Python keras.backend.reshape() Examples

def categorical_crossentropy_with_class_rebal(y_true, y_pred):
    y_true = K.reshape(y_true, (-1, num_classes))
    y_pred = K.reshape(y_pred, (-1, num_classes))

    idx_max = K.argmax(y_true, axis=1)
    weights = K.gather(factor, idx_max)
    weights = K.reshape(weights, (-1, 1))

    # multiply y_true by weights
    y_true = y_true * weights

    cross_ent = K.categorical_crossentropy(y_pred, y_true)
    cross_ent = K.mean(cross_ent, axis=-1)

    return cross_ent


# getting the number of GPUs 

def call(self, x):
        mean = K.mean(x, axis=-1)
        std = K.std(x, axis=-1)

        if len(x.shape) == 3:
            mean = K.permute_dimensions(
                K.repeat(mean, x.shape.as_list()[-1]),
                [0,2,1]
            )
            std = K.permute_dimensions(
                K.repeat(std, x.shape.as_list()[-1]),
                [0,2,1] 
            )
            
        elif len(x.shape) == 2:
            mean = K.reshape(
                K.repeat_elements(mean, x.shape.as_list()[-1], 0),
                (-1, x.shape.as_list()[-1])
            )
            std = K.reshape(
                K.repeat_elements(mean, x.shape.as_list()[-1], 0),
                (-1, x.shape.as_list()[-1])
            )
        
        return self._g * (x - mean) / (std + self._epsilon) + self._b 

def softmax_sparse_crossentropy_ignoring_first_label(y_true, y_pred):
    y_pred = K.reshape(y_pred, (-1, K.int_shape(y_pred)[-1]))
    log_softmax = tf.nn.log_softmax(y_pred)

    y_true = K.one_hot(tf.to_int32(K.flatten(y_true)), K.int_shape(y_pred)[-1]+1)
    unpacked = tf.unstack(y_true, axis=-1)
    legal_labels = ~tf.cast(unpacked[0], tf.bool)
    y_true = tf.stack(unpacked[1:], axis=-1)

    cross_entropy = -K.sum(y_true * log_softmax, axis=1)
    cross_entropy_mean = K.sum(cross_entropy) / K.sum(tf.to_float(legal_labels))

    return cross_entropy_mean


# Accuracy for segmentation (ignoring first label) 

def sparse_accuracy(y_true, y_pred):
    classes = K.int_shape(y_pred)[-1]
    y_pred = K.reshape(y_pred, (-1, classes))

    y_true = K.one_hot(tf.to_int32(K.flatten(y_true)),
                       classes + 1)
    unpacked = tf.unstack(y_true, axis=-1)
    legal_labels = ~tf.cast(unpacked[0], tf.bool)
    y_true = tf.stack(unpacked[1:], axis=-1)

    return K.sum(tf.to_float(legal_labels & K.equal(K.argmax(y_true, axis=-1), K.argmax(y_pred, axis=-1)))) / K.sum(tf.to_float(legal_labels))


# Define different models


# 3D-FCN model 

def categorical_crossentropy_color(y_true, y_pred):
    q = 313
    y_true = K.reshape(y_true, (-1, q))
    y_pred = K.reshape(y_pred, (-1, q))

    idx_max = K.argmax(y_true, axis=1)
    weights = K.gather(prior_factor, idx_max)
    weights = K.reshape(weights, (-1, 1))

