Python keras.backend.update_add() Examples

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = self.learning_rate * (K.sqrt(1. - K.pow(self.beta_2, t)) / (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            p_t = lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            self.updates.append(K.update_sub(p, p_t))
        return self.updates 

def __call__(self, y_true, y_pred):
        y_true = K.cast(K.round(y_true), "int32")
        y_pred = K.cast(K.round(y_pred), "int32")
        neg_y_pred = 1 - y_pred

        tp = K.sum(K.transpose(y_true * y_pred), axis=-1)
        fn = K.sum(K.transpose(y_true * neg_y_pred), axis=-1)

        current_tp = K.cast(self.tp + tp, self.epsilon.dtype)
        current_fn = K.cast(self.fn + fn, self.epsilon.dtype)

        tp_update = K.update_add(self.tp, tp)
        fn_update = K.update_add(self.fn, fn)

        self.add_update(tp_update, inputs=[y_true, y_pred])
        self.add_update(fn_update, inputs=[y_true, y_pred])

        return K.mean(truediv(current_tp, current_tp + current_fn + self.epsilon)) 

def __call__(self, y_true, y_pred):
        y_true, y_pred = self.cast(y_true, y_pred)
        neg_y_pred = 1 - y_pred

        fn = K.sum(y_true * neg_y_pred)
        current_fn = self.fn * 1

        fn_update = K.update_add(self.fn, fn)
        self.add_update(fn_update, inputs=[y_true, y_pred])

        return fn + current_fn 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr *= (1. / (1. + self.decay * K.cast(self.iterations,
                                                  K.dtype(self.decay))))
        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + moments
        for p, g, m in zip(params, grads, moments):
            
            matched_layer = [x for x in self.lr_multipliers.keys() if x in p.name]
            if matched_layer:
                new_lr = lr * self.lr_multipliers[matched_layer[0]]
            else:
                new_lr = lr

            v = self.momentum * m - new_lr * g  # velocity
            self.updates.append(K.update(m, v))

            if self.nesterov:
                new_p = p + self.momentum * v - new_lr * g
            else:
                new_p = p + v

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr *= (1. / (1. + self.decay * K.cast(self.iterations,
                                                  K.dtype(self.decay))))
        
        accum_switch = K.equal(self.iterations % self.accum_iters, 0)
        print(accum_switch)
        accum_switch = K.cast(accum_switch, dtype='float32')

        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape) for shape in shapes]
        temp_grads = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + moments
        for p, cg, m, tg in zip(params, grads, moments, temp_grads):
            g = cg + tg
            v = self.momentum * m - (lr * g / self.accum_iters)  # velocity
            self.updates.append(K.update(m, (1 - accum_switch) * m + accum_switch * v))
            self.updates.append(K.update(tg, (1 - accum_switch) * g))

            if self.nesterov:
                new_p = p + self.momentum * v - (lr * g / self.accum_iters)
            else:
                new_p = p + v

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, (1 - accum_switch) * p + accum_switch * new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr *= (1. / (1. + self.decay * K.cast(self.iterations,
                                                  K.dtype(self.decay))))
        
        accum_switch = K.equal(self.iterations % self.accum_iters, 0)
        print(accum_switch)
        accum_switch = K.cast(accum_switch, dtype='float32')

        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape) for shape in shapes]
        temp_grads = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + moments
        for p, cg, m, tg in zip(params, grads, moments, temp_grads):
            g = cg + tg
            v = self.momentum * m - (lr * g / self.accum_iters)  # velocity
            self.updates.append(K.update(m, (1 - accum_switch) * m + accum_switch * v))
            self.updates.append(K.update(tg, (1 - accum_switch) * g))

            if self.nesterov:
                new_p = p + self.momentum * v - (lr * g / self.accum_iters)
            else:
                new_p = p + v

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, (1 - accum_switch) * p + accum_switch * new_p))
        return self.updates 

def call(self, inputs, training=None):
        noise_shape = self._get_noise_shape(inputs)
        t = K.cast(self.iterations, K.floatx()) + 1
        p = t / float(self.decay_interval)

        keep_rate = self.initial_keep_rate * K.pow(self.decay_rate, p)

        def dropped_inputs():
            self.add_update([K.update_add(self.iterations, [1])], inputs)
            return K.dropout(inputs, 1 - keep_rate[0], noise_shape, seed=self.seed)
        return K.in_train_phase(dropped_inputs, inputs, training=training) 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr *= (1. / (1. + self.decay * K.cast(self.iterations,
                                                  K.dtype(self.decay))))
        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + moments
        for p, g, m in zip(params, grads, moments):
            v = self.momentum * m + g  # velocity
            self.updates.append(K.update(m, v))

            if self.nesterov:
                new_p = p - lr * (self.momentum * v + g)
            else:
                new_p = p - lr * v

