Python utils.add_grads_and_vars() Examples

The following are 16 code examples of utils.add_grads_and_vars(). You can vote up the ones you like or vote down the ones you don't like, and go to the original project or source file by following the links above each example. You may also want to check out all available functions/classes of the module utils , or try the search function .
Example #1
Source File: rebar.py    From yolo_v2 with Apache License 2.0 5 votes vote down vote up 
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
###
Example #2
Source File: rebar.py    From Gun-Detector with Apache License 2.0 5 votes vote down vote up 
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
###
Example #3
Source File: rebar.py    From hands-detection with MIT License 5 votes vote down vote up 
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
###
Example #4
Source File: rebar.py    From object_detection_kitti with Apache License 2.0 5 votes vote down vote up 
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
###
Example #5
Source File: rebar.py    From object_detection_with_tensorflow with MIT License 5 votes vote down vote up 
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
###
Example #6
Source File: rebar.py    From g-tensorflow-models with Apache License 2.0 5 votes vote down vote up 
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
###
Example #7
Source File: rebar.py    From models with Apache License 2.0 5 votes vote down vote up 
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
###
Example #8
Source File: rebar.py    From multilabel-image-classification-tensorflow with MIT License 5 votes vote down vote up 
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
###
Example #9
Source File: rebar.py    From yolo_v2 with Apache License 2.0 4 votes vote down vote up 
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
Example #10
Source File: rebar.py    From Gun-Detector with Apache License 2.0 4 votes vote down vote up 
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
Example #11
Source File: rebar.py    From hands-detection with MIT License 4 votes vote down vote up 
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
Example #12
Source File: rebar.py    From object_detection_kitti with Apache License 2.0 4 votes vote down vote up 
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
Example #13
Source File: rebar.py    From object_detection_with_tensorflow with MIT License 4 votes vote down vote up 
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
Example #14
Source File: rebar.py    From g-tensorflow-models with Apache License 2.0 4 votes vote down vote up 
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
Example #15
Source File: rebar.py    From models with Apache License 2.0 4 votes vote down vote up 
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
Example #16
Source File: rebar.py    From multilabel-image-classification-tensorflow with MIT License 4 votes vote down vote up 
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