    # multiply y_true by weights
    y_true = y_true * weights

    cross_ent = K.categorical_crossentropy(y_pred, y_true)
    cross_ent = K.mean(cross_ent, axis=-1)

    return cross_ent


# getting the number of GPUs 

def get_output(self, train=False):
        def format_shape(shape):
            if K._BACKEND == 'tensorflow':
                def trf(x):
                    try:
                        return int(x)
                    except TypeError:
                        return x

                return map(trf, shape)
            return shape

        X = self.get_input(train)

        in_shape = format_shape(K.shape(X))
        batch_flatten_len = K.prod(in_shape[:2])
        cast_in_shape = (batch_flatten_len, ) + tuple(in_shape[i] for i in range(2, K.ndim(X)))
        
        pre_outs = self.layer(K.reshape(X, cast_in_shape))
        
        out_shape = format_shape(K.shape(pre_outs))
        cast_out_shape = (in_shape[0], in_shape[1]) + tuple(out_shape[i] for i in range(1, K.ndim(pre_outs)))
        
        outputs = K.reshape(pre_outs, cast_out_shape)
        return outputs 

def split_heads_2d(self, ip):
        tensor_shape = K.shape(ip)

        # batch, height, width, channels for axis = -1
        tensor_shape = [tensor_shape[i] for i in range(len(self._shape))]

        batch = tensor_shape[0]
        height = tensor_shape[1]
        width = tensor_shape[2]
        channels = tensor_shape[3]

        # Save the spatial tensor dimensions
        self._batch = batch
        self._height = height
        self._width = width

        ret_shape = K.stack([batch, height, width,  self.num_heads, channels // self.num_heads])
        split = K.reshape(ip, ret_shape)
        transpose_axes = (0, 3, 1, 2, 4)
        split = K.permute_dimensions(split, transpose_axes)

        return split 

def mmd(n_conditions, beta, kernel_method='multi-scale-rbf', computation_way="general"):
    def mmd_loss(real_labels, y_pred):
        with tf.variable_scope("mmd_loss", reuse=tf.AUTO_REUSE):
            real_labels = K.reshape(K.cast(real_labels, 'int32'), (-1,))
            conditions_mmd = tf.dynamic_partition(y_pred, real_labels, num_partitions=n_conditions)
            loss = 0.0
            if computation_way.isdigit():
                boundary = int(computation_way)
                for i in range(boundary):
                    for j in range(boundary, n_conditions):
                        loss += compute_mmd(conditions_mmd[i], conditions_mmd[j], kernel_method)
            else:
                for i in range(len(conditions_mmd)):
                    for j in range(i):
                        loss += compute_mmd(conditions_mmd[j], conditions_mmd[j + 1], kernel_method)
            return _nan2zero(beta * loss)

    return mmd_loss 

def step(self, x, states):
        x_shape = K.shape(x)
        h_tm1 = states[0]
        c_tm1 = states[1]

        e = self.V_a(K.tanh(self.W_a(h_tm1) + self.U_a(x)))
        a = K.reshape(K.softmax(K.batch_flatten(e)), (x_shape[0], 1, x_shape[2], x_shape[3]))
        x_tilde = x * K.repeat_elements(a, x_shape[1], 1)

        x_i = self.W_i(x_tilde)
        x_f = self.W_f(x_tilde)
        x_c = self.W_c(x_tilde)
        x_o = self.W_o(x_tilde)

        i = self.inner_activation(x_i + self.U_i(h_tm1))
        f = self.inner_activation(x_f + self.U_f(h_tm1))
        c = f * c_tm1 + i * self.activation(x_c + self.U_c(h_tm1))
        o = self.inner_activation(x_o + self.U_o(h_tm1))

        h = o * self.activation(c)
        return h, [h, c] 

def call(self, inputs):
        def wrapper(rois, mrcnn_class, mrcnn_bbox, image_meta):
            detections_batch = []
            _, _, window, _ = parse_image_meta(image_meta)
            for b in range(self.config.BATCH_SIZE):
                detections = refine_detections(
                    rois[b], mrcnn_class[b], mrcnn_bbox[b], window[b], self.config)
                # Pad with zeros if detections < DETECTION_MAX_INSTANCES
                gap = self.config.DETECTION_MAX_INSTANCES - detections.shape[0]
                assert gap >= 0
                if gap > 0:
                    detections = np.pad(
                        detections, [(0, gap), (0, 0)], 'constant', constant_values=0)
                detections_batch.append(detections)