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]
        wd = self.wd # decoupled weight decay (3/4)

        lr = self.lr
        if self.initial_decay > 0:
            lr *= (1. / (1. + self.decay * K.cast(self.iterations,
                                                  K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                     (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon) - lr * wd * p # decoupled weight decay (4/4)

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        tower_gradvars = []
        gdev_list = self._gdev_list

        global_scope = tf.get_variable_scope()
        for idev, device in enumerate(gdev_list):
            with tf.device(device), \
                    tf.variable_scope(global_scope, reuse=idev > 0), \
                    tf.name_scope('tower_%i' % idev):
                grads = self.optimizer.compute_gradients(loss, params)

            gradvars = zip(grads, params)
            tower_gradvars.append(gradvars)

        tower_gradvars = all_avg_gradients(tower_gradvars,
                                           gdev_list,
                                           usenccl=False)

        self.updates = [K.update_add(self.iterations, 1)]

        for device_num, device in enumerate(gdev_list):
            with tf.device(device):
                gradvars = tower_gradvars[device_num]
                opt_update = self.optimizer.apply_gradients(
                    grads, global_step=self.iterations)
            self.updates.append(opt_update)

        return self.updates 

def get_updates(self, params, constraints, loss):
        grads = self.get_gradients(loss, params)
        self.updates = []
        self.updates.append(K.update_add(self.iterations, 1))
        for p, g in zip(params, grads):
            self.updates.append((p, p - self.lr * g))
        return self.updates 

def __call__(self, y_true, y_pred):
        y_true, y_pred = self.cast(y_true, y_pred)
        neg_y_true = 1 - y_true

        fp = K.sum(neg_y_true * y_pred)
        current_fp = self.fp * 1

        fp_update = K.update_add(self.fp, fp)
        self.add_update(fp_update, inputs=[y_true, y_pred])

        return fp + current_fp 

def __call__(self, y_true, y_pred):
        y_true, y_pred = self.cast(y_true, y_pred)

        neg_y_true = 1 - y_true
        neg_y_pred = 1 - y_pred

        tn = K.sum(neg_y_true * neg_y_pred)
        current_tn = self.tn * 1

        tn_update = K.update_add(self.tn, tn)
        self.add_update(tn_update, inputs=[y_true, y_pred])

        return tn + current_tn 

def __call__(self, y_true, y_pred):
        y_true, y_pred = self.cast(y_true, y_pred)

        tp = K.sum(y_true * y_pred)
        current_tp = self.tp * 1

        tp_update = K.update_add(self.tp, tp)
        self.add_update(tp_update, inputs=[y_true, y_pred])

        return tp + current_tp 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                     (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vhats = [K.zeros(1) for _ in params]
        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
            g2 = K.square(g)
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = v - (1. - self.beta_2) * K.sign(v - g2) * g2
            p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.inital_decay > 0:
            lr *= (1. / (1. + self.decay * self.iterations))

        t = self.iterations + 1

        lr_t = lr / (1. - K.pow(self.beta_1, t))

        shapes = [K.int_shape(p) for p in params]
        zs = [K.zeros(shape) for shape in shapes]
        vs = [K.zeros(shape) for shape in shapes]
        ds = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + zs + vs + ds

        for p, g, z, v, d in zip(params, grads, zs, vs, ds):
            v_t = self.beta_2 * v + (1. - self.beta_2) * K.square(g)
            d_t = (K.sqrt(v_t / (1. - K.pow(self.beta_2, t)))
                   + self.epsilon) / lr_t
            sigma_t = d_t - self.beta_1 * d
            z_t = self.beta_1 * z + (1. - self.beta_1) * g - sigma_t * p

            p_t = - z_t / d_t

            self.updates.append(K.update(z, z_t))
            self.updates.append(K.update(v, v_t))
            self.updates.append(K.update(d, d_t))

