Python keras.backend.squeeze() Examples

def _compute_probabilities(self, energy, previous_attention=None):
        if self.is_monotonic:
            # add presigmoid noise to encourage discreteness
            sigmoid_noise = K.in_train_phase(1., 0.)
            noise = K.random_normal(K.shape(energy), mean=0.0, stddev=sigmoid_noise)
            # encourage discreteness in train
            energy = K.in_train_phase(energy + noise, energy)

            p = K.in_train_phase(K.sigmoid(energy),
                                 K.cast(energy > 0, energy.dtype))
            p = K.squeeze(p, -1)
            p_prev = K.squeeze(previous_attention, -1)
            # monotonic attention function from tensorflow
            at = K.in_train_phase(
                tf.contrib.seq2seq.monotonic_attention(p, p_prev, 'parallel'),
                tf.contrib.seq2seq.monotonic_attention(p, p_prev, 'hard'))
            at = K.expand_dims(at, -1)
        else:
            # softmax
            at = keras.activations.softmax(energy, axis=1)

        return at 

def _backward(gamma, mask):
    """Backward recurrence of the linear chain crf."""
    gamma = K.cast(gamma, 'int32')

    def _backward_step(gamma_t, states):
        y_tm1 = K.squeeze(states[0], 0)
        y_t = batch_gather(gamma_t, y_tm1)
        return y_t, [K.expand_dims(y_t, 0)]

    initial_states = [K.expand_dims(K.zeros_like(gamma[:, 0, 0]), 0)]
    _, y_rev, _ = K.rnn(_backward_step,
                        gamma,
                        initial_states,
                        go_backwards=True)
    y = K.reverse(y_rev, 1)

    if mask is not None:
        mask = K.cast(mask, dtype='int32')
        # mask output
        y *= mask
        # set masked values to -1
        y += -(1 - mask)
    return y 

def segmentation_gaussian_binary_crossentropy(
        y_true,
        y_pred,
        gaussian_sigma=3):
    with K.name_scope(name='segmentation_gaussian_binary_crossentropy') as scope:
        if keras.backend.ndim(y_true) == 4:
            # sometimes the dimensions are expanded from 2 to 4
            # to meet Keras' expectations.
            # In that case reduce them back to 2
            y_true = K.squeeze(y_true, axis=-1)
            y_true = K.squeeze(y_true, axis=-1)
        results = segmentation_gaussian_measurement_batch(
            y_true, y_pred,
            measurement=segmentation_losses.binary_crossentropy,
            gaussian_sigma=gaussian_sigma)
        return results 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def step(self, x_input, states):
    	#print "x_input:", x_input, x_input.shape
    	# <TensorType(float32, matrix)>
    	
        input_shape = self.input_spec[0].shape
        en_seq = states[-1]
        _, [h, c] = super(PointerLSTM, self).step(x_input, states[:-1])

        # vt*tanh(W1*e+W2*d)
        dec_seq = K.repeat(h, input_shape[1])
        Eij = time_distributed_dense(en_seq, self.W1, output_dim=1)
        Dij = time_distributed_dense(dec_seq, self.W2, output_dim=1)
        U = self.vt * tanh(Eij + Dij)
        U = K.squeeze(U, 2)

        # make probability tensor
        pointer = softmax(U)
        return pointer, [h, c] 

def dot_product(x, kernel):
    """
    Wrapper for dot product operation used inthe attention layers, in order to be compatible with both
    Theano and Tensorflow
    Args:
        x (): input
        kernel (): weights
    Returns:
    """
    if K.backend() == 'tensorflow':
        # todo: check that this is correct
        return K.squeeze(K.dot(x, K.expand_dims(kernel)), axis=-1)
    else:
        return K.dot(x, kernel)