            # Stack detections and cast to float32
            # TODO: track where float64 is introduced
            detections_batch = np.array(detections_batch).astype(np.float32)
            # Reshape output
            # [batch, num_detections, (y1, x1, y2, x2, class_score)] in pixels
            return np.reshape(detections_batch, [self.config.BATCH_SIZE, self.config.DETECTION_MAX_INSTANCES, 6])

        # Return wrapped function
        return tf.py_func(wrapper, inputs, tf.float32) 

def call(self , x, mask=None):
        
        e1=x[0].T
        e2=x[1].T
        
        batch_size = K.shape(x[0])[0]
        sim = []
        V_out = K.dot(self.V, K.concatenate([e1,e2],axis=0))     

        for i in range(self.k): 
            temp = K.batch_dot(K.dot(e1.T,self.W[i,:,:]),e2.T,axes=1)
            sim.append(temp)
        sim=K.reshape(sim,(self.k,batch_size))

        tensor_bi_product = self.activation(V_out+sim)
        tensor_bi_product = K.dot(self.U.T, tensor_bi_product).T

        return tensor_bi_product 

def call(self, x, mask=None):
        x = K.expand_dims(x, -1)  # add a dimension of the right
        x = K.permute_dimensions(x, (0, 2, 1, 3))

        # TF uses the last dimension as channel dimension,
        # instead of the 2nd one.
        # TH kernel shape: (depth, input_depth, rows, cols)
        # TF kernel shape: (rows, cols, input_depth, depth)

        # for us, we need to switch the rows with the columns?
        W = tf.transpose(self.tied_to.W, (1, 0, 2, 3))
        output = K.conv2d(x, W, strides=self.subsample,
                          border_mode=self.border_mode,
                          dim_ordering='th')
        if self.bias:
            output += K.reshape(self.b, (1, self.nb_filter, 1, 1))
        output = K.squeeze(output, 3)  # remove the dummy 3rd dimension
        output = K.permute_dimensions(output, (0, 2, 1))
        output = self.activation(output)
        return output 

def call(self, inputs):
        """Following the routing algorithm from Hinton's paper,
        but replace b = b + <u,v> with b = <u,v>.

        This change can improve the feature representation of Capsule.

        However, you can replace
            b = K.batch_dot(outputs, hat_inputs, [2, 3])
        with
            b += K.batch_dot(outputs, hat_inputs, [2, 3])
        to realize a standard routing.
        """

        if self.share_weights:
            hat_inputs = K.conv1d(inputs, self.kernel)
        else:
            hat_inputs = K.local_conv1d(inputs, self.kernel, [1], [1])

        batch_size = K.shape(inputs)[0]
        input_num_capsule = K.shape(inputs)[1]
        hat_inputs = K.reshape(hat_inputs,
                               (batch_size, input_num_capsule,
                                self.num_capsule, self.dim_capsule))
        hat_inputs = K.permute_dimensions(hat_inputs, (0, 2, 1, 3))

        b = K.zeros_like(hat_inputs[:, :, :, 0])
        for i in range(self.routings):
            c = softmax(b, 1)
            o = self.activation(K.batch_dot(c, hat_inputs, [2, 2]))
            if i < self.routings - 1:
                b = K.batch_dot(o, hat_inputs, [2, 3])
                if K.backend() == 'theano':
                    o = K.sum(o, axis=1)

        return o 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_boxes_and_scores(feats, anchors, num_classes, input_shape, image_shape):
    '''Process Conv layer output'''
    box_xy, box_wh, box_confidence, box_class_probs = yolo_head(feats,
        anchors, num_classes, input_shape)
    boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)
    boxes = K.reshape(boxes, [-1, 4])
    box_scores = box_confidence * box_class_probs
    box_scores = K.reshape(box_scores, [-1, num_classes])
    return boxes, box_scores 

