def loss_func(y_true, y_pred_mll):
y_true = y_true[:, 0]
y_pred = y_pred_mll[:, 0]
mll_pred = y_pred_mll[:, 1]
mll_loss = K.mean(K.abs(mll_pred - 91.2))
mll_sigma_loss = K.abs(K.std(mll_pred) - 7.67)
return binary_crossentropy(
y_true, y_pred) + c * mll_loss + c * mll_sigma_loss
def _compute_cost_huber(self, q, a, r, t, q2):
preds = slice_tensor_tensor(q, a)
bootstrap = K.max if not self.use_mean else K.mean
targets = r + (1 - t) * self.gamma * bootstrap(q2, axis=1)
err = targets - preds
cond = K.abs(err) > 1.0
L2 = 0.5 * K.square(err)
L1 = (K.abs(err) - 0.5)
cost = tf.where(cond, L2, L1)
return K.mean(cost)
def custom_mean_squared_loss(y_true, y_pred):
print(y_true)
print(y_pred)
diff = K.abs(y_true - y_pred)
angle_diff = K.minimum(diff[:, :, 6:], 360 - diff[:, :, 6:])
error = tf.concat([diff[:, :, :6], angle_diff], axis=-1)
return K.mean(K.square(error), axis=-1)
def confidence_reconstruction_loss(y_true, y_pred, mask, num_steps,
gaussian_kernel_size, gaussian_kernel_std):
mask_blurred = gaussian_utils.blur_mask(mask, num_steps,
gaussian_kernel_size,
gaussian_kernel_std)
valid_mask = 1 - mask
diff = K.abs(y_true - y_pred)
l1 = K.mean(diff * valid_mask + diff * mask_blurred, axis=[1, 2, 3])
return l1
def loss_func(y_true, y_pred_mll):
y_true = y_true[:, 0]
y_pred = y_pred_mll[:, 0]
mll_pred = y_pred_mll[:, 1]
mll_loss = K.mean(K.abs(mll_pred - 91.2))
# pseudomll = K.random_normal_variable(shape=(1,1), mean=91.2, scale=2)
# mll_loss = K.mean((mll_pred - pseudomll)**2)
return binary_crossentropy(y_true, y_pred) + c * mll_loss
def reconstruction_loss(y_true, y_pred):
diff = K.abs(y_pred - y_true)
l1 = K.mean(diff, axis=[1, 2, 3])
return l1
def jaccard(x):
return x[0] * x[1] / (
K.sum(x[0]**2, axis=1, keepdims=True) +
K.sum(x[1]**2, axis=1, keepdims=True) -
K.sum(K.abs(x[0] * x[1]), axis=1, keepdims=True))