            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        weights = self.get_weights()
        self.updates = [K.update_add(self.iterations, 1)]
        scaled_lr = self.lr
        w_norm = K.sqrt(K.sum([K.sum(K.square(weight))
                               for weight in weights]))
        g_norm = K.sqrt(K.sum([K.sum(K.square(grad))
                               for grad in grads]))
        scaled_lr = K.switch(K.greater(w_norm * g_norm, K.zeros([1])),
                             K.expand_dims((self.eeta * w_norm /
                                            (g_norm + self.weight_decay * w_norm +
                                             self.epsilon)) * self.lr),
                             K.ones([1]) * self.lr)
        if K.backend() == 'theano':
            scaled_lr = scaled_lr[0]  # otherwise theano raise broadcasting error
        # momentum
        moments = [K.zeros(K.int_shape(param), dtype=K.dtype(param))
                   for param in params]
        self.weights = [self.iterations] + moments
        for param, grad, moment in zip(params, grads, moments):
            v0 = (moment * self.momentum)
            v1 = scaled_lr * grad  # velocity
            veloc = v0 - v1
            self.updates.append(K.update(moment, veloc))

            if self.nesterov:
                new_param = param + (veloc * self.momentum) - v1
            else:
                new_param = param + veloc

            # Apply constraints.
            if getattr(param, 'constraint', None) is not None:
                new_param = param.constraint(new_param)

            self.updates.append(K.update(param, new_param))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        beta_1_t = K.pow(self.beta_1, t)
        beta_2_t = K.pow(self.beta_2, t)
        rho = 2 / (1 - self.beta_2) - 1
        rho_t = rho - 2 * t * beta_2_t / (1 - beta_2_t)
        r_t = K.sqrt(
            K.relu(rho_t - 4) * K.relu(rho_t - 2) * rho / ((rho - 4) * (rho - 2) * rho_t)
        )
        flag = K.cast(rho_t > 4, K.floatx())

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            mhat_t = m_t / (1 - beta_1_t)
            vhat_t = K.sqrt(v_t / (1 - beta_2_t))
            p_t = p - lr * mhat_t * (flag * r_t / (vhat_t + self.epsilon) + (1 - flag))

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
    grads = self.get_gradients(loss, params)
    self.updates = [K.update_add(self.iterations, 1)]

    lr = self.lr
    if self.initial_decay > 0:
      lr *= (1. / (1. + self.decay * K.cast(self.iterations, K.dtype(self.decay))))

    t = K.cast(self.iterations, K.floatx()) + 1
    lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /(1. - K.pow(self.beta_1, t)))

    ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
    vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
    vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params] 
    self.weights = [self.iterations] + ms + vs + vhats

    for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
      m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
      v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
      vhat_t = K.maximum(vhat, v_t)
      p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)

      self.updates.append(K.update(m, m_t))
      self.updates.append(K.update(v, v_t))
      self.updates.append(K.update(vhat, vhat_t))
      new_p = p_t

      # Apply constraints.
      if getattr(p, 'constraint', None) is not None:
        new_p = p.constraint(new_p)

      self.updates.append(K.update(p, new_p))
    return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.learning_rate
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))
        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape, name='moment_' + str(i))
                   for (i, shape) in enumerate(shapes)]
        self.weights = [self.iterations] + moments

        for p, g, m in zip(params, grads, moments):
            # Learning rate multipliers
            lr_t = self.learning_rate
            if self.lr_multipliers is not None:
                lr_t = _apply_lr_multiplier(self, lr_t, p)

            v = self.momentum * m - self.eta_t * lr_t * g  # velocity
            self.updates.append(K.update(m, v))

            if self.nesterov:
                p_t = p + self.momentum * v - self.eta_t * lr_t * g
            else:
                p_t = p + v

            # Weight decays
            if p.name in self.weight_decays.keys():
                p_t = _apply_weight_decays(self, p, p_t)
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)
            self.updates.append(K.update(p, new_p))

        # Cosine annealing
        _update_t_cur_eta_t(self)
        self.lr_t = lr_t * self.eta_t  # for external tracking

        self._init_notified = True
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                     (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]

        if self.amsgrad:
            vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        else:
            vhats = [K.zeros(1) for _ in params]
        self.weights = [self.iterations] + ms + vs + vhats

        lr_t_premult = lr_t
        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
            # Learning rate multipliers
            if self.lr_multipliers is not None:
                lr_t = _apply_lr_multiplier(self, lr_t_premult, p)

            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))

            if self.amsgrad:
                vhat_t = K.maximum(vhat, v_t)
                p_t = p - self.eta_t * lr_t * m_t / (
                    K.sqrt(vhat_t) + self.epsilon)
                self.updates.append(K.update(vhat, vhat_t))
            else:
                p_t = p - self.eta_t * lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

