)
def evaluate(self, inp, out):
"""Prediction and error estimation for given input and output"""
with torch.no_grad():
# Switch to PyTorch's evaluation mode.
# Some layers, which are used for regularization, e.g., dropout or batch norm layers,
# behave differently, i.e., are turnd off, in evaluation mode
# to prevent influencing the prediction accuracy.
if isinstance(self.model, (list, np.ndarray)):
for idx, __ in enumerate(self.model):
self.model[idx].eval()
else:
self.model.eval()
pred_out = self.model(torch.Tensor(inp).to(DEVICE))
pred_out = torch.sigmoid(pred_out)
# RMSE is the default accuracy metric
error = torch.sqrt(
self.loss(pred_out,
torch.Tensor(out).to(DEVICE)))
return pred_out, (error * 100.0)
def evaluate(self, inp, out):
"""Predition and error estimation for given input and output."""
with torch.no_grad():
# Switch to PyTorch's evaluation mode.
# Some layers, which are used for regularization, e.g., dropout or batch norm layers,
# behave differently, i.e., are turnd off, in evaluation mode
# to prevent influencing the prediction accuracy.
self.model.eval()
pred_out = self.predict(inp)
# RMSE is the default accuracy metric
error = torch.sqrt(
self.loss(pred_out,
torch.Tensor(out).to(DEVICE)))
return pred_out, (error * 100.0)
def learn_from_epoch(self):
"""Training method"""
epoch_loss = 0
try:
inp_batches, out_batches = self.get_batches_fn()
except AttributeError:
print("Error: No nb_batches_fn defined in preprocessor. "
"This attribute is required by the training routine.")
for batch_idx in tqdm(range(len(inp_batches))):
pred_out = self.model(
torch.Tensor(inp_batches[batch_idx]).to(DEVICE))
pred_out = torch.sigmoid(pred_out)
batch_loss = self.loss(
pred_out,
torch.Tensor(out_batches[batch_idx]).to(DEVICE))
self.optimizer.zero_grad()
batch_loss.backward()
self.optimizer.step()
epoch_loss += batch_loss.item()
epoch_loss /= len(inp_batches)
return epoch_loss
def learn_from_epoch(self):
"""Training method."""
epoch_loss = 0
nb_scenarios = 0
inp_batches, out_batches = self.get_batches_fn()
for batch_idx in tqdm(range(len(inp_batches))):
pred_out = self.predict(inp_batches[batch_idx])
batch_loss = self.loss(
pred_out,
torch.Tensor(out_batches[batch_idx]).to(DEVICE))
self.optimizer.zero_grad()
batch_loss.backward()
self.optimizer.step()
epoch_loss += batch_loss.item()
epoch_loss /= nb_scenarios
return epoch_loss