def matrix2tensor(x, w):
    ns = x.shape[0]
    nx = int(x.shape[1])
    ny = int(x.shape[2])
    nw = int(x.shape[3])
    assert nw == 1
    assert nx%w == 0
    assert ny%w == 0
    y = K.reshape(x, (-1, nx//w, w, nx//w, w))
    z = K.permute_dimensions(y, (0,1,3,2,4))
    return K.reshape(z, (-1, nx//w, ny//w, w**2)) 

def tensor2matrix(x, w):
    ns = x.shape[0]
    nx = int(x.shape[1])
    ny = int(x.shape[2])
    w2 = int(x.shape[3])
    assert w2 == w**2
    y = K.reshape(x, (-1, nx, ny, w, w))
    z = K.permute_dimensions(y, (0, 1, 3, 2, 4))
    return K.reshape(z, (-1, nx*w, ny*w)) 

def reshape_interpolation(x):
    ns = x.shape[0]
    nx = int(x.shape[1])
    ny = int(x.shape[2])
    n_chan = int(x.shape[3])//4
    y = K.reshape(x, (-1, nx, ny, n_chan, 2, 2))
    z = K.permute_dimensions(y, (0, 1, 4, 2, 5, 3))
    return K.reshape(z, (-1, 2*nx, 2*ny, n_chan))

# === calculate the relative error of the training/test data sets 

def matrix2tensor(x, w):
    ns = x.shape[0]
    nx = int(x.shape[1])
    ny = int(x.shape[2])
    nw = int(x.shape[3])
    assert nw == 1
    assert nx%w == 0
    assert ny%w == 0
    y = K.reshape(x, (-1, nx//w, w, nx//w, w))
    z = K.permute_dimensions(y, (0,1,3,2,4))
    return K.reshape(z, (-1, nx//w, ny//w, w**2)) 

def tensor2matrix(x, w):
    ns = x.shape[0]
    nx = int(x.shape[1])
    ny = int(x.shape[2])
    w2 = int(x.shape[3])
    assert w2 == w**2
    y = K.reshape(x, (-1, nx, ny, w, w))
    z = K.permute_dimensions(y, (0, 1, 3, 2, 4))
    return K.reshape(z, (-1, nx*w, ny*w)) 

def get_constants(self, inputs, training=None):
        constants = []
        if self.implementation != 0 and 0 < self.dropout < 1:
            input_shape = K.int_shape(inputs)
            input_dim = input_shape[-1]
            ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
            ones = K.tile(ones, (1, int(input_dim)))

            def dropped_inputs():
                return K.dropout(ones, self.dropout)

            dp_mask = [K.in_train_phase(dropped_inputs,
                                        ones,
                                        training=training) for _ in range(4)]
            constants.append(dp_mask)
        else:
            constants.append([K.cast_to_floatx(1.) for _ in range(4)])

        if 0 < self.recurrent_dropout < 1:
            ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
            ones = K.tile(ones, (1, self.units))

            def dropped_inputs():
                return K.dropout(ones, self.recurrent_dropout)
            rec_dp_mask = [K.in_train_phase(dropped_inputs,
                                            ones,
                                            training=training) for _ in range(4)]
            constants.append(rec_dp_mask)
        else:
            constants.append([K.cast_to_floatx(1.) for _ in range(4)])

        # append the input as well for use later
        constants.append(inputs)
        return constants 

def _part_to_whole_predictions(self, x):
        """
        Estimate the pose of the whole given the pose of the part.
        :param x: set of poses to transform
        """
        # inputs.shape=[ input_num_capsule, input_num_instance, input_dim_capsule]
        # output.shape=[num_instance*num_capsule, num_parts*input_num_capsule*input_num_instance,dim_capsule]
        # xt.shape = [ input_num_capsule, num_instance, input_num_instance, input_dim_capsule]
        # xpart.shape = [ num_instance, input_num_instance, num_capsule, num_part, dim_x,input_num_capsule]
        # gpose.shape = [ input_num_capsule, num_instance, input_num_instance, dim_geom+1]
        xt = K.tile(K.expand_dims(x,1),[1,self.num_instance,1,1])

