예제 #1
0
class BNN:
    """Neural network models which model aleatoric uncertainty (and possibly epistemic uncertainty
    with ensembling).
    """
    def __init__(self, params):
        """Initializes a class instance.

        Arguments:
            params (DotMap): A dotmap of model parameters.
                .name (str): Model name, used for logging/use in variable scopes.
                    Warning: Models with the same name will overwrite each other.
                .num_networks (int): (optional) The number of networks in the ensemble. Defaults to 1.
                    Ignored if model is being loaded.
                .model_dir (str/None): (optional) Path to directory from which model will be loaded, and
                    saved by default. Defaults to None.
                .load_model (bool): (optional) If True, model will be loaded from the model directory,
                    assuming that the files are generated by a model of the same name. Defaults to False.
                .sess (tf.Session/None): The session that this model will use.
                    If None, creates a session with its own associated graph. Defaults to None.
        """
        self.name = get_required_argument(params, 'name', 'Must provide name.')
        self.model_dir = params.get('model_dir', None)

        if params.get('sess', None) is None:
            config = tf.ConfigProto()
            config.gpu_options.allow_growth = True
            config.gpu_options.per_process_gpu_memory_fraction = 0.4
            self._sess = tf.Session(config=config)
        else:
            self._sess = params.get('sess')

        # Instance variables
        self.finalized = False
        self.layers, self.max_logvar, self.min_logvar = [], None, None
        self.decays, self.optvars, self.nonoptvars = [], [], []
        self.end_act, self.end_act_name = None, None
        self.scaler = None

        # Training objects
        self.optimizer = None
        self.sy_train_in, self.sy_train_targ = None, None
        self.train_op, self.mse_loss = None, None

        # Prediction objects
        self.sy_pred_in2d, self.sy_pred_mean2d_fac, self.sy_pred_var2d_fac = None, None, None
        self.sy_pred_mean2d, self.sy_pred_var2d = None, None
        self.sy_pred_in3d, self.sy_pred_mean3d_fac, self.sy_pred_var3d_fac = None, None, None

        if params.get('load_model', False):
            if self.model_dir is None:
                raise ValueError(
                    "Cannot load model without providing model directory.")
            self._load_structure()
            self.num_nets, self.model_loaded = self.layers[
                0].get_ensemble_size(), True
            print("Model loaded from %s." % self.model_dir)
        else:
            self.num_nets = params.get('num_networks', 1)
            self.model_loaded = False

        if self.num_nets == 1:
            print("Created a neural network with variance predictions.")
        else:
            print(
                "Created an ensemble of %d neural networks with variance predictions."
                % (self.num_nets))

    @property
    def is_probabilistic(self):
        return True

    @property
    def is_tf_model(self):
        return True

    @property
    def sess(self):
        return self._sess

    ###################################
    # Network Structure Setup Methods #
    ###################################

    def add(self, layer):
        """Adds a new layer to the network.

        Arguments:
            layer: (layer) The new layer to be added to the network.
                   If this is the first layer, the input dimension of the layer must be set.

        Returns: None.
        """
        if self.finalized:
            raise RuntimeError(
                "Cannot modify network structure after finalizing.")
        if len(self.layers) == 0 and layer.get_input_dim() is None:
            raise ValueError("Must set input dimension for the first layer.")
        if self.model_loaded:
            raise RuntimeError("Cannot add layers to a loaded model.")

        layer.set_ensemble_size(self.num_nets)
        if len(self.layers) > 0:
            layer.set_input_dim(self.layers[-1].get_output_dim())
        self.layers.append(layer.copy())

    def pop(self):
        """Removes and returns the most recently added layer to the network.

        Returns: (layer) The removed layer.
        """
        if len(self.layers) == 0:
            raise RuntimeError("Network is empty.")
        if self.finalized:
            raise RuntimeError(
                "Cannot modify network structure after finalizing.")
        if self.model_loaded:
            raise RuntimeError("Cannot remove layers from a loaded model.")

        return self.layers.pop()

    def finalize(self, optimizer, optimizer_args=None, *args, **kwargs):
        """Finalizes the network.

        Arguments:
            optimizer: (tf.train.Optimizer) An optimizer class from those available at tf.train.Optimizer.
            optimizer_args: (dict) A dictionary of arguments for the __init__ method of the chosen optimizer.

        Returns: None
        """
        if len(self.layers) == 0:
            raise RuntimeError("Cannot finalize an empty network.")
        if self.finalized:
            raise RuntimeError("Can only finalize a network once.")

        optimizer_args = {} if optimizer_args is None else optimizer_args
        self.optimizer = optimizer(**optimizer_args)

        # Add variance output.
        self.layers[-1].set_output_dim(2 * self.layers[-1].get_output_dim())

        # Remove last activation to isolate variance from activation function.
        self.end_act = self.layers[-1].get_activation()
        self.end_act_name = self.layers[-1].get_activation(as_func=False)
        self.layers[-1].unset_activation()

        # Construct all variables.
        with self.sess.as_default():
            with tf.variable_scope(self.name):
                self.scaler = TensorStandardScaler(
                    self.layers[0].get_input_dim())
                self.max_logvar = tf.Variable(
                    np.ones([1, self.layers[-1].get_output_dim() // 2]) / 2.,
                    dtype=tf.float32,
                    name="max_log_var")
                self.min_logvar = tf.Variable(
                    -np.ones([1, self.layers[-1].get_output_dim() // 2]) * 10.,
                    dtype=tf.float32,
                    name="min_log_var")
                for i, layer in enumerate(self.layers):
                    with tf.variable_scope("Layer%i" % i):
                        layer.construct_vars()
                        self.decays.extend(layer.get_decays())
                        self.optvars.extend(layer.get_vars())
        self.optvars.extend([self.max_logvar, self.min_logvar])
        self.nonoptvars.extend(self.scaler.get_vars())

        # Set up training
        with tf.variable_scope(self.name):
            self.optimizer = optimizer(**optimizer_args)
            self.sy_train_in = tf.placeholder(
                dtype=tf.float32,
                shape=[self.num_nets, None, self.layers[0].get_input_dim()],
                name="training_inputs")
            self.sy_train_targ = tf.placeholder(
                dtype=tf.float32,
                shape=[
                    self.num_nets, None, self.layers[-1].get_output_dim() // 2
                ],
                name="training_targets")
            train_loss = tf.reduce_sum(
                self._compile_losses(self.sy_train_in,
                                     self.sy_train_targ,
                                     inc_var_loss=True))
            train_loss += tf.add_n(self.decays)
            # regularization to ensure max_logvar not grow too much beyond training distribution
            # and min_logvar not drop below training distribution
            train_loss += 0.01 * tf.reduce_sum(
                self.max_logvar) - 0.01 * tf.reduce_sum(self.min_logvar)
            self.mse_loss = self._compile_losses(self.sy_train_in,
                                                 self.sy_train_targ,
                                                 inc_var_loss=False)

            self.train_op = self.optimizer.minimize(train_loss,
                                                    var_list=self.optvars)

        # Initialize all variables
        self.sess.run(
            tf.variables_initializer(self.optvars + self.nonoptvars +
                                     self.optimizer.variables()))

