def evaluate_mae(self, X_test, y_test, device, batch_size=4096):
X_test = torch.as_tensor(X_test, device=device)
y_test = check_numpy(y_test)
self.model.train(False)
with torch.no_grad():
prediction = process_in_chunks(self.model,
X_test,
batch_size=batch_size)
prediction = check_numpy(prediction)
error_rate = np.abs(y_test - prediction).mean()
return error_rate
def evaluate_logloss(self, X_test, y_test, device, batch_size=512):
X_test = torch.as_tensor(X_test, device=device)
y_test = check_numpy(y_test)
self.model.train(False)
with torch.no_grad():
logits = F.softmax(process_in_chunks(self.model,
X_test,
batch_size=batch_size),
dim=1)
logits = check_numpy(logits)
y_test = torch.from_numpy(y_test)
logloss = log_loss(check_numpy(to_one_hot(y_test)), logits)
return logloss
def evaluate_classification_error(self,
X_test,
y_test,
device,
batch_size=4096):
X_test = torch.as_tensor(X_test, device=device)
y_test = check_numpy(y_test)
self.model.train(False)
with torch.no_grad():
logits = process_in_chunks(self.model,
X_test,
batch_size=batch_size)
logits = check_numpy(logits)
error_rate = (y_test != np.argmax(logits, axis=1)).mean()
return error_rate