        tmp = K.reshape( xt[:,:,:,:1],[self.input_num_capsule,self.num_instance,self.input_num_instance,1,1,1])
        tmp = K.tile(tmp,[1,1,1,self.num_capsule,self.num_part,1])
        ppart=K.permute_dimensions(tmp,[1,2,3,4,5,0])

        gpose = K.concatenate([xt[:,:,:,1:dim_geom+1],K.ones_like(xt[:,:,:,:1])]) # add 1 col to allow x-y translate
        gpart = K.concatenate([K.expand_dims(K.dot(gpose[i],self.W1[i]),-1) for i in range(self.input_num_capsule)])
        apart = K.concatenate([K.expand_dims(K.dot(xt[i,:,:,dim_geom+1:],self.W2[i]),-1) for i in range(self.input_num_capsule)])
        whole=K.concatenate([ppart,gpart,apart],4)
        output=K.permute_dimensions(whole,[0,2,3,5,1,4])
        output=K.reshape(output,[self.num_instance*self.num_capsule,
                                 self.num_part*self.input_num_capsule*self.input_num_instance,self.dim_capsule])
        # output = tf.Print(output, [tf.shape(x)], message='x', summarize=16)
        # output = tf.Print(output, [x[0,18,1:3]], message='x ', summarize=3)
        # output = tf.Print(output, [gpose[0,0,0,:]], message='x gpose ', summarize=5)
        # output = tf.Print(output, [gpose[0,1,0,:]], message='x gpose ', summarize=5)
        # output = tf.Print(output, [gpart[0,0,0,0,0,:]], message='x gpart ', summarize=5)
        # output = tf.Print(output, [gpart[0,1,0,0,0,:]], message='x gpart ', summarize=5)
        return output 

def call(self, inputs, training=None):
        # inputs.shape=[None, input_num_capsule, input_num_instance, input_dim_capsule]
        # inputs_hat.shape=[None,num_instance*num_capsule,num_parts*input_num_capsule*input_num_instance,dim_capsule]

        inputs_hat = K.map_fn(lambda x: self._part_to_whole_predictions(x), elems=inputs)

        # Begin: Routing algorithm ---------------------------------------------------------------------#
        # The prior for coupling coefficient, initialized as zeros.
        # b.shape = [None, self.num_capsule, self.num_parts, self.input_num_capsule].
        b = K.tf.zeros(shape=[K.shape(inputs_hat)[0], self.num_instance*self.num_capsule,
                                 self.num_part*self.input_num_capsule*self.input_num_instance])

        assert self.routings > 0, 'The routings should be > 0.'
        for i in range(self.routings):
            # c.shape=[batch_size, num_instance*num_capsule, input_num_capsule]
            tmpb = K.reshape(b, [-1,self.num_capsule * self.num_instance*self.num_part,
                                        self.input_num_capsule * self.input_num_instance])

            # softmax for all outputs of each input_capsule*input_instance
            tmpc = K.tf.nn.softmax(tmpb, dim=1)
            c=K.reshape(tmpc,[-1,self.num_capsule * self.num_instance,
                                self.num_part*self.input_num_capsule * self.input_num_instance])

            #outputs.shape=[None,num_instance * num_capsule,dim_capsule]
            outputs = self._best_guess(c, inputs_hat)

            if i < self.routings - 1: #
                b += self._agreement(outputs, inputs_hat)