        # Set up prediction
        with tf.variable_scope(self.name):
            self.sy_pred_in2d = tf.placeholder(
                dtype=tf.float32,
                shape=[None, self.layers[0].get_input_dim()],
                name="2D_training_inputs")
            self.sy_pred_mean2d_fac, self.sy_pred_var2d_fac = \
                self.create_prediction_tensors(self.sy_pred_in2d, factored=True)
            self.sy_pred_mean2d = tf.reduce_mean(self.sy_pred_mean2d_fac,
                                                 axis=0)
            self.sy_pred_var2d = tf.reduce_mean(self.sy_pred_var2d_fac, axis=0) + \
                tf.reduce_mean(tf.square(self.sy_pred_mean2d_fac - self.sy_pred_mean2d), axis=0)

            self.sy_pred_in3d = tf.placeholder(
                dtype=tf.float32,
                shape=[self.num_nets, None, self.layers[0].get_input_dim()],
                name="3D_training_inputs")
            self.sy_pred_mean3d_fac, self.sy_pred_var3d_fac = \
                self.create_prediction_tensors(self.sy_pred_in3d, factored=True)

        # Load model if needed
        # self.optimizer.variables() no need to save and load
        if self.model_loaded:
            with self.sess.as_default():
                params_dict = loadmat(
                    os.path.join(self.model_dir, "%s.mat" % self.name))
                all_vars = self.nonoptvars + self.optvars
                for i, var in enumerate(all_vars):
                    var.load(params_dict[str(i)])
        self.finalized = True

    #################
    # Model Methods #
    #################

    def train(self,
              inputs,
              targets,
              batch_size=32,
              epochs=100,
              hide_progress=False,
              holdout_ratio=0.0,
              max_logging=5000):
        """Trains/Continues network training

        Arguments:
            inputs (np.ndarray): Network inputs in the training dataset in rows.
            targets (np.ndarray): Network target outputs in the training dataset in rows corresponding
                to the rows in inputs.
            batch_size (int): The minibatch size to be used for training.
            epochs (int): Number of epochs (full network passes that will be done.
            hide_progress (bool): If True, hides the progress bar shown at the beginning of training.

        Returns: None
        """
        def shuffle_rows(arr):
            idxs = np.argsort(np.random.uniform(size=arr.shape), axis=-1)
            return arr[np.arange(arr.shape[0])[:, None], idxs]

        # Split into training and holdout sets
        num_holdout = min(int(inputs.shape[0] * holdout_ratio), max_logging)
        # shuffle np.range(inputs.shape[0])
        permutation = np.random.permutation(inputs.shape[0])
        inputs, holdout_inputs = inputs[permutation[num_holdout:]], inputs[
            permutation[:num_holdout]]
        targets, holdout_targets = targets[permutation[num_holdout:]], targets[
            permutation[:num_holdout]]
        holdout_inputs = np.tile(holdout_inputs[None], [self.num_nets, 1, 1])
        holdout_targets = np.tile(holdout_targets[None], [self.num_nets, 1, 1])

        with self.sess.as_default():
            self.scaler.fit(inputs)
        #range bewteen 0-inuputs.shape[0]
        # [ensemble size, train_sample_size]
        idxs = np.random.randint(inputs.shape[0],
                                 size=[self.num_nets, inputs.shape[0]])
        if hide_progress:
            epoch_range = range(epochs)
        else:
            epoch_range = trange(epochs,
                                 unit="epoch(s)",
                                 desc="Network training")
        for _ in epoch_range:
            for batch_num in range(int(np.ceil(idxs.shape[-1] / batch_size))):
                batch_idxs = idxs[:, batch_num * batch_size:(batch_num + 1) *
                                  batch_size]
                self.sess.run(self.train_op,
                              feed_dict={
                                  self.sy_train_in: inputs[batch_idxs],
                                  self.sy_train_targ: targets[batch_idxs]
                              })
            idxs = shuffle_rows(idxs)
            if not hide_progress:
                if holdout_ratio < 1e-12:
                    epoch_range.set_postfix({
                        "Training loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in:
                                          inputs[idxs[:, :max_logging]],
                                          self.sy_train_targ:
                                          targets[idxs[:, :max_logging]]
                                      })
                    })
                else:
                    epoch_range.set_postfix({
                        "Training loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in:
                                          inputs[idxs[:, :max_logging]],
                                          self.sy_train_targ:
                                          targets[idxs[:, :max_logging]]
                                      }),
                        "Holdout loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in: holdout_inputs,
                                          self.sy_train_targ: holdout_targets
                                      })
                    })

    def predict(self, inputs, factored=False, *args, **kwargs):
        """Returns the distribution predicted by the model for each input vector in inputs.
        Behavior is affected by the dimensionality of inputs and factored as follows:

        inputs is 2D, factored=True: Each row is treated as an input vector.
            Returns a mean of shape [ensemble_size, batch_size, output_dim] and variance of shape
            [ensemble_size, batch_size, output_dim], where N(mean[i, j, :], diag([i, j, :])) is the
            predicted output distribution by the ith model in the ensemble on input vector j.

        inputs is 2D, factored=False: Each row is treated as an input vector.
            Returns a mean of shape [batch_size, output_dim] and variance of shape
            [batch_size, output_dim], where aggregation is performed as described in the paper.

        inputs is 3D, factored=True/False: Each row in the last dimension is treated as an input vector.
            Returns a mean of shape [ensemble_size, batch_size, output_dim] and variance of sha
            [ensemble_size, batch_size, output_dim], where N(mean[i, j, :], diag([i, j, :])) is the
            predicted output distribution by the ith model in the ensemble on input vector [i, j].

        Arguments:
            inputs (np.ndarray): An array of input vectors in rows. See above for behavior.
            factored (bool): See above for behavior.
        """
        if len(inputs.shape) == 2:
            if factored:
                return self.sess.run(
                    [self.sy_pred_mean2d_fac, self.sy_pred_var2d_fac],
                    feed_dict={self.sy_pred_in2d: inputs})
            else:
                return self.sess.run([self.sy_pred_mean2d, self.sy_pred_var2d],
                                     feed_dict={self.sy_pred_in2d: inputs})
        else:
            return self.sess.run(
                [self.sy_pred_mean3d_fac, self.sy_pred_var3d_fac],
                feed_dict={self.sy_pred_in3d: inputs})

    def create_prediction_tensors(self,
                                  inputs,
                                  factored=False,
                                  *args,
                                  **kwargs):
        """See predict() above for documentation.
        """
        factored_mean, factored_variance = self._compile_outputs(inputs)
        if inputs.shape.ndims == 2 and not factored:
            # DS
            # [ensemble_size, nparticle, dO] -> [nparticle, dO]
            mean = tf.reduce_mean(factored_mean, axis=0)
            variance = tf.reduce_mean(tf.square(factored_mean - mean), axis=0) + \
                       tf.reduce_mean(factored_variance, axis=0)
            return mean, variance
        # TS1, TSInf
        # [ensemble_size, nparticle / ensemble_size, dO]
        return factored_mean, factored_variance

    def save(self, savedir=None):
        """Saves all information required to recreate this model in two files in savedir
        (or self.model_dir if savedir is None), one containing the model structuure and the other
        containing all variables in the network.

        savedir (str): (Optional) Path to which files will be saved. If not provided, self.model_dir
            (the directory provided at initialization) will be used.
        """
        if not self.finalized:
            raise RuntimeError()
        model_dir = self.model_dir if savedir is None else savedir