            # Weight decays
            if p.name in self.weight_decays.keys():
                p_t = _apply_weight_decays(self, p, p_t)
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)
            self.updates.append(K.update(p, new_p))

        # Cosine annealing
        _update_t_cur_eta_t(self)
        self.lr_t = lr_t * self.eta_t  # for external tracking

        self._init_notified = True
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        t = K.cast(self.iterations, K.floatx()) + 1

        # Due to the recommendations in [2], i.e. warming momentum schedule
        momentum_cache_t = self.beta_1 * (1. - 0.5 * (
            K.pow(K.cast_to_floatx(0.96), t * self.schedule_decay)))
        momentum_cache_t_1 = self.beta_1 * (1. - 0.5 * (
            K.pow(K.cast_to_floatx(0.96), (t + 1) * self.schedule_decay)))
        m_schedule_new = self.m_schedule * momentum_cache_t
        m_schedule_next = self.m_schedule * momentum_cache_t * momentum_cache_t_1
        self.updates.append((self.m_schedule, m_schedule_new))

        shapes = [K.int_shape(p) for p in params]
        ms = [K.zeros(shape) for shape in shapes]
        vs = [K.zeros(shape) for shape in shapes]

        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            # Learning rate multipliers
            lr_t = self.lr
            if self.lr_multipliers is not None:
                lr_t = _apply_lr_multiplier(self, lr_t, p)

            # the following equations given in [1]
            g_prime = g / (1. - m_schedule_new)
            m_t = self.beta_1 * m + (1. - self.beta_1) * g
            m_t_prime = m_t / (1. - m_schedule_next)
            v_t = self.beta_2 * v + (1. - self.beta_2) * K.square(g)
            v_t_prime = v_t / (1. - K.pow(self.beta_2, t))
            m_t_bar = (1. - momentum_cache_t) * g_prime + (
                momentum_cache_t_1 * m_t_prime)

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))

            p_t = p - self.eta_t * lr_t * m_t_bar / (
                K.sqrt(v_t_prime) + self.epsilon)

            # Weight decays
            if p.name in self.weight_decays.keys():
                p_t = _apply_weight_decays(self, p, p_t)
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)
            self.updates.append(K.update(p, new_p))

        # Cosine annealing
        _update_t_cur_eta_t(self)
        self.lr_t = lr_t * self.eta_t  # for external tracking

        self._init_notified = True
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))
        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + moments

        for p, g, m in zip(params, grads, moments):
            # Learning rate multipliers
            lr_t = self.lr
            if self.lr_multipliers is not None:
                lr_t = _apply_lr_multiplier(self, lr_t, p)

            v = self.momentum * m - self.eta_t * lr_t * g  # velocity
            self.updates.append(K.update(m, v))

            if self.nesterov:
                p_t = p + self.momentum * v - self.eta_t * lr_t * g
            else:
                p_t = p + v

            # Weight decays
            if p.name in self.weight_decays.keys():
                p_t = _apply_weight_decays(self, p, p_t)
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)
            self.updates.append(K.update(p, new_p))

        # Cosine annealing
        _update_t_cur_eta_t(self)
        self.lr_t = lr_t * self.eta_t  # for external tracking

        self._init_notified = True
        return self.updates 

def get_updates(self, params, loss):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.inital_decay > 0:
            lr *= (1. / (1. + self.decay * self.iterations))

        t = self.iterations + 1
        lr_t = lr * K.sqrt(1. - K.pow(self.beta_2, t)) / (1. - K.pow(self.beta_1, t))

        shapes = [K.get_variable_shape(p) for p in params]
        ms = [K.zeros(shape) for shape in shapes]
        vs = [K.zeros(shape) for shape in shapes]
        f = K.variable(0)
        d = K.variable(1)
        self.weights = [self.iterations] + ms + vs + [f, d]

        cond = K.greater(t, K.variable(1))
        small_delta_t = K.switch(K.greater(loss, f), self.small_k + 1, 1. / (self.big_K + 1))
        big_delta_t = K.switch(K.greater(loss, f), self.big_K + 1, 1. / (self.small_k + 1))

        c_t = K.minimum(K.maximum(small_delta_t, loss / (f + self.epsilon)), big_delta_t)
        f_t = c_t * f
        r_t = K.abs(f_t - f) / (K.minimum(f_t, f))
        d_t = self.beta_3 * d + (1 - self.beta_3) * r_t

        f_t = K.switch(cond, f_t, loss)
        d_t = K.switch(cond, d_t, K.variable(1.))