        # End: Routing algorithm -----------------------------------------------------------------------#
        outputs=K.reshape(outputs,[-1,self.num_instance,self.num_capsule,self.dim_capsule])
        outputs=K.permute_dimensions(outputs,[0,2,1,3])
        return outputs 

def _generate_recurrent_dropout_mask(self, inputs, training=None):
        if 0 < self.recurrent_dropout < 1:
            ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
            ones = K.tile(ones, (1, self.units))

            def dropped_inputs():
                return K.dropout(ones, self.dropout)

            self._recurrent_dropout_mask = [K.in_train_phase(
                dropped_inputs,
                ones,
                training=training)
                for _ in range(2)]
        else:
            self._recurrent_dropout_mask = None 

def call(self, x, mask=None):
        input_shape = K.int_shape(x)

        features_dim = self.features_dim
        # step_dim = self.step_dim
        step_dim = input_shape[1]

        eij = K.reshape(K.dot(K.reshape(x, (-1, features_dim)), K.reshape(self.W, (features_dim, 1))), (-1, step_dim))

        if self.bias:
            eij += self.b[:input_shape[1]]

        eij = K.tanh(eij)

        a = K.exp(eij)

        # apply mask after the exp. will be re-normalized next
        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            a *= K.cast(mask, K.floatx())

        # in some cases especially in the early stages of training the sum may be almost zero
        # and this results in NaN's. A workaround is to add a very small positive number ε to the sum.
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        a = K.expand_dims(a)
        weighted_input = x * a
    	# print weigthted_input.shape
        return K.sum(weighted_input, axis=1) 

def call(self, inputs, mask=None):
        en = inputs[0]
        de = inputs[1]
        de_shape = K.int_shape(de)
        step_dim = de_shape[1]

        hid_en = K.dot(en, self.W_en1)
        hid_de = K.dot(de, self.W_en2)
        if self.bias:
            hid_en += self.b_en1
            hid_de += self.b_en2
        hid = K.tanh(K.expand_dims(hid_en, axis=1) + hid_de)
        eij = K.reshape(K.dot(hid, K.reshape(self.W_de, (self.hid_size, 1))), (-1, step_dim))
        if self.bias:
            eij += self.b_de[:step_dim]

        a = K.exp(eij - K.max(eij, axis=-1, keepdims=True))

        # apply mask after the exp. will be re-normalized next
        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            a *= K.cast(mask[1], K.floatx())

        # in some cases especially in the early stages of training the sum may be almost zero
        # and this results in NaN's. A workaround is to add a very small positive number ε to the sum.
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        a = K.expand_dims(a)
        weighted_input = de * a
        return K.sum(weighted_input, axis=1) 

def load_data():
    # load training data
    directory = os.path.join(Utils.data_path, 'Train_fcn_all_' + str(Utils.test_frac) + '.h5')
    images, train_labels = read_file(directory, 'train')
    # load test data
    directory = os.path.join(Utils.data_path, 'Test_fcn_all_one_' + str(Utils.test_frac) + '.h5')
    _, test_labels = read_file(directory, 'test')


#    images = np.array(images[:144, :144])
#    train_labels = np.array(train_labels[:, :144, :144])
#    test_labels = np.array(test_labels[:, :144, :144])


    train_labels = np.reshape(train_labels, (train_labels.shape[0], train_labels.shape[1],  train_labels.shape[2], 1))
    test_labels = np.reshape(test_labels, (test_labels.shape[0], test_labels.shape[1],  test_labels.shape[2], 1))

    if model_name == 'fcn_3d':
        images = np.reshape(images, (images.shape[0], images.shape[1], images.shape[2], 1))
        input_size = (images.shape[1], images.shape[2], images.shape[2], 1)
#    test_images = np.reshape(test_images, (1, test_images.shape[0], test_images.shape[1], test_images.shape[2], 1))
    else:
        input_size = (images.shape[0], images.shape[1], images.shape[2])

    return images, train_labels, test_labels, input_size


# visualizing losses and accuracy 

def sparse_accuracy(y_true, y_pred):
    classes = K.int_shape(y_pred)[-1]
    y_pred = K.reshape(y_pred, (-1, classes))