        # Write structure to file
        with open(os.path.join(model_dir, "%s.nns" % self.name), "w+") as f:
            for layer in self.layers[:-1]:
                f.write("%s\n" % repr(layer))
            last_layer_copy = self.layers[-1].copy()
            last_layer_copy.set_activation(self.end_act_name)
            last_layer_copy.set_output_dim(last_layer_copy.get_output_dim() //
                                           2)
            f.write("%s\n" % repr(last_layer_copy))

        # Save network parameters (including scalers) in a .mat file
        var_vals = {}
        for i, var_val in enumerate(
                self.sess.run(self.nonoptvars + self.optvars)):
            var_vals[str(i)] = var_val
        savemat(os.path.join(model_dir, "%s.mat" % self.name), var_vals)

    def _load_structure(self):
        """Uses the saved structure in self.model_dir with the name of this network to initialize
        the structure of this network.
        """
        structure = []
        with open(os.path.join(self.model_dir, "%s.nns" % self.name),
                  "r") as f:
            for line in f:
                kwargs = {
                    key: val
                    for (key, val) in
                    [argval.split("=") for argval in line[3:-2].split(", ")]
                }
                kwargs["input_dim"] = int(kwargs["input_dim"])
                kwargs["output_dim"] = int(kwargs["output_dim"])
                kwargs["weight_decay"] = None if kwargs[
                    "weight_decay"] == "None" else float(
                        kwargs["weight_decay"])
                kwargs["activation"] = None if kwargs[
                    "activation"] == "None" else kwargs["activation"][1:-1]
                kwargs["ensemble_size"] = int(kwargs["ensemble_size"])
                structure.append(FC(**kwargs))
        self.layers = structure

    #######################
    # Compilation methods #
    #######################

    def _compile_outputs(self, inputs, ret_log_var=False):
        """Compiles the output of the network at the given inputs.

        If inputs is 2D, returns a 3D tensor where output[i] is the output of the ith network in the ensemble.
        If inputs is 3D, returns a 3D tensor where output[i] is the output of the ith network on the ith input matrix.

        Arguments:
            inputs: (tf.Tensor) A tensor representing the inputs to the network
            ret_log_var: (bool) If True, returns the log variance instead of the variance.

        Returns: (tf.Tensors) The mean and variance/log variance predictions at inputs for each network
            in the ensemble.
        """
        dim_output = self.layers[-1].get_output_dim()
        cur_out = self.scaler.transform(inputs)
        for layer in self.layers:
            cur_out = layer.compute_output_tensor(cur_out)

        mean = cur_out[:, :, :dim_output // 2]
        if self.end_act is not None:
            mean = self.end_act(mean)

        # notice the gradient flows here through max_logvar and min_logvar
        # equal as exp(max_logvar) * exp(logvar) / exp(max_logvar) + exp(logvar)
        logvar = self.max_logvar - tf.nn.softplus(
            self.max_logvar - cur_out[:, :, dim_output // 2:])
        # equal as exp(logvar) + exp(min_var)
        logvar = self.min_logvar + tf.nn.softplus(logvar - self.min_logvar)

        if ret_log_var:
            return mean, logvar
        else:
            return mean, tf.exp(logvar)

    def _compile_losses(self, inputs, targets, inc_var_loss=True):
        """Helper method for compiling the loss function.

        The loss function is obtained from the log likelihood, assuming that the output
        distribution is Gaussian, with both mean and (diagonal) covariance matrix being determined
        by network outputs.

        Arguments:
            inputs: (tf.Tensor) A tensor representing the input batch
            targets: (tf.Tensor) The desired targets for each input vector in inputs.
            inc_var_loss: (bool) If True, includes log variance loss.

        Returns: (tf.Tensor) A tensor representing the loss on the input arguments.
        """
        mean, log_var = self._compile_outputs(inputs, ret_log_var=True)
        inv_var = tf.exp(-log_var)

        if inc_var_loss:
            # max log liklihood in for Gaussian
            mse_losses = tf.reduce_mean(tf.reduce_mean(
                tf.square(mean - targets) * inv_var, axis=-1),
                                        axis=-1)
            var_losses = tf.reduce_mean(tf.reduce_mean(log_var, axis=-1),
                                        axis=-1)
            total_losses = mse_losses + var_losses
        else:
            total_losses = tf.reduce_mean(tf.reduce_mean(tf.square(mean -
                                                                   targets),
                                                         axis=-1),
                                          axis=-1)

        return total_losses
예제 #2
0
class NN:
    """Neural network models which cannot capture aleatoric uncertainty (but possibly epistemic uncertainty
    with ensembling).
    """
    def __init__(self, params):
        """Initializes a class instance.

        Arguments:
            params (DotMap): A dotmap of model parameters.
                .name (str): Model name, used for logging/use in variable scopes.
                    Warning: Models with the same name will overwrite each other.
                .num_networks (int): (optional) The number of networks in the ensemble. Defaults to 1.
                    Ignored if model is being loaded.
                .model_dir (str/None): (optional) Path to directory from which model will be loaded, and
                    saved by default. Defaults to None.
                .load_model (bool): (optional) If True, model will be loaded from the model directory,
                    assuming that the files are generated by a model of the same name. Defaults to False.
                .sess (tf.Session/None): The session that this model will use.
                    If None, creates a session with its own associated graph. Defaults to None.
        """
        self.name = get_required_argument(params, 'name', 'Must provide name.')
        self.model_dir = params.get('model_dir', None)

        if params.get('sess', None) is None:
            config = tf.ConfigProto()
            # config.gpu_options.allow_growth = True
            self._sess = tf.Session(config=config)
        else:
            self._sess = params.get('sess')

        # Instance variables
        self.finalized = False
        self.layers, self.decays, self.optvars, self.nonoptvars = [], [], [], []
        self.scaler = None

        # Training objects
        self.optimizer = None
        self.sy_train_in, self.sy_train_targ = None, None
        self.train_op, self.mse_loss = None, None

        # Prediction objects
        self.sy_pred_in2d, self.sy_pred_mean2d_fac = None, None
        self.sy_pred_mean2d, self.sy_pred_var2d = None, None
        self.sy_pred_in3d, self.sy_pred_mean3d_fac = None, None

        if params.get('load_model', False):
            if self.model_dir is None:
                raise ValueError(
                    "Cannot load model without providing model directory.")
            self._load_structure()
            self.num_nets, self.model_loaded = self.layers[
                0].get_ensemble_size(), True
            print("Model loaded from %s." % self.model_dir)
        else:
            self.num_nets = params.get('num_networks', 1)
            self.model_loaded = False

        if self.num_nets == 1:
            print("Created a neural network without variance predictions.")
        else:
            print(
                "Created an ensemble of %d neural networks without variance predictions."
                % (self.num_nets))

    @property
    def is_probabilistic(self):
        return True if self.num_nets > 1 else False

    @property
    def is_tf_model(self):
        return True

    @property
    def sess(self):
        return self._sess

    ###################################
    # Network Structure Setup Methods #
    ###################################

    def add(self, layer):
        """Adds a new layer to the network.

        Arguments:
            layer: (layer) The new layer to be added to the network.
                   If this is the first layer, the input dimension of the layer must be set.

        Returns: None.
        """
        if self.finalized:
            raise RuntimeError(
                "Cannot modify network structure after finalizing.")
        if len(self.layers) == 0 and layer.get_input_dim() is None:
            raise ValueError("Must set input dimension for the first layer.")
        if self.model_loaded:
            raise RuntimeError("Cannot add layers to a loaded model.")

        layer.set_ensemble_size(self.num_nets)
        if len(self.layers) > 0:
            layer.set_input_dim(self.layers[-1].get_output_dim())
        self.layers.append(layer.copy())

    def pop(self):
        """Removes and returns the most recently added layer to the network.

