def __call__(self, y_true, y_pred): # There are additional parameters for this function # Note: some of the 'modes' for edge behavior do not yet have a gradient definition in the Theano tree # and cannot be used for learning kernel = [self.kernel_size, self.kernel_size] y_true = KC.reshape(y_true, [-1] + list(self.__int_shape(y_pred)[1:])) y_pred = KC.reshape(y_pred, [-1] + list(self.__int_shape(y_pred)[1:])) patches_pred = KC.extract_image_patches(y_pred, kernel, kernel, 'valid', self.dim_ordering) patches_true = KC.extract_image_patches(y_true, kernel, kernel, 'valid', self.dim_ordering) # Reshape to get the var in the cells bs, w, h, c1, c2, c3 = self.__int_shape(patches_pred) patches_pred = KC.reshape(patches_pred, [-1, w, h, c1 * c2 * c3]) patches_true = KC.reshape(patches_true, [-1, w, h, c1 * c2 * c3]) # Get mean u_true = KC.mean(patches_true, axis=-1) u_pred = KC.mean(patches_pred, axis=-1) # Get variance var_true = K.var(patches_true, axis=-1) var_pred = K.var(patches_pred, axis=-1) # Get std dev covar_true_pred = K.mean(patches_true * patches_pred, axis=-1) - u_true * u_pred ssim = (2 * u_true * u_pred + self.c1) * (2 * covar_true_pred + self.c2) denom = (K.square(u_true) + K.square(u_pred) + self.c1) * (var_pred + var_true + self.c2) ssim /= denom # no need for clipping, c1 and c2 make the denom non-zero return K.mean((1.0 - ssim) / 2.0)
def total_variation_loss(x): assert K.ndim(x) == 4 if K.image_data_format() == 'channels_first': a = K.square(x[:, :, :img_h - 1, :img_w - 1] - x[:, :, 1:, :img_w - 1]) b = K.square(x[:, :, :img_h - 1, :img_w - 1] - x[:, :, :img_h - 1, 1:]) else: # Move the image pixel by pixel, and calculate the variance a = K.square(x[:, :img_h - 1, :img_w - 1, :] - x[:, 1:, :img_w - 1, :]) b = K.square(x[:, :img_h - 1, :img_w - 1, :] - x[:, :img_h - 1, 1:, :]) return K.sum(K.pow(a + b, 1.25))
def _get_coords_for_joint(joint_idx, parent_idx, child_angle_idx, coords): if parent_idx is None: # joint_idx should be 0 coords[joint_idx] = K.zeros(base_shape[:-2] + [3, 1]) parent_bone = K.constant( np.concatenate([ np.ones(base_shape), np.zeros(base_shape), np.zeros(base_shape) ], axis=-2)) else: parent_bone = coords[parent_idx] - coords[joint_idx] parent_bone_norm = K.sqrt( K.sum(K.square(parent_bone), axis=-2, keepdims=True) + K.epsilon()) parent_bone = parent_bone / parent_bone_norm for child_idx in body_graph[joint_idx]: child_bone = tf.matmul(rotmat_list[child_angle_idx], parent_bone) child_bone_idx = bone_idcs[(joint_idx, child_idx)] child_bone = child_bone * K.reshape( bone_len_list[child_bone_idx], (child_bone.shape[0], 1, 1, 1)) coords[child_idx] = child_bone + coords[joint_idx] child_angle_idx += 1 for child_idx in body_graph[joint_idx]: child_angle_idx, coords = _get_coords_for_joint( child_idx, joint_idx, child_angle_idx, coords) return child_angle_idx, coords
def _get_bone_len(arg): bone_list = tf.unstack(arg[:, :, 0, :], axis=1) bones = [ bone_list[j] - bone_list[i] for i, j in zip(members_from, members_to) ] bones = K.stack(bones, axis=1) return K.sqrt(K.sum(K.square(bones), axis=-1) + K.epsilon())
def normSAD2(y_true, y_pred): y_true2 = K.l2_normalize(y_true + K.epsilon(), axis=-1) y_pred2 = K.l2_normalize(y_pred + K.epsilon(), axis=-1) mse = K.mean(K.square(y_true - y_pred), axis=-1) # sad = -K.log(1.0-K.mean(y_true2 * y_pred2/np.pi, axis=-1)) sad = K.mean(y_true2 * y_pred2, axis=-1) # sid = SID(y_true,y_pred) return 0.005 * mse - 0.75 * sad
def style_loss(style, gen): assert K.ndim(style) == 3 assert K.ndim(gen) == 3 S = gram_matrix(style) G = gram_matrix(gen) channels = 3 size = img_h * img_w # Euclidean distance of the gram matrices multiplied by the constant return K.sum(K.square(S - G)) / (4. * (channels**2) * (size**2))