    y_true = K.one_hot(tf.to_int32(K.flatten(y_true)),
                       classes + 1)
    unpacked = tf.unstack(y_true, axis=-1)
    legal_labels = ~tf.cast(unpacked[0], tf.bool)
    y_true = tf.stack(unpacked[1:], axis=-1)

    return K.sum(tf.to_float(legal_labels & K.equal(K.argmax(y_true, axis=-1), K.argmax(y_pred, axis=-1)))) / K.sum(tf.to_float(legal_labels)) 

def generate_batches(x, y, num_batch):
    while 1:
        np.random.shuffle(y)
        for i in range(y.shape[0] // num_batch):
            batch = slice(num_batch * i, num_batch * (i + 1))
            temp = np.reshape(x, (x.shape[0], x.shape[1], x.shape[2], 1))
            temp = np.tile(temp, num_batch)
            temp = np.transpose(temp, (3, 0, 1, 2))
            yield temp, y[batch]


# Define different models

# 3D-FCN model 

def fcn_3d(input_shape):
    inputs = Input(input_shape)

    conv1 = Conv3D(8, kernel_size=(3, 3, 20), strides=(1, 1, 10), activation='relu')(inputs)
    pool1 = MaxPooling3D(pool_size=(3, 3, 1), strides=(3, 3, 1))(conv1)

    conv2 = Conv3D(32, kernel_size=(6, 6, 10), strides=(1, 1, 2), activation='relu')(pool1)
    pool2 = MaxPooling3D(pool_size=(3, 3, 1), strides=(3, 3, 1))(conv2)

    conv3 = Conv3D(128, kernel_size=(3, 3, 5), strides=(1, 1, 1), activation='relu')(pool2)
    pool3 = MaxPooling3D(pool_size=(3, 3, 1), strides=(3, 3, 1))(conv3)

    reshape = Reshape((4, 4, 128))(pool3)

    up1 = UpSampling2D(size=(3, 3))(reshape)
    deconv1 = Conv2DTranspose(32, 3, activation='relu')(up1)

    up2 = UpSampling2D(size=(3, 3))(deconv1)
    deconv2 = Conv2DTranspose(8, 6, activation='relu')(up2)

    up3 = UpSampling2D(size=(3, 3))(deconv2)
    deconv3 = Conv2DTranspose(nb_classes, 3)(up3)
    deconv4 = Conv2DTranspose(nb_classes, 3)(deconv3)

    model = Model(inputs=inputs, outputs=deconv4)
    adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)

    model.compile(loss=softmax_sparse_crossentropy_ignoring_first_label,
                  optimizer=adam,
                  metrics=[sparse_accuracy])
    return model


# 2D-FCN model 

def fcn_3d(input_shape):
    inputs = Input(input_shape)

    conv1 = Conv3D(8, kernel_size=(3, 3, 20), strides=(1, 1, 10), activation='relu')(inputs)
    pool1 = MaxPooling3D(pool_size=(3, 3, 3), strides=(1, 1, 1), padding='same')(conv1)

    conv2 = Conv3D(32, kernel_size=(6, 6, 10), strides=(1, 1, 2), activation='relu')(pool1)
    pool2 = MaxPooling3D(pool_size=(3, 3, 3), strides=(1, 1, 1), padding='same')(conv2)

    conv3 = Conv3D(128, kernel_size=(3, 3, 6), activation='relu')(pool2)

    reshape = Reshape((2, 2, 128))(conv3)

    deconv1 = Conv2DTranspose(32, 3, activation='relu')(reshape)

#    up1 = UpSampling2D(size=(3, 3))(deconv1)
    deconv2 = Conv2DTranspose(8, 6, activation='relu')(deconv1)

#    up2 = UpSampling2D(size=(3, 3))(deconv2)
    deconv3 = Conv2DTranspose(nb_classes, 3)(deconv2)

    model = Model(input=inputs, output=deconv3)
    adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)

    model.compile(loss=softmax_sparse_crossentropy_ignoring_first_label,
                  optimizer=adam,
                  metrics=[sparse_accuracy])
    return model