        Returns: (layer) The removed layer.
        """
        if len(self.layers) == 0:
            raise RuntimeError("Network is empty.")
        if self.finalized:
            raise RuntimeError(
                "Cannot modify network structure after finalizing.")
        if self.model_loaded:
            raise RuntimeError("Cannot remove layers from a loaded model.")

        return self.layers.pop()

    def finalize(self,
                 optimizer,
                 optimizer_args=None,
                 suffix="",
                 *args,
                 **kwargs):
        """Finalizes the network.

        Arguments:
            optimizer: (tf.train.Optimizer) An optimizer class from those available at tf.train.Optimizer.
            optimizer_args: (dict) A dictionary of arguments for the __init__ method of the chosen optimizer.

        Returns: None
        """
        if len(self.layers) == 0:
            raise RuntimeError("Cannot finalize an empty network.")
        if self.finalized:
            raise RuntimeError("Can only finalize a network once.")

        optimizer_args = {} if optimizer_args is None else optimizer_args
        self.optimizer = optimizer(**optimizer_args)

        # Construct all variables.
        with self.sess.as_default():
            with tf.variable_scope(self.name):
                self.scaler = TensorStandardScaler(
                    self.layers[0].get_input_dim(), suffix)
                for i, layer in enumerate(self.layers):
                    with tf.variable_scope(("Layer%i" + suffix) % i):
                        layer.construct_vars()
                        self.decays.extend(layer.get_decays())
                        self.optvars.extend(layer.get_vars())
        self.nonoptvars.extend(self.scaler.get_vars())

        # Setup training
        with tf.variable_scope(self.name):
            self.optimizer = optimizer(**optimizer_args)
            self.sy_train_in = tf.placeholder(
                dtype=tf.float32,
                shape=[self.num_nets, None, self.layers[0].get_input_dim()],
                name="training_inputs")
            self.sy_train_targ = tf.placeholder(
                dtype=tf.float32,
                shape=[self.num_nets, None, self.layers[-1].get_output_dim()],
                name="training_targets")
            train_loss = tf.reduce_sum(
                self._compile_losses(self.sy_train_in, self.sy_train_targ))
            train_loss += tf.add_n(self.decays)
            self.mse_loss = self._compile_losses(self.sy_train_in,
                                                 self.sy_train_targ)

            self.train_op = self.optimizer.minimize(train_loss,
                                                    var_list=self.optvars)

        # Initialize all variables
        self.sess.run(
            tf.variables_initializer(self.optvars + self.nonoptvars +
                                     self.optimizer.variables()))

        # Setup prediction
        with tf.variable_scope(self.name):
            self.sy_pred_in2d = tf.placeholder(
                dtype=tf.float32,
                shape=[None, self.layers[0].get_input_dim()],
                name="2D_training_inputs")
            self.sy_pred_mean2d_fac = self.create_prediction_tensors(
                self.sy_pred_in2d, factored=True)[0]
            self.sy_pred_mean2d = tf.reduce_mean(self.sy_pred_mean2d_fac,
                                                 axis=0)
            self.sy_pred_var2d = tf.reduce_mean(
                tf.square(self.sy_pred_mean2d_fac - self.sy_pred_mean2d),
                axis=0)

            self.sy_pred_in3d = tf.placeholder(
                dtype=tf.float32,
                shape=[self.num_nets, None, self.layers[0].get_input_dim()],
                name="3D_training_inputs")
            self.sy_pred_mean3d_fac = \
                self.create_prediction_tensors(self.sy_pred_in3d, factored=True)[0]

        # Load model if needed
        if self.model_loaded:
            with self.sess.as_default():
                params_dict = loadmat(
                    os.path.join(self.model_dir, "%s.mat" % self.name))
                all_vars = self.nonoptvars + self.optvars
                for i, var in enumerate(all_vars):
                    var.load(params_dict[str(i)])

        self.finalized = True

    #################
    # Model Methods #
    #################

    def train(self,
              inputs,
              targets,
              batch_size=32,
              epochs=100,
              hide_progress=False,
              holdout_ratio=0.0,
              max_logging=5000):
        def shuffle_rows(arr):
            idxs = np.argsort(np.random.uniform(size=arr.shape), axis=-1)
            return arr[np.arange(arr.shape[0])[:, None], idxs]

        # Split into training and holdout sets
        num_holdout = min(int(inputs.shape[0] * holdout_ratio), max_logging)
        permutation = np.random.permutation(inputs.shape[0])
        inputs, holdout_inputs = inputs[permutation[num_holdout:]], inputs[
            permutation[:num_holdout]]
        targets, holdout_targets = targets[permutation[num_holdout:]], targets[
            permutation[:num_holdout]]
        holdout_inputs = np.tile(holdout_inputs[None], [self.num_nets, 1, 1])
        holdout_targets = np.tile(holdout_targets[None], [self.num_nets, 1, 1])

        with self.sess.as_default():
            self.scaler.fit(inputs)

        idxs = np.random.randint(inputs.shape[0],
                                 size=[self.num_nets, inputs.shape[0]])
        if hide_progress:
            epoch_range = range(epochs)
        else:
            epoch_range = trange(epochs,
                                 unit="epoch(s)",
                                 desc="Network training")
        for _ in epoch_range:
            for batch_num in range(int(np.ceil(idxs.shape[-1] / batch_size))):
                batch_idxs = idxs[:, batch_num * batch_size:(batch_num + 1) *
                                  batch_size]
                self.sess.run(self.train_op,
                              feed_dict={
                                  self.sy_train_in: inputs[batch_idxs],
                                  self.sy_train_targ: targets[batch_idxs]
                              })
            idxs = shuffle_rows(idxs)
            if not hide_progress:
                if holdout_ratio < 1e-12:
                    epoch_range.set_postfix({
                        "Current loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in:
                                          inputs[idxs[:, :max_logging]],
                                          self.sy_train_targ:
                                          targets[idxs[:, :max_logging]]
                                      }),
                    })
                else:
                    epoch_range.set_postfix({
                        "Current loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in:
                                          inputs[idxs[:, :max_logging]],
                                          self.sy_train_targ:
                                          targets[idxs[:, :max_logging]]
                                      }),
                        "Holdout loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in: holdout_inputs,
                                          self.sy_train_targ: holdout_targets
                                      })
                    })

    def predict(self, inputs, factored=False, *args, **kwargs):
        """Returns the distribution predicted by the model for each input vector in inputs.
        Behavior is affected by the dimensionality of inputs and factored as follows:

        inputs is 2D, factored=True: Each row is treated as an input vector.
            Returns a mean of shape [ensemble_size, batch_size, output_dim] and variance of shape
            [ensemble_size, batch_size, output_dim], where N(mean[i, j, :], diag([i, j, :])) is the
            predicted output distribution by the ith model in the ensemble on input vector j.

        inputs is 2D, factored=False: Each row is treated as an input vector.
            Returns a mean of shape [batch_size, output_dim] and variance of shape
            [batch_size, output_dim], where aggregation is performed as described in the paper.

        inputs is 3D, factored=True/False: Each row in the last dimension is treated as an input vector.
            Returns a mean of shape [ensemble_size, batch_size, output_dim] and variance of sha
            [ensemble_size, batch_size, output_dim], where N(mean[i, j, :], diag([i, j, :])) is the
            predicted output distribution by the ith model in the ensemble on input vector [i, j].