def normSAD2(y_true, y_pred): # y_true2 = K.l2_normalize(y_true + K.epsilon(), axis=-1) # y_pred2 = K.l2_normalize(y_pred + K.epsilon(), axis=-1) mse = K.mean(K.square(y_true - y_pred)) sad = SAD(y_true, y_pred) # sad = -K.log(1.0-SAD(y_true, y_pred)/np.pi) # sid = SID(y_true,y_pred) # return 0.005 * mse + 0.75 * sad return 0.005 * mse + 10.0 * sad
def normSAD(y_true, y_pred): # y_true2 = K.l2_normalize(y_true + K.epsilon(), axis=-1) # y_pred2 = K.l2_normalize(y_pred + K.epsilon(), axis=-1) mse = K.mean(K.square(y_true - y_pred)) # sad = -K.log(1.0-K.mean(y_true2 * y_pred2/np.pi, axis=-1)) sad = SAD(y_true, y_pred) # sid = SID(y_true,y_pred) # return 0.008*mse-1.0*sad return 0.008 * mse + 1.0 * sad
def _get_avg_bone_len(arg): bone_list = tf.unstack(arg[:, :, 0, :], axis=1) bones = [ bone_list[j] - bone_list[i] for i, j in zip(members_from, members_to) ] bones = K.expand_dims(K.stack(bones, axis=1), axis=2) bone_len = K.sqrt( K.sum(K.square(bones), axis=-1, keepdims=True) + K.epsilon()) return K.mean(bone_len, axis=1, keepdims=True)
def critic_optimizer(self): discounted_prediction = K.placeholder(shape=(None, )) value = self.critic.output # loss = MSE(discounted_prediction, value) loss = K.mean(K.square(discounted_prediction - value)) optimizer = Adam(lr=self.critic_lr) updates = optimizer.get_updates(loss, self.critic.trainable_weights) train = K.function([self.critic.input, discounted_prediction], [loss], updates=updates) return train
def MSE_KL(y_true, y_pred): # y_true=y_true[:,-162:] y_true = K.switch( K.min(y_true) < 0, y_true - K.min(y_true) + K.epsilon(), y_true + K.epsilon()) y_pred = K.switch( K.min(y_pred) < 0, y_pred - K.min(y_pred) + K.epsilon(), y_pred + K.epsilon()) p_n = y_true / K.max(y_true, axis=1, keepdims=True) q_n = y_pred / K.max(y_pred, axis=1, keepdims=True) return K.mean(K.square(y_true - y_pred), axis=-1) + 0.5 * (K.sum(p_n * K.log(p_n / q_n)) + K.sum( (1.001 - p_n) * K.log((1.01 - p_n) / (1.001 - q_n))))
def add_loss(model, W): inputs = model.inputs[0] abnormal = model.inputs[1] # abnormal = K.print_tensor(abnormal, message='abnormal = ') outputs = model.outputs[0] z_mean = model.get_layer('z_mean').output z_log_var = model.get_layer('z_log_var').output beta = K.sum(1.0 - abnormal, axis=-1, keepdims=True) / W # beta = K.print_tensor(beta, message='beta = ') reconstruction_loss = mean_squared_error(inputs, outputs) reconstruction_loss *= W kl_loss = 1 + z_log_var - beta * K.square(z_mean) - K.exp(z_log_var) kl_loss = K.sum(kl_loss, axis=-1) kl_loss *= -0.5 vae_loss = K.mean(reconstruction_loss + kl_loss) model.add_loss(vae_loss)
def tsne(P, activations): # d = K.shape(activations)[1] v = d - 1. eps = K.variable( 10e-15 ) # needs to be at least 10e-8 to get anything after Q /= K.sum(Q) sum_act = K.sum(K.square(activations), axis=1) Q = K.reshape(sum_act, [-1, 1]) + -2 * K.dot(activations, K.transpose(activations)) Q = (sum_act + Q) / v Q = K.pow(1 + Q, -(v + 1) / 2) Q *= K.variable(1 - np.eye(n)) Q /= K.sum(Q) Q = K.maximum(Q, eps) C = K.log((P + eps) / (Q + eps)) C = K.sum(P * C) return C
def rmse(y_true, y_pred): return backend.sqrt(backend.mean(backend.square(y_pred - y_true), axis=-1))
def content_loss(content, gen): assert K.ndim(content) == 3 assert K.ndim(gen) == 3 # Euclidean distance return K.sum(K.square(gen - content))
def __call__(self, p): p *= K.cast(p >= 0., K.floatx()) return p / (K.epsilon() + K.sqrt(K.sum(K.square(p), axis=self.axis, keepdims=True)))
def edm(x, y=None): with K.name_scope('edm'): y = x if y is None else y x = K.expand_dims(x, axis=1) y = K.expand_dims(y, axis=2) return K.sqrt(K.sum(K.square(x - y), axis=-1) + K.epsilon())
def edm_loss(y_true, y_pred): return K.mean(K.sum(K.square(edm(y_true) - edm(y_pred)), axis=[1, 2]))
def normMSE(y_true, y_pred): y_true2 = K.l2_normalize(y_true + K.epsilon(), axis=-1) y_pred2 = K.l2_normalize(y_pred + K.epsilon(), axis=-1) mse = K.mean(K.square(y_true - y_pred)) return mse