# read list of words (one per line), e.g. stopwords, badwords 

def _backward(gamma, mask):
    '''Backward recurrence of the linear chain crf.'''
    gamma = K.cast(gamma, 'int32')

    def _backward_step(gamma_t, states):
        y_tm1 = K.squeeze(states[0], 0)
        y_t = batch_gather(gamma_t, y_tm1)
        return y_t, [K.expand_dims(y_t, 0)]

    initial_states = [K.expand_dims(K.zeros_like(gamma[:, 0, 0]), 0)]
    _, y_rev, _ = K.rnn(_backward_step,
                        gamma,
                        initial_states,
                        go_backwards=True)
    y = K.reverse(y_rev, 1)

    if mask is not None:
        mask = K.cast(mask, dtype='int32')
        # mask output
        y *= mask
        # set masked values to -1
        y += -(1 - mask)
    return y 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.
    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def _backward(gamma, mask):
    '''Backward recurrence of the linear chain crf.'''
    gamma = K.cast(gamma, 'int32')

    def _backward_step(gamma_t, states):
        y_tm1 = K.squeeze(states[0], 0)
        y_t = batch_gather(gamma_t, y_tm1)
        return y_t, [K.expand_dims(y_t, 0)]

    initial_states = [K.expand_dims(K.zeros_like(gamma[:, 0, 0]), 0)]
    _, y_rev, _ = K.rnn(_backward_step,
                        gamma,
                        initial_states,
                        go_backwards=True)
    y = K.reverse(y_rev, 1)

    if mask is not None:
        mask = K.cast(mask, dtype='int32')
        # mask output
        y *= mask
        # set masked values to -1
        y += -(1 - mask)
    return y 

def gripper_coordinate_y_true(y_true, y_pred=None):
    """ Get the label found in y_true which also contains coordinates.

    # Arguments

        y_true: [ground_truth_label, y_height_coordinate, x_width_coordinate]
            Shape of y_true is [batch_size, 3].
        y_pred: Predicted values with shape [batch_size, img_height, img_width, 1].
    """
    with K.name_scope(name="gripper_coordinate_y_true") as scope:
        if keras.backend.ndim(y_true) == 4:
            # sometimes the dimensions are expanded from 2 to 4
            # to meet Keras' expectations.
            # In that case reduce them back to 2
            y_true = K.squeeze(y_true, axis=-1)
            y_true = K.squeeze(y_true, axis=-1)
        label = K.cast(y_true[:, :1], 'float32')
        return label 

def _target_class_loss(
            self,
            target_class,
            box_scores,
            box_class_probs_logits):
        """ Evaluate target_class_loss w.r.t. the input.

        """
        box_scores = K.squeeze(box_scores, axis=0)
        box_class_probs_logits = K.squeeze(box_class_probs_logits, axis=0)
        import tensorflow as tf
        boi_idx = tf.where(box_scores[:, target_class] > self._score)
        loss_box_class_conf = tf.reduce_mean(
            tf.gather(box_class_probs_logits[:, target_class], boi_idx))

        # Avoid the propagation of nan
        return tf.cond(
            tf.is_nan(loss_box_class_conf),
            lambda: tf.constant(0.),
            lambda: loss_box_class_conf) 

def gripper_coordinate_y_pred(y_true, y_pred):
    """ Get the predicted value at the coordinate found in y_true.

    # Arguments

        y_true: [ground_truth_label, y_height_coordinate, x_width_coordinate]
            Shape of y_true is [batch_size, 3].
        y_pred: Predicted values with shape [batch_size, img_height, img_width, 1].
    """
    with K.name_scope(name="gripper_coordinate_y_pred") as scope:
        if keras.backend.ndim(y_true) == 4:
            # sometimes the dimensions are expanded from 2 to 4
            # to meet Keras' expectations.
            # In that case reduce them back to 2
            y_true = K.squeeze(y_true, axis=-1)
            y_true = K.squeeze(y_true, axis=-1)
        yx_coordinate = K.cast(y_true[:, 1:], 'int32')
        yx_shape = K.shape(yx_coordinate)
        sample_index = K.expand_dims(K.arange(yx_shape[0]), axis=-1)
        byx_coordinate = K.concatenate([sample_index, yx_coordinate], axis=-1)