        Arguments:
            inputs (np.ndarray): An array of input vectors in rows. See above for behavior.
            factored (bool): See above for behavior.
        """
        if len(inputs.shape) == 2:
            if factored:
                mean = self.sess.run(self.sy_pred_mean2d_fac,
                                     feed_dict={self.sy_pred_in2d: inputs})
                return mean, None
            else:
                return self.sess.run([self.sy_pred_mean2d, self.sy_pred_var2d],
                                     feed_dict={self.sy_pred_in2d: inputs})
        else:
            mean = self.sess.run(self.sy_pred_mean3d_fac,
                                 feed_dict={self.sy_pred_in3d: inputs})
            return mean, None

    def create_prediction_tensors(self,
                                  inputs,
                                  factored=False,
                                  *args,
                                  **kwargs):
        """See predict() above for documentation.
        """
        factored_mean = self._compile_outputs(inputs)
        if inputs.shape.ndims == 2 and not factored:
            mean = tf.reduce_mean(factored_mean, axis=0)
            variance = tf.reduce_mean(tf.square(factored_mean - mean), axis=0)
            return mean, variance
        return factored_mean, None

    def save(self, savedir=None):
        """Saves all information required to recreate this model in two files in savedir
        (or self.model_dir if savedir is None), one containing the model structuure and the other
        containing all variables in the network.

        savedir (str): (Optional) Path to which files will be saved. If not provided, self.model_dir
            (the directory provided at initialization) will be used.
        """
        if not self.finalized:
            raise RuntimeError()
        model_dir = self.model_dir if savedir is None else savedir

        # Write structure to file
        with open(os.path.join(model_dir, "%s.nns" % self.name), "w+") as f:
            for layer in self.layers:
                f.write("%s\n" % repr(layer))

        # Save network parameters (including scalers) in a .mat file
        var_vals = {}
        for i, var_val in enumerate(
                self.sess.run(self.nonoptvars + self.optvars)):
            var_vals[str(i)] = var_val
        savemat(os.path.join(model_dir, "%s.mat" % self.name), var_vals)

    def _load_structure(self):
        """Uses the saved structure in self.model_dir with the name of this network to initialize
        the structure of this network.
        """
        structure = []
        with open(os.path.join(self.model_dir, "%s.nns" % self.name),
                  "r") as f:
            for line in f:
                kwargs = {
                    key: val
                    for (key, val) in
                    [argval.split("=") for argval in line[3:-2].split(", ")]
                }
                kwargs["input_dim"] = int(kwargs["input_dim"])
                kwargs["output_dim"] = int(kwargs["output_dim"])
                kwargs["weight_decay"] = None if kwargs[
                    "weight_decay"] == "None" else float(
                        kwargs["weight_decay"])
                kwargs["activation"] = None if kwargs[
                    "activation"] == "None" else kwargs["activation"][1:-1]
                kwargs["ensemble_size"] = int(kwargs["ensemble_size"])
                structure.append(FC(**kwargs))
        self.layers = structure

    #######################
    # Compilation methods #
    #######################

    def _compile_outputs(self, inputs):
        cur_out = self.scaler.transform(inputs)
        for layer in self.layers:
            cur_out = layer.compute_output_tensor(cur_out)
        return cur_out

    def _compile_losses(self, inputs, targets):
        mean = self._compile_outputs(inputs)
        return tf.reduce_mean(tf.reduce_mean(tf.square(mean - targets) / 2,
                                             axis=-1),
                              axis=-1)
예제 #3
0
class BNN:
    """Neural network models which model aleatoric uncertainty (and possibly epistemic uncertainty
    with ensembling).
    """
    def __init__(self, params):
        """Initializes a class instance.

        Arguments:
            params (DotMap): A dotmap of model parameters.
                .name (str): Model name, used for logging/use in variable scopes.
                    Warning: Models with the same name will overwrite each other.
                .num_networks (int): (optional) The number of networks in the ensemble. Defaults to 1.
                    Ignored if model is being loaded.
                .model_dir (str/None): (optional) Path to directory from which model will be loaded, and
                    saved by default. Defaults to None.
                .load_model (bool): (optional) If True, model will be loaded from the model directory,
                    assuming that the files are generated by a model of the same name. Defaults to False.
                .sess (tf.Session/None): The session that this model will use.
                    If None, creates a session with its own associated graph. Defaults to None.
        """
        self.name = get_required_argument(params, 'name', 'Must provide name.')
        self.model_dir = params.get('model_dir', None)

        if params.get('sess', None) is None:
            config = tf.ConfigProto()
            # config.gpu_options.allow_growth = True
            self._sess = tf.Session(config=config)
        else:
            self._sess = params.get('sess')

        # Instance variables
        self.finalized = False
        self.layers, self.max_logvar, self.min_logvar = [], None, None
        self.decays, self.optvars, self.nonoptvars = [], [], []
        self.end_act, self.end_act_name = None, None
        self.scaler = None

        # Training objects
        self.optimizer = None
        self.sy_train_in, self.sy_train_targ = None, None
        self.train_op, self.mse_loss = None, None

        # Prediction objects
        self.sy_pred_in2d, self.sy_pred_mean2d_fac, self.sy_pred_var2d_fac = None, None, None
        self.sy_pred_mean2d, self.sy_pred_var2d = None, None
        self.sy_pred_in3d, self.sy_pred_mean3d_fac, self.sy_pred_var3d_fac = None, None, None

        if params.get('load_model', False):
            if self.model_dir is None:
                raise ValueError(
                    "Cannot load model without providing model directory.")
            self._load_structure()
            self.num_nets, self.model_loaded = self.layers[
                0].get_ensemble_size(), True
            print("Model loaded from %s." % self.model_dir)
        else:
            self.num_nets = params.get('num_networks', 1)
            self.model_loaded = False

        if self.num_nets == 1:
            print("Created a neural network with variance predictions.")
        else:
            print(
                "Created an ensemble of %d neural networks with variance predictions."
                % (self.num_nets))

    @property
    def is_probabilistic(self):
        return True

    @property
    def is_tf_model(self):
        return True

    @property
    def sess(self):
        return self._sess

    ###################################
    # Network Structure Setup Methods #
    ###################################

    def add(self, layer):
        """Adds a new layer to the network.

        Arguments:
            layer: (layer) The new layer to be added to the network.
                   If this is the first layer, the input dimension of the layer must be set.

        Returns: None.
        """
        if self.finalized:
            raise RuntimeError(
                "Cannot modify network structure after finalizing.")
        if len(self.layers) == 0 and layer.get_input_dim() is None:
            raise ValueError("Must set input dimension for the first layer.")
        if self.model_loaded:
            raise RuntimeError("Cannot add layers to a loaded model.")

        layer.set_ensemble_size(self.num_nets)
        if len(self.layers) > 0:
            layer.set_input_dim(self.layers[-1].get_output_dim())
        self.layers.append(layer.copy())

    def pop(self):
        """Removes and returns the most recently added layer to the network.

        Returns: (layer) The removed layer.
        """
        if len(self.layers) == 0:
            raise RuntimeError("Network is empty.")
        if self.finalized:
            raise RuntimeError(
                "Cannot modify network structure after finalizing.")
        if self.model_loaded:
            raise RuntimeError("Cannot remove layers from a loaded model.")

        return self.layers.pop()

    def finalize(self, optimizer, optimizer_args=None, *args, **kwargs):
        """Finalizes the network.

        Arguments:
            optimizer: (tf.train.Optimizer) An optimizer class from those available at tf.train.Optimizer.
            optimizer_args: (dict) A dictionary of arguments for the __init__ method of the chosen optimizer.

        Returns: None
        """
        if len(self.layers) == 0:
            raise RuntimeError("Cannot finalize an empty network.")
        if self.finalized:
            raise RuntimeError("Can only finalize a network once.")

        optimizer_args = {} if optimizer_args is None else optimizer_args
        self.optimizer = optimizer(**optimizer_args)

        out_dim = self.layers[-1].get_output_dim()
        # Add variance output.
        self.layers[-1].set_output_dim(2 * out_dim)

        # Remove last activation to isolate variance from activation function.
        self.end_act = self.layers[-1].get_activation()
        self.end_act_name = self.layers[-1].get_activation(as_func=False)
        self.layers[-1].unset_activation()

        self.recalibrator = RecalibrationLayer(out_dim)

        self.cal_vars = self.recalibrator.get_vars()

        # Construct all variables.
        with self.sess.as_default():
            with tf.variable_scope(self.name):
                self.scaler = TensorStandardScaler(
                    self.layers[0].get_input_dim())
                self.max_logvar = tf.Variable(
                    np.ones([1, self.layers[-1].get_output_dim() // 2]) / 2.,
                    dtype=tf.float32,
                    name="max_log_var")
                self.min_logvar = tf.Variable(
                    -np.ones([1, self.layers[-1].get_output_dim() // 2]) * 10.,
                    dtype=tf.float32,
                    name="min_log_var")
                for i, layer in enumerate(self.layers):
                    with tf.variable_scope("Layer%i" % i):
                        layer.construct_vars()
                        self.decays.extend(layer.get_decays())
                        self.optvars.extend(layer.get_vars())
        self.optvars.extend([self.max_logvar, self.min_logvar])
        self.nonoptvars.extend(self.scaler.get_vars())

        # Set up training
        with tf.variable_scope(self.name):
            self.optimizer = optimizer(**optimizer_args)
            self.sy_train_in = tf.placeholder(
                dtype=tf.float32,
                shape=[self.num_nets, None, self.layers[0].get_input_dim()],
                name="training_inputs")
            self.sy_train_targ = tf.placeholder(
                dtype=tf.float32,
                shape=[
                    self.num_nets, None, self.layers[-1].get_output_dim() // 2
                ],
                name="training_targets")
            train_loss = tf.reduce_sum(
                self._compile_losses(self.sy_train_in,
                                     self.sy_train_targ,
                                     inc_var_loss=True))
            train_loss += tf.add_n(self.decays)
            train_loss += 0.01 * tf.reduce_sum(
                self.max_logvar) - 0.01 * tf.reduce_sum(self.min_logvar)
            self.mse_loss = self._compile_losses(self.sy_train_in,
                                                 self.sy_train_targ,
                                                 inc_var_loss=False)

            self.train_op = self.optimizer.minimize(train_loss,
                                                    var_list=self.optvars)

        with tf.variable_scope('calibration'):
            self.sy_cdf_in = tf.placeholder(
                dtype=tf.float32,
                shape=[None, self.recalibrator.get_output_dim()],
                name="training_inputs_cdf")

            self.sy_cdf_true = tf.placeholder(
                dtype=tf.float32,
                shape=[None, self.recalibrator.get_output_dim()],
                name="training_targets_cdf")

            self.cal_optimizer = tf.train.AdamOptimizer(learning_rate=5e-2)

            cdf_pred = self.recalibrator(self.sy_cdf_in, activation=False)
            cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(
                labels=self.sy_cdf_true, logits=cdf_pred)
            self.cal_loss = tf.reduce_mean(tf.reduce_mean(cross_entropy,
                                                          axis=-1),
                                           axis=-1)

            self.cal_train_op = self.cal_optimizer.minimize(
                self.cal_loss, var_list=self.cal_vars)

        # Initialize all variables
        self.sess.run(
            tf.variables_initializer(self.optvars + self.nonoptvars +
                                     self.optimizer.variables() +
                                     self.cal_vars +
                                     self.cal_optimizer.variables()))

        # Set up prediction
        with tf.variable_scope(self.name):
            self.sy_pred_in2d = tf.placeholder(
                dtype=tf.float32,
                shape=[None, self.layers[0].get_input_dim()],
                name="2D_training_inputs")
            self.sy_pred_mean2d_fac, self.sy_pred_var2d_fac = \
                self.create_prediction_tensors(self.sy_pred_in2d, factored=True)
            self.sy_pred_mean2d = tf.reduce_mean(self.sy_pred_mean2d_fac,
                                                 axis=0)
            self.sy_pred_var2d = tf.reduce_mean(self.sy_pred_var2d_fac, axis=0) + \
                tf.reduce_mean(tf.square(self.sy_pred_mean2d_fac - self.sy_pred_mean2d), axis=0)

            self.sy_pred_in3d = tf.placeholder(
                dtype=tf.float32,
                shape=[self.num_nets, None, self.layers[0].get_input_dim()],
                name="3D_training_inputs")
            self.sy_pred_mean3d_fac, self.sy_pred_var3d_fac = \
                self.create_prediction_tensors(self.sy_pred_in3d, factored=True)

        # Load model if needed
        if self.model_loaded:
            with self.sess.as_default():
                params_dict = loadmat(
                    os.path.join(self.model_dir, "%s.mat" % self.name))
                all_vars = self.nonoptvars + self.optvars + self.cal_vars
                for i, var in enumerate(all_vars):
                    var.load(params_dict[str(i)])
        self.finalized = True

    #################
    # Model Methods #
    #################
    def calibrate(self,
                  inputs,
                  targets,
                  batch_size=32,
                  epochs=1000,
                  hide_progress=False,
                  holdout_ratio=0.0,
                  max_logging=5000):
        """Calibrates network post-training

        Arguments:
            inputs (np.ndarray): Network inputs in the training dataset in rows.
            targets (np.ndarray): Network target outputs in the training dataset in rows corresponding
                to the rows in inputs.
            batch_size (int): The minibatch size to be used for training.
            epochs (int): Number of epochs (full network passes that will be done.
            hide_progress (bool): If True, hides the progress bar shown at the beginning of training.