        self.updates.append(K.update(f, f_t))
        self.updates.append(K.update(d, d_t))

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            p_t = p - lr_t * m_t / (d_t * K.sqrt(v_t) + self.epsilon)

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))

            new_p = p_t
            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)

            m_t_hat = m_t / (1. - K.pow(self.beta_1, t))
            v_t_hat = v_t / (1. - K.pow(self.beta_2, t))

            p_dash = m_t_hat / (K.sqrt(v_t_hat + self.epsilon))

            if self.weight_decay > 0.:
                wd = self.weight_decay * p
                p_dash = p_dash + wd

            r1 = K.sqrt(K.sum(K.square(p)))
            r2 = K.sqrt(K.sum(K.square(p_dash)))

            r = tf.where(tf.greater(r1, 0.),
                         tf.where(tf.greater(r2, 0.),
                                  r1 / r2,
                                  1.0),
                         1.0)
            # r = r1 / r2
            eta = r * lr

            p_t = p - eta * p_dash

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr

        completed_updates = K.cast(K.tf.floor(self.iterations / self.accum_iters), K.floatx())

        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * completed_updates))

        t = completed_updates + 1

        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) / (1. - K.pow(self.beta_1, t)))

        # self.iterations incremented after processing a batch
        # batch:              1 2 3 4 5 6 7 8 9
        # self.iterations:    0 1 2 3 4 5 6 7 8
        # update_switch = 1:        x       x    (if accum_iters=4)  
        update_switch = K.equal((self.iterations + 1) % self.accum_iters, 0)
        update_switch = K.cast(update_switch, K.floatx())

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        gs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]

        if self.amsgrad:
            vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        else:
            vhats = [K.zeros(1) for _ in params]

        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat, tg in zip(params, grads, ms, vs, vhats, gs):

            sum_grad = tg + g
            avg_grad = sum_grad / self.accum_iters_float

            m_t = (self.beta_1 * m) + (1. - self.beta_1) * avg_grad
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(avg_grad)

            if self.amsgrad:
                vhat_t = K.maximum(vhat, v_t)
                p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
                self.updates.append(K.update(vhat, (1 - update_switch) * vhat + update_switch * vhat_t))
            else:
                p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

            self.updates.append(K.update(m, (1 - update_switch) * m + update_switch * m_t))
            self.updates.append(K.update(v, (1 - update_switch) * v + update_switch * v_t))
            self.updates.append(K.update(tg, (1 - update_switch) * sum_grad))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, (1 - update_switch) * p + update_switch * new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]
        lr = self.lr
        if self.initial_decay > 0:
            lr *= (1. / (1. + self.decay * K.cast(self.iterations, K.floatx())))
            pass
        t = K.cast(self.iterations + 1, K.floatx())
        lr_t = lr * K.sqrt(1. - K.pow(self.beta_2, t)) / (1. - K.pow(self.beta_1, t))
        shapes = [K.get_variable_shape(p) for p in params]
        ms = [K.zeros(shape) for shape in shapes]
        vs = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + ms + vs
        for p, g, m, v in zip(params, grads, ms, vs):
            # if a weight tensor (len > 1) use weight normalized parameterization
            # this is the only part changed w.r.t. keras.optimizers.Adam
            ps = K.get_variable_shape(p)
            if len(ps)>1:
                # get weight normalization parameters
                V, V_norm, V_scaler, g_param, grad_g, grad_V = get_weightnorm_params_and_grads(p, g)
                # Adam containers for the 'g' parameter
                V_scaler_shape = K.get_variable_shape(V_scaler)
                m_g = K.zeros(V_scaler_shape)
                v_g = K.zeros(V_scaler_shape)
                # update g parameters
                m_g_t = (self.beta_1 * m_g) + (1. - self.beta_1) * grad_g
                v_g_t = (self.beta_2 * v_g) + (1. - self.beta_2) * K.square(grad_g)
                new_g_param = g_param - lr_t * m_g_t / (K.sqrt(v_g_t) + self.epsilon)
                self.updates.append(K.update(m_g, m_g_t))
                self.updates.append(K.update(v_g, v_g_t))
                # update V parameters
                m_t = (self.beta_1 * m) + (1. - self.beta_1) * grad_V
                v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(grad_V)
                new_V_param = V - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
                self.updates.append(K.update(m, m_t))
                self.updates.append(K.update(v, v_t))
                # if there are constraints we apply them to V, not W
                if getattr(p, 'constraint', None) is not None:
                    new_V_param = p.constraint(new_V_param)
                    pass
                # wn param updates --> W updates
                add_weightnorm_param_updates(self.updates, new_V_param, new_g_param, p, V_scaler)
                pass
            else: # do optimization normally
                m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
                v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
                p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
                self.updates.append(K.update(m, m_t))
                self.updates.append(K.update(v, v_t))
                new_p = p_t
                # apply constraints
                if getattr(p, 'constraint', None) is not None:
                    new_p = p.constraint(new_p)
                    pass
                self.updates.append(K.update(p, new_p))
                pass
            pass
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1