        # maybe need to transpose yx_coordinate?
        gripper_coordinate_y_predicted = tf.gather_nd(y_pred, byx_coordinate)
        return gripper_coordinate_y_predicted 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def call(self, inputs, **kwargs):
        """Evaluate YOLO model on given input and return filtered boxes."""
        yolo_outputs = inputs[0:-1]
        input_image_shape = K.squeeze(inputs[-1], axis=0)
        num_layers = len(yolo_outputs)
        anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] if num_layers == 3 else [[3, 4, 5],
                                                                                 [1, 2, 3]]  # default setting
        input_shape = K.shape(yolo_outputs[0])[1:3] * 32
        boxes = []
        box_scores = []
        for l in range(num_layers):
            _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l], self.anchors[anchor_mask[l]], self.num_classes,
                                                        input_shape, input_image_shape)
            boxes.append(_boxes)
            box_scores.append(_box_scores)
        boxes = K.concatenate(boxes, axis=0)
        box_scores = K.concatenate(box_scores, axis=0)
        return [boxes, box_scores] 

def preprocess_input(self, inputs, training=None):
        if self.window_size > 1:
            inputs = K.temporal_padding(inputs, (self.window_size - 1, 0))
        inputs = K.expand_dims(inputs, 2)  # add a dummy dimension

        output = K.conv2d(inputs, self.kernel, strides=self.strides,
                          padding='valid',
                          data_format='channels_last')
        output = K.squeeze(output, 2)  # remove the dummy dimension
        if self.use_bias:
            output = K.bias_add(output, self.bias, data_format='channels_last')

        if self.dropout is not None and 0. < self.dropout < 1.:
            z = output[:, :, :self.units]
            f = output[:, :, self.units:2 * self.units]
            o = output[:, :, 2 * self.units:]
            f = K.in_train_phase(1 - _dropout(1 - f, self.dropout), f, training=training)
            return K.concatenate([z, f, o], -1)
        else:
            return output 

def step(self, x, states):  
        h = states[0]
        # states[1] necessary?
        
        # comes from the constants
        X_static = states[-2]
        # equals K.dot(static_x, self._W1) + self._b2 with X.shape=[bs, L, static_input_dim]
        total_x_static_prod = states[-1]

        # expand dims to add the vector which is only valid for this time step
        # to total_x_prod which is valid for all time steps
        hw = K.expand_dims(K.dot(h, self._W2), 1)
        additive_atn = total_x_static_prod + hw
        attention = K.softmax(K.dot(additive_atn, self._V), axis=1)
        static_x_weighted = K.sum(attention * X_static, [1])
        
        x = K.dot(K.concatenate([x, static_x_weighted], 1), self._W3) + self._b3

        h, new_states = self.layer.cell.call(x, states[:-2])
        
        # append attention to the states to "smuggle" it out of the RNN wrapper
        attention = K.squeeze(attention, -1)
        h = K.concatenate([h, attention])

        return h, new_states 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def _backward(gamma, mask):
    '''Backward recurrence of the linear chain crf.'''
    gamma = K.cast(gamma, 'int32')

    def _backward_step(gamma_t, states):
        y_tm1 = K.squeeze(states[0], 0)
        y_t = batch_gather(gamma_t, y_tm1)
        return y_t, [K.expand_dims(y_t, 0)]

    initial_states = [K.expand_dims(K.zeros_like(gamma[:, 0, 0]), 0)]
    _, y_rev, _ = K.rnn(_backward_step,
                        gamma,
                        initial_states,
                        go_backwards=True)
    y = K.reverse(y_rev, 1)

    if mask is not None:
        mask = K.cast(mask, dtype='int32')
        # mask output
        y *= mask
        # set masked values to -1
        y += -(1 - mask)
    return y 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    '''RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    '''
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def call(self, x, mask=None):
        # size of x :[batch_size, sel_len, attention_dim]
        # size of u :[batch_size, attention_dim]
        # uit = tanh(xW+b)
        uit = K.tanh(K.bias_add(K.dot(x, self.W), self.b))
        ait = K.dot(uit, self.u)
        ait = K.squeeze(ait, -1)

        ait = K.exp(ait)

        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            ait *= K.cast(mask, K.floatx())
        ait /= K.cast(K.sum(ait, axis=1, keepdims=True) + K.epsilon(), K.floatx())
        ait = K.expand_dims(ait)
        weighted_input = x * ait
        output = K.sum(weighted_input, axis=1)

        return output 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for BG/FG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def call(self, x, mask=None):
        return K.squeeze(x, axis=self.axis) 

def viterbi_decoding(self, X, mask=None):
        input_energy = self.activation(K.dot(X, self.kernel) + self.bias)
        if self.use_boundary:
            input_energy = self.add_boundary_energy(
                input_energy, mask, self.left_boundary, self.right_boundary)

        argmin_tables = self.recursion(input_energy, mask, return_logZ=False)
        argmin_tables = K.cast(argmin_tables, 'int32')