        Returns: None
        """
        with self.sess.as_default():
            self.scaler.fit(inputs)

        all_mus, all_vars = self.predict(inputs)
        all_ys = targets

        train_x = np.zeros_like(all_ys)
        train_y = np.zeros_like(all_ys)

        for d in range(all_mus.shape[1]):
            mu = all_mus[:, d]
            var = all_vars[:, d]
            ys = all_ys[:, d]

            cdf_pred = norm.cdf(ys, loc=mu, scale=np.sqrt(var))
            cdf_true = np.array(
                [np.sum(cdf_pred < p) / len(cdf_pred) for p in cdf_pred])

            train_x[:, d] = cdf_pred
            train_y[:, d] = cdf_true

        epochs = 200

        if hide_progress:
            epoch_range = range(epochs)
        else:
            epoch_range = trange(epochs,
                                 unit="epoch(s)",
                                 desc="Calibration training")

        def iterate_minibatches(inp, targs, batchsize, shuffle=True):
            assert inp.shape[0] == targs.shape[0]
            indices = np.arange(inp.shape[0])
            if shuffle:
                np.random.shuffle(indices)

            last_idx = 0

            for curr_idx in range(0, inp.shape[0] - batchsize + 1, batchsize):
                curr_batch = indices[curr_idx:curr_idx + batchsize]
                last_idx = curr_idx + batchsize
                yield inp[curr_batch], targs[curr_batch]

            if inp.shape[0] % batchsize != 0:
                last_batch = indices[last_idx:]
                yield inp[last_batch], targs[last_batch]

        for _ in epoch_range:
            for x_batch, y_batch in iterate_minibatches(
                    train_x, train_y, batch_size):
                self.sess.run(self.cal_train_op,
                              feed_dict={
                                  self.sy_cdf_in: x_batch,
                                  self.sy_cdf_true: y_batch
                              })

            if not hide_progress:
                epoch_range.set_postfix({
                    "Training loss(es)":
                    self.sess.run(self.cal_loss,
                                  feed_dict={
                                      self.sy_cdf_in: train_x,
                                      self.sy_cdf_true: train_y
                                  })
                })

    def save_calibration_info(self, inputs, targets, save_dir, calibrate=True):

        all_mus, all_vars = self.predict(inputs)
        all_ys = targets

        all_cdfs_pred = norm.cdf(all_ys, loc=all_mus, scale=np.sqrt(all_vars))
        all_cdfs_pred_cal = self.sess.run(
            self.recalibrator(all_cdfs_pred)) if calibrate else []

        # Save network parameters (including scalers) in a .mat file
        save_vals = {
            'all_mus': all_mus,
            'all_vars': all_mus,
            'all_ys': targets,
            'all_cdfs_pred': all_cdfs_pred,
            'all_cdfs_pred_cal': all_cdfs_pred_cal
        }

        savemat(os.path.join(save_dir, "calib_logs.mat"), save_vals)

    def plot_calibration(self, inputs, targets, save_dir):

        all_mus, all_vars = self.predict(inputs)
        all_ys = targets

        timestamp = datetime.now().strftime("%Y%m%d-%H%M%S")

        all_cdfs_pred = norm.cdf(all_ys, loc=all_mus, scale=np.sqrt(all_vars))
        all_cdfs_pred_cal = self.sess.run(self.recalibrator(all_cdfs_pred))

        for d in range(all_mus.shape[1]):
            mu = all_mus[:, d]
            var = all_vars[:, d]
            ys = all_ys[:, d]

            cdf_pred = all_cdfs_pred[:, d]
            cdf_pred_cal = all_cdfs_pred_cal[:, d]

            cal_ps = np.linspace(0, 1, num=30)

            cdf_emp = [np.sum(cdf_pred < p) / len(cdf_pred) for p in cal_ps]
            cdf_emp_cal = [
                np.sum(cdf_pred_cal < p) / len(cdf_pred_cal) for p in cal_ps
            ]

            plt.close('all')
            plt.figure()
            ax1 = plt.subplot2grid((3, 1), (0, 0), rowspan=2)
            ax2 = plt.subplot2grid((3, 1), (2, 0))

            ax1.set_title('Calibration Curve BNN (dim={}, t={})'.format(
                d, timestamp))
            ax1.set_xlabel('Expected confidence level')
            ax1.set_ylabel('Observed confidence level')
            ax1.plot(cal_ps, cdf_emp, "s-", label='Uncalibrated')
            ax1.plot(cal_ps, cdf_emp_cal, "s-", label='Calibrated')
            ax1.plot(cal_ps, cal_ps, alpha=0.6, color='gray')
            ax1.legend()

            ax2.hist(cdf_pred,
                     range=(0, 1),
                     bins=10,
                     label='Uncalibrated',
                     histtype="step",
                     lw=2)
            ax2.hist(cdf_pred_cal,
                     range=(0, 1),
                     bins=10,
                     label='Calibrated',
                     histtype="step",
                     lw=2)
            ax2.set_xlabel("Probability of predicted value")
            ax2.set_ylabel("Count")
            ax2.legend(loc="upper left", ncol=2)

            plt.tight_layout()

            print('Saving dim={}'.format(d))
            plt.savefig(
                os.path.join(save_dir,
                             'cal_{}_dim_{}.png'.format(timestamp, d)))

    def sample_predictions(self, means, var, calibrate=True):
        """
            Input shape of mean and var is N x d where N
            is batch size and d is size of state space dimension
        """
        if not calibrate:
            return means + tf.random_normal(
                shape=tf.shape(means), mean=0, stddev=1) * tf.sqrt(var)

        ps = tf.random.uniform(shape=means.shape)

        ps = self.recalibrator.inv_call(ps, activation=True)
        ps = tf.clip_by_value(ps, 1e-6, 1 - 1e-6)

        dist = tfp.distributions.Normal(loc=means, scale=tf.sqrt(var))
        ret = dist.quantile(ps)

        return ret

    def train(self,
              inputs,
              targets,
              batch_size=32,
              epochs=100,
              hide_progress=False,
              holdout_ratio=0.0,
              max_logging=5000):
        """Trains/Continues network training

        Arguments:
            inputs (np.ndarray): Network inputs in the training dataset in rows.
            targets (np.ndarray): Network target outputs in the training dataset in rows corresponding
                to the rows in inputs.
            batch_size (int): The minibatch size to be used for training.
            epochs (int): Number of epochs (full network passes that will be done.
            hide_progress (bool): If True, hides the progress bar shown at the beginning of training.

        Returns: None
        """
        def shuffle_rows(arr):
            idxs = np.argsort(np.random.uniform(size=arr.shape), axis=-1)
            return arr[np.arange(arr.shape[0])[:, None], idxs]

        # Split into training and holdout sets
        num_holdout = min(int(inputs.shape[0] * holdout_ratio), max_logging)
        permutation = np.random.permutation(inputs.shape[0])
        inputs, holdout_inputs = inputs[permutation[num_holdout:]], inputs[
            permutation[:num_holdout]]
        targets, holdout_targets = targets[permutation[num_holdout:]], targets[
            permutation[:num_holdout]]
        holdout_inputs = np.tile(holdout_inputs[None], [self.num_nets, 1, 1])
        holdout_targets = np.tile(holdout_targets[None], [self.num_nets, 1, 1])

        with self.sess.as_default():
            self.scaler.fit(inputs)

        idxs = np.random.randint(inputs.shape[0],
                                 size=[self.num_nets, inputs.shape[0]])
        if hide_progress:
            epoch_range = range(epochs)
        else:
            epoch_range = trange(epochs,
                                 unit="epoch(s)",
                                 desc="Network training")
        for _ in epoch_range:
            for batch_num in range(int(np.ceil(idxs.shape[-1] / batch_size))):
                batch_idxs = idxs[:, batch_num * batch_size:(batch_num + 1) *
                                  batch_size]
                self.sess.run(self.train_op,
                              feed_dict={
                                  self.sy_train_in: inputs[batch_idxs],
                                  self.sy_train_targ: targets[batch_idxs]
                              })
            idxs = shuffle_rows(idxs)
            if not hide_progress:
                if holdout_ratio < 1e-12:
                    epoch_range.set_postfix({
                        "Training loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in:
                                          inputs[idxs[:, :max_logging]],
                                          self.sy_train_targ:
                                          targets[idxs[:, :max_logging]]
                                      })
                    })
                else:
                    epoch_range.set_postfix({
                        "Training loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in:
                                          inputs[idxs[:, :max_logging]],
                                          self.sy_train_targ:
                                          targets[idxs[:, :max_logging]]
                                      }),
                        "Holdout loss(es)":
                        self.sess.run(self.mse_loss,
                                      feed_dict={
                                          self.sy_train_in: holdout_inputs,
                                          self.sy_train_targ: holdout_targets
                                      })
                    })

    def predict(self, inputs, factored=False, *args, **kwargs):
        """Returns the distribution predicted by the model for each input vector in inputs.
        Behavior is affected by the dimensionality of inputs and factored as follows:

        inputs is 2D, factored=True: Each row is treated as an input vector.
            Returns a mean of shape [ensemble_size, batch_size, output_dim] and variance of shape
            [ensemble_size, batch_size, output_dim], where N(mean[i, j, :], diag([i, j, :])) is the
            predicted output distribution by the ith model in the ensemble on input vector j.

        inputs is 2D, factored=False: Each row is treated as an input vector.
            Returns a mean of shape [batch_size, output_dim] and variance of shape
            [batch_size, output_dim], where aggregation is performed as described in the paper.

        inputs is 3D, factored=True/False: Each row in the last dimension is treated as an input vector.
            Returns a mean of shape [ensemble_size, batch_size, output_dim] and variance of sha
            [ensemble_size, batch_size, output_dim], where N(mean[i, j, :], diag([i, j, :])) is the
            predicted output distribution by the ith model in the ensemble on input vector [i, j].

        Arguments:
            inputs (np.ndarray): An array of input vectors in rows. See above for behavior.
            factored (bool): See above for behavior.
        """
        if len(inputs.shape) == 2:
            if factored:
                return self.sess.run(
                    [self.sy_pred_mean2d_fac, self.sy_pred_var2d_fac],
                    feed_dict={self.sy_pred_in2d: inputs})
            else:
                return self.sess.run([self.sy_pred_mean2d, self.sy_pred_var2d],
                                     feed_dict={self.sy_pred_in2d: inputs})
        else:
            return self.sess.run(
                [self.sy_pred_mean3d_fac, self.sy_pred_var3d_fac],
                feed_dict={self.sy_pred_in3d: inputs})

    def create_prediction_tensors(self,
                                  inputs,
                                  factored=False,
                                  *args,
                                  **kwargs):
        """See predict() above for documentation.
        """
        factored_mean, factored_variance = self._compile_outputs(inputs)
        if inputs.shape.ndims == 2 and not factored:
            mean = tf.reduce_mean(factored_mean, axis=0)
            variance = tf.reduce_mean(tf.square(factored_mean - mean), axis=0) + \
                       tf.reduce_mean(factored_variance, axis=0)
            return mean, variance
        return factored_mean, factored_variance

    def save(self, savedir=None):
        """Saves all information required to recreate this model in two files in savedir
        (or self.model_dir if savedir is None), one containing the model structuure and the other
        containing all variables in the network.

        savedir (str): (Optional) Path to which files will be saved. If not provided, self.model_dir
            (the directory provided at initialization) will be used.
        """
        if not self.finalized:
            raise RuntimeError()
        model_dir = self.model_dir if savedir is None else savedir

        # Write structure to file
        with open(os.path.join(model_dir, "%s.nns" % self.name), "w+") as f:
            for layer in self.layers[:-1]:
                f.write("%s\n" % repr(layer))
            last_layer_copy = self.layers[-1].copy()
            last_layer_copy.set_activation(self.end_act_name)
            last_layer_copy.set_output_dim(last_layer_copy.get_output_dim() //
                                           2)
            f.write("%s\n" % repr(last_layer_copy))

        # Save network parameters (including scalers) in a .mat file
        var_vals = {}
        for i, var_val in enumerate(
                self.sess.run(self.nonoptvars + self.optvars + self.cal_vars)):
            var_vals[str(i)] = var_val
        savemat(os.path.join(model_dir, "%s.mat" % self.name), var_vals)

    def _load_structure(self):
        """Uses the saved structure in self.model_dir with the name of this network to initialize
        the structure of this network.
        """
        structure = []
        with open(os.path.join(self.model_dir, "%s.nns" % self.name),
                  "r") as f:
            for line in f:
                kwargs = {
                    key: val
                    for (key, val) in
                    [argval.split("=") for argval in line[3:-2].split(", ")]
                }
                kwargs["input_dim"] = int(kwargs["input_dim"])
                kwargs["output_dim"] = int(kwargs["output_dim"])
                kwargs["weight_decay"] = None if kwargs[
                    "weight_decay"] == "None" else float(
                        kwargs["weight_decay"])
                kwargs["activation"] = None if kwargs[
                    "activation"] == "None" else kwargs["activation"][1:-1]
                kwargs["ensemble_size"] = int(kwargs["ensemble_size"])
                structure.append(FC(**kwargs))
        self.layers = structure

    #######################
    # Compilation methods #
    #######################

    def _compile_outputs(self, inputs, ret_log_var=False):
        """Compiles the output of the network at the given inputs.

        If inputs is 2D, returns a 3D tensor where output[i] is the output of the ith network in the ensemble.
        If inputs is 3D, returns a 3D tensor where output[i] is the output of the ith network on the ith input matrix.

        Arguments:
            inputs: (tf.Tensor) A tensor representing the inputs to the network
            ret_log_var: (bool) If True, returns the log variance instead of the variance.

        Returns: (tf.Tensors) The mean and variance/log variance predictions at inputs for each network
            in the ensemble.
        """
        dim_output = self.layers[-1].get_output_dim()
        cur_out = self.scaler.transform(inputs)
        for layer in self.layers:
            cur_out = layer.compute_output_tensor(cur_out)

        mean = cur_out[:, :, :dim_output // 2]
        if self.end_act is not None:
            mean = self.end_act(mean)

        logvar = self.max_logvar - tf.nn.softplus(
            self.max_logvar - cur_out[:, :, dim_output // 2:])
        logvar = self.min_logvar + tf.nn.softplus(logvar - self.min_logvar)

        if ret_log_var:
            return mean, logvar
        else:
            return mean, tf.exp(logvar)

    def _compile_losses(self, inputs, targets, inc_var_loss=True):
        """Helper method for compiling the loss function.

        The loss function is obtained from the log likelihood, assuming that the output
        distribution is Gaussian, with both mean and (diagonal) covariance matrix being determined
        by network outputs.

        Arguments:
            inputs: (tf.Tensor) A tensor representing the input batch
            targets: (tf.Tensor) The desired targets for each input vector in inputs.
            inc_var_loss: (bool) If True, includes log variance loss.

        Returns: (tf.Tensor) A tensor representing the loss on the input arguments.
        """
        mean, log_var = self._compile_outputs(inputs, ret_log_var=True)
        inv_var = tf.exp(-log_var)

        if inc_var_loss:
            mse_losses = tf.reduce_mean(tf.reduce_mean(
                tf.square(mean - targets) * inv_var, axis=-1),
                                        axis=-1)
            var_losses = tf.reduce_mean(tf.reduce_mean(log_var, axis=-1),
                                        axis=-1)
            total_losses = mse_losses + var_losses
        else:
            total_losses = tf.reduce_mean(tf.reduce_mean(tf.square(mean -
                                                                   targets),
                                                         axis=-1),
                                          axis=-1)

        return total_losses