        # Applies bounds on actual learning rate
        step_size = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                          (1. - K.pow(self.beta_1, t)))

        final_lr = self.final_lr * lr / self.base_lr
        lower_bound = final_lr * (1. - 1. / (self.gamma * t + 1.))
        upper_bound = final_lr * (1. + 1. / (self.gamma * t))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        if self.amsbound:
            vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        else:
            vhats = [K.zeros(1) for _ in params]
        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
            # apply weight decay
            if self.weight_decay != 0.:
                g += self.weight_decay * K.stop_gradient(p)

            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)

            if self.amsbound:
                vhat_t = K.maximum(vhat, v_t)
                denom = (K.sqrt(vhat_t) + self.epsilon)
                self.updates.append(K.update(vhat, vhat_t))
            else:
                denom = (K.sqrt(v_t) + self.epsilon)

            # Compute the bounds
            step_size_p = step_size * K.ones_like(denom)
            step_size_p_bound = step_size_p / denom
            bounded_lr_t = m_t * K.minimum(K.maximum(step_size_p_bound,
                                                     lower_bound), upper_bound)

            p_t = p - bounded_lr_t

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = tf.train.polynomial_decay(
            self.lr,
            self.iterations,
            self.num_train_steps,
            end_learning_rate=0.0,
            power=1.0,
            cycle=False
        )

        # Implements linear warmup. I.e., if global_step < num_warmup_steps, the
        # learning rate will be `global_step/num_warmup_steps * init_lr`.
        t = K.cast(self.iterations, K.floatx()) + 1
        warmup_percent_done = K.cast(t / self.num_warmup_steps, dtype=K.floatx())
        warmup_lr = self.lr * warmup_percent_done
        is_warmup = K.cast(t < self.num_warmup_steps, dtype=K.floatx())
        lr = ((1.0 - is_warmup) * lr) + is_warmup * warmup_lr

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]

        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            if self.bias_corrected:
                m_t /= 1 - K.pow(self.beta_1, t)
                v_t /= 1 - K.pow(self.beta_2, t)

            update = m_t / (K.sqrt(v_t) + self.epsilon)

            # Just adding the square of the weights to the loss function is *not*
            # the correct way of using L2 regularization/weight decay with Adam,
            # since that will interact with the m and v parameters in strange ways.
            #
            # Instead we want ot decay the weights in a manner that doesn't interact
            # with the m/v parameters. This is equivalent to adding the square
            # of the weights to the loss with plain (non-momentum) SGD.
            param_name = self._get_variable_name(p.name)
            if self._do_use_weight_decay(param_name):
                update += self.weight_decay_rate * p

            p_t = p - lr * update

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr *= (1. / (1. + self.decay * K.cast(self.iterations,
                                                  K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                     (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        if self.amsgrad:
            vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        else:
            vhats = [K.zeros(1) for _ in params]
        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):

            # Learning rate multipliers
            if self.multipliers:
                multiplier = [mult for mult in self.multipliers if mult in p.name]
            else:
                multiplier = None
            if multiplier:
                new_lr_t = lr_t * self.multipliers[multiplier[0]]
                if self.debug_verbose:
                    print('Setting {} to learning rate {}'.format(multiplier[0], new_lr_t))
                    print(K.get_value(new_lr_t))
            else:
                new_lr_t = lr_t
                if self.debug_verbose:
                    print('No change in learning rate {}'.format(p.name))
                    print(K.get_value(new_lr_t))
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            if self.amsgrad:
                vhat_t = K.maximum(vhat, v_t)
                p_t = p - new_lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
                self.updates.append(K.update(vhat, vhat_t))
            else:
                p_t = p - new_lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates