def fit_eval(self):
train, test = self.data.train_test_split()
loss = self.model.learn(train[0], train[1])
preds = self.model.predict(test[0])
crit = self.model.criterion()
val_loss = crit(to_torch(preds), to_torch(test[1])).item()
acc = -1
try:
acc = accuracy_score(test[1], preds)
except:
pass
out = {'accuracy': acc, 'training_loss': loss, 'val_loss': val_loss}
out['val_vc'] = np.unique(test[1], return_counts=True)
out['val_pred_vc'] = np.unique(preds, return_counts=True)
return out
def learn(self, x, y):
x, y = to_torch(x), to_torch(y)
criterion = self.criterion()
optimizer = self.optimizer(self.parameters(),
lr=self.lr,
weight_decay=self.wd)
closs = 100000
for i in range(self.tr_epochs):
print(i, "start")
pred = self.forward(x)
loss = criterion(pred, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
closs = loss.item()
print(i, closs, "\n=============")
return closs
def predict(self, x):
x = to_torch(x)
out = None
with no_grad():
out = self.forward(x).numpy()
return out