Python utils.vectorize方法代碼示例

# 需要導入模塊: import utils [as 別名]
# 或者: from utils import vectorize [as 別名]
def get_rebar_gradient(self):
    """Get the rebar gradient."""
    hardELBO, nvil_gradient, logQHard = self._create_hard_elbo()
    if self.hparams.quadratic:
      gumbel_cv, _ = self._create_gumbel_control_variate_quadratic(logQHard)
    else:
      gumbel_cv, _ = self._create_gumbel_control_variate(logQHard)

    f_grads = self.optimizer_class.compute_gradients(tf.reduce_mean(-nvil_gradient))

    eta = {}
    h_grads, eta_statistics = self.multiply_by_eta_per_layer(
        self.optimizer_class.compute_gradients(tf.reduce_mean(gumbel_cv)),
        eta)

    model_grads = U.add_grads_and_vars(f_grads, h_grads)
    total_grads = model_grads

    # Construct the variance objective
    variance_objective = tf.reduce_mean(tf.square(U.vectorize(model_grads, set_none_to_zero=True)))

    debug = { 'ELBO': hardELBO,
             'etas': eta_statistics,
             'variance_objective': variance_objective,
             }
    return total_grads, debug, variance_objective

###
# Create varaints
### 

# 需要導入模塊: import utils [as 別名]
# 或者: from utils import vectorize [as 別名]
def compute_gradient_moments(self, grads_and_vars):
    first_moment = U.vectorize(grads_and_vars, set_none_to_zero=True)
    second_moment = tf.square(first_moment)
    self.maintain_ema_ops.append(self.ema.apply([first_moment, second_moment]))

    return self.ema.average(first_moment), self.ema.average(second_moment) 

# 需要導入模塊: import utils [as 別名]
# 或者: from utils import vectorize [as 別名]
def _create_train_op(self, grads_and_vars, extra_grads_and_vars=[]):
    '''
    Args:
      grads_and_vars: gradients to apply and compute running average variance
      extra_grads_and_vars: gradients to apply (not used to compute average variance)
    '''
    # Variance summaries
    first_moment = U.vectorize(grads_and_vars, skip_none=True)
    second_moment = tf.square(first_moment)
    self.maintain_ema_ops.append(self.ema.apply([first_moment, second_moment]))

    # Add baseline losses
    if len(self.baseline_loss) > 0:
      mean_baseline_loss = tf.reduce_mean(tf.add_n(self.baseline_loss))
      extra_grads_and_vars += self.optimizer_class.compute_gradients(
          mean_baseline_loss,
          var_list=tf.get_collection('BASELINE'))

    # Ensure that all required tensors are computed before updates are executed
    extra_optimizer = tf.train.AdamOptimizer(
        learning_rate=10*self.hparams.learning_rate,
        beta2=self.hparams.beta2)
    with tf.control_dependencies(
        [tf.group(*[g for g, _ in (grads_and_vars + extra_grads_and_vars) if g is not None])]):

      # Filter out the P_COLLECTION variables if we're in eval mode
      if self.eval_mode:
        grads_and_vars = [(g, v) for g, v in grads_and_vars
                          if v not in tf.get_collection(P_COLLECTION)]

      train_op = self.optimizer_class.apply_gradients(grads_and_vars,
                                                      global_step=self.global_step)

      if len(extra_grads_and_vars) > 0:
        extra_train_op = extra_optimizer.apply_gradients(extra_grads_and_vars)
      else:
        extra_train_op = tf.no_op()

      self.optimizer = tf.group(train_op, extra_train_op, *self.maintain_ema_ops)

    # per parameter variance
    variance_estimator = (self.ema.average(second_moment) -
        tf.square(self.ema.average(first_moment)))
    self.grad_variance = tf.reduce_mean(variance_estimator) 

# 需要導入模塊: import utils [as 別名]
# 或者: from utils import vectorize [as 別名]
def get_dynamic_rebar_gradient(self):
    """Get the dynamic rebar gradient (t, eta optimized)."""
    tiled_pre_temperature = tf.tile([self.pre_temperature_variable],
                                [self.batch_size])
    temperature = tf.exp(tiled_pre_temperature)

    hardELBO, nvil_gradient, logQHard = self._create_hard_elbo()
    if self.hparams.quadratic:
      gumbel_cv, extra  = self._create_gumbel_control_variate_quadratic(logQHard, temperature=temperature)
    else:
      gumbel_cv, extra  = self._create_gumbel_control_variate(logQHard, temperature=temperature)

    f_grads = self.optimizer_class.compute_gradients(tf.reduce_mean(-nvil_gradient))

    eta = {}
    h_grads, eta_statistics = self.multiply_by_eta_per_layer(
        self.optimizer_class.compute_gradients(tf.reduce_mean(gumbel_cv)),
        eta)

    model_grads = U.add_grads_and_vars(f_grads, h_grads)
    total_grads = model_grads

    # Construct the variance objective
    g = U.vectorize(model_grads, set_none_to_zero=True)
    self.maintain_ema_ops.append(self.ema.apply([g]))
    gbar = 0  #tf.stop_gradient(self.ema.average(g))
    variance_objective = tf.reduce_mean(tf.square(g - gbar))

    reinf_g_t = 0
    if self.hparams.quadratic:
      for layer in xrange(self.hparams.n_layer):
        gumbel_learning_signal, _ = extra[layer]
        df_dt = tf.gradients(gumbel_learning_signal, tiled_pre_temperature)[0]
        reinf_g_t_i, _ = self.multiply_by_eta_per_layer(
            self.optimizer_class.compute_gradients(tf.reduce_mean(tf.stop_gradient(df_dt) * logQHard[layer])),
            eta)
        reinf_g_t += U.vectorize(reinf_g_t_i, set_none_to_zero=True)

      reparam = tf.add_n([reparam_i for _, reparam_i in extra])
    else:
      gumbel_learning_signal, reparam = extra
      df_dt = tf.gradients(gumbel_learning_signal, tiled_pre_temperature)[0]
      reinf_g_t, _ = self.multiply_by_eta_per_layer(
          self.optimizer_class.compute_gradients(tf.reduce_mean(tf.stop_gradient(df_dt) * tf.add_n(logQHard))),
          eta)
      reinf_g_t = U.vectorize(reinf_g_t, set_none_to_zero=True)

    reparam_g, _ = self.multiply_by_eta_per_layer(
        self.optimizer_class.compute_gradients(tf.reduce_mean(reparam)),
        eta)
    reparam_g = U.vectorize(reparam_g, set_none_to_zero=True)
    reparam_g_t = tf.gradients(tf.reduce_mean(2*tf.stop_gradient(g - gbar)*reparam_g), self.pre_temperature_variable)[0]

    variance_objective_grad = tf.reduce_mean(2*(g - gbar)*reinf_g_t) + reparam_g_t

    debug = { 'ELBO': hardELBO,
             'etas': eta_statistics,
             'variance_objective': variance_objective,
             }
    return total_grads, debug, variance_objective, variance_objective_grad