        # backward to find best path, `initial_best_idx` can be any,
        # as all elements in the last argmin_table are the same
        argmin_tables = K.reverse(argmin_tables, 1)
        # matrix instead of vector is required by tf `K.rnn`
        initial_best_idx = [K.expand_dims(argmin_tables[:, 0, 0])]
        if K.backend() == 'theano':
            initial_best_idx = [K.T.unbroadcast(initial_best_idx[0], 1)]

        def gather_each_row(params, indices):
            n = K.shape(indices)[0]
            if K.backend() == 'theano':
                return params[K.T.arange(n), indices]
            else:
                indices = K.transpose(K.stack([K.tf.range(n), indices]))
                return K.tf.gather_nd(params, indices)

        def find_path(argmin_table, best_idx):
            next_best_idx = gather_each_row(argmin_table, best_idx[0][:, 0])
            next_best_idx = K.expand_dims(next_best_idx)
            if K.backend() == 'theano':
                next_best_idx = K.T.unbroadcast(next_best_idx, 1)
            return next_best_idx, [next_best_idx]

        _, best_paths, _ = K.rnn(find_path, argmin_tables, initial_best_idx,
                                 input_length=K.int_shape(X)[1], unroll=self.unroll)
        best_paths = K.reverse(best_paths, 1)
        best_paths = K.squeeze(best_paths, 2)

        return K.one_hot(best_paths, self.units) 

def rpn_bbox_loss_graph(config, target_bbox, rpn_match, rpn_bbox):
    """Return the RPN bounding box loss graph.

    config: the model config object.
    target_bbox: [batch, max positive anchors, (dy, dx, log(dh), log(dw))].
        Uses 0 padding to fill in unsed bbox deltas.
    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_bbox: [batch, anchors, (dy, dx, log(dh), log(dw))]
    """
    # Positive anchors contribute to the loss, but negative and
    # neutral anchors (match value of 0 or -1) don't.
    rpn_match = K.squeeze(rpn_match, -1)
    indices = tf.where(K.equal(rpn_match, 1))

    # Pick bbox deltas that contribute to the loss
    rpn_bbox = tf.gather_nd(rpn_bbox, indices)

    # Trim target bounding box deltas to the same length as rpn_bbox.
    batch_counts = K.sum(K.cast(K.equal(rpn_match, 1), tf.int32), axis=1)
    target_bbox = batch_pack_graph(target_bbox, batch_counts,
                                   config.IMAGES_PER_GPU)

    # TODO: use smooth_l1_loss() rather than reimplementing here
    #       to reduce code duplication
    diff = K.abs(target_bbox - rpn_bbox)
    less_than_one = K.cast(K.less(diff, 1.0), "float32")
    loss = (less_than_one * 0.5 * diff**2) + (1 - less_than_one) * (diff - 0.5)

    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_bbox_loss_graph(config, target_bbox, rpn_match, rpn_bbox):
    """Return the RPN bounding box loss graph.

    config: the model config object.
    target_bbox: [batch, max positive anchors, (dy, dx, log(dh), log(dw))].
        Uses 0 padding to fill in unsed bbox deltas.
    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_bbox: [batch, anchors, (dy, dx, log(dh), log(dw))]
    """
    # Positive anchors contribute to the loss, but negative and
    # neutral anchors (match value of 0 or -1) don't.
    rpn_match = K.squeeze(rpn_match, -1)
    indices = tf.where(K.equal(rpn_match, 1))

    # Pick bbox deltas that contribute to the loss
    rpn_bbox = tf.gather_nd(rpn_bbox, indices)

    # Trim target bounding box deltas to the same length as rpn_bbox.
    batch_counts = K.sum(K.cast(K.equal(rpn_match, 1), tf.int32), axis=1)
    target_bbox = batch_pack_graph(target_bbox, batch_counts,
                                   config.IMAGES_PER_GPU)

    # TODO: use smooth_l1_loss() rather than reimplementing here
    #       to reduce code duplication
    diff = K.abs(target_bbox - rpn_bbox)
    less_than_one = K.cast(K.less(diff, 1.0), "float32")
    loss = (less_than_one * 0.5 * diff**2) + (1 - less_than_one) * (diff - 0.5)

    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss