def mean_squared_error(y_pred, y):
if utils.isNumpy(y):
y = utils.numpy2var(y)
if utils.isNumpy(y_pred):
y_pred = utils.numpy2var(y_pred)
return torch.sum((y_pred - y)**2)
def accuracy(y_pred, y):
if utils.isNumpy(y):
y = utils.numpy2var(y)
if utils.isNumpy(y_pred):
y_pred = utils.numpy2var(y_pred)
_, y_pred = y_pred.topk(1)
y_pred = torch.squeeze(y_pred).long()
y = torch.squeeze(y).long()
acc = (y_pred == y).sum().data[0]
return acc
def forward_backward(self, XList, y, loss_function=None, verbose=0):
y_pred = self.forward_pass(XList)
if isinstance(y, (np.ndarray, list)):
y = utt.numpy2var(y)
if self.my_optimizer is None:
print "optimizer set..."
self.set_optimizer()
# ----- 2.OPTIMIZER
self.my_optimizer.zero_grad()
if loss_function is None:
loss = loss_dict[self.loss_name](y_pred, *y)
else:
loss = loss_function(y_pred, *y)
loss.backward()
self.my_optimizer.step()
loss_value = loss.data[0]
if verbose:
print "loss: %.3f" % loss_value
return loss_value
def score(self, X, y, batch_size=100, score_function=None):
if score_function is None:
if self.problem_type == "classification":
# Returns accuracy
score_function = "acc"
elif self.problem_type == "regression":
# Returns mean square error
score_function = "mse"
elif self.problem_type == "segmentation":
# Returns Jaccard
score_function = "jacc"
# DEFINE SCORE FUNCTION
score_function = sf.SCORE_DICT[score_function]
if utt.isNumpy(X):
X = utt.numpy2var(X)
n = X.size()[0]
score = 0.
for b in range(0, n, batch_size):
score += score_function(self.forward_pass(X[b:b + batch_size]),
y[b:b + batch_size])
return score / float(n)
def get_heatmap(self, x, output=1):
n, _, n_rows, n_cols = x.shape
x = tu.numpy2var(x)
x = F.relu(F.max_pool2d(self.conv1(x), kernel_size=2))
x = F.relu(F.max_pool2d(self.conv2(x), kernel_size=2))
x = F.max_pool2d(self.conv3(x), kernel_size=2)
x = F.sigmoid(x)
x = tu.get_numpy(x)
return x
def forward_pass(self, XList):
self.start_gpu()
self.train()
if not isinstance(XList, list):
XList = [XList]
if isinstance(XList[0], np.ndarray):
XList = utt.numpy2var(*XList)
if not isinstance(XList, list):
XList = [XList]
y_pred = self(*XList)
return y_pred