# 2D-FCN model 

def call(self, x, mask=None):

        assert self.built, 'Layer must be built before being called'
        input_shape = K.int_shape(x)

        reduction_axes = list(range(len(input_shape)))
        del reduction_axes[self.axis]
        broadcast_shape = [1] * len(input_shape)
        broadcast_shape[self.axis] = input_shape[self.axis]

        if sorted(reduction_axes) == range(K.ndim(x))[:-1]:
            x_normed = K.batch_normalization(
                x, self.running_mean, self.running_std,
                self.beta, self.gamma,
                epsilon=self.epsilon)
        else:
            # need broadcasting
            broadcast_running_mean = K.reshape(self.running_mean, broadcast_shape)
            broadcast_running_std = K.reshape(self.running_std, broadcast_shape)
            broadcast_beta = K.reshape(self.beta, broadcast_shape)
            broadcast_gamma = K.reshape(self.gamma, broadcast_shape)
            x_normed = K.batch_normalization(
                x, broadcast_running_mean, broadcast_running_std,
                broadcast_beta, broadcast_gamma,
                epsilon=self.epsilon)

        return x_normed 

def yolo_head(feats, anchors, num_classes, input_shape):
    """
    Convert final layer features to bounding box parameters.
    """
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3]  # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
                    [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
                    [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust predictions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    return box_xy, box_wh, box_confidence, box_class_probs 

def _get_anchors(self):
        with open(anchors_path) as f:
            anchors = f.readline()
        anchors = [float(x) for x in anchors.split(',')]
        return np.array(anchors).reshape(-1, 2) 

def overlaps_graph(boxes1, boxes2):
    """Computes IoU overlaps between two sets of boxes.
    boxes1, boxes2: [N, (y1, x1, y2, x2)].
    """
    # 1. Tile boxes2 and repeat boxes1. This allows us to compare
    # every boxes1 against every boxes2 without loops.
    # TF doesn't have an equivalent to np.repeat() so simulate it
    # using tf.tile() and tf.reshape.
    b1 = tf.reshape(tf.tile(tf.expand_dims(boxes1, 1),
                            [1, 1, tf.shape(boxes2)[0]]), [-1, 4])
    b2 = tf.tile(boxes2, [tf.shape(boxes1)[0], 1])
    # 2. Compute intersections
    b1_y1, b1_x1, b1_y2, b1_x2 = tf.split(b1, 4, axis=1)
    b2_y1, b2_x1, b2_y2, b2_x2 = tf.split(b2, 4, axis=1)
    y1 = tf.maximum(b1_y1, b2_y1)
    x1 = tf.maximum(b1_x1, b2_x1)
    y2 = tf.minimum(b1_y2, b2_y2)
    x2 = tf.minimum(b1_x2, b2_x2)
    intersection = tf.maximum(x2 - x1, 0) * tf.maximum(y2 - y1, 0)
    # 3. Compute unions
    b1_area = (b1_y2 - b1_y1) * (b1_x2 - b1_x1)
    b2_area = (b2_y2 - b2_y1) * (b2_x2 - b2_x1)
    union = b1_area + b2_area - intersection
    # 4. Compute IoU and reshape to [boxes1, boxes2]
    iou = intersection / union
    overlaps = tf.reshape(iou, [tf.shape(boxes1)[0], tf.shape(boxes2)[0]])
    return overlaps 

def call(self, inputs):
        rois = inputs[0]
        mrcnn_class = inputs[1]
        mrcnn_bbox = inputs[2]
        image_meta = inputs[3]

        # Get windows of images in normalized coordinates. Windows are the area
        # in the image that excludes the padding.
        # Use the shape of the first image in the batch to normalize the window
        # because we know that all images get resized to the same size.
        m = parse_image_meta_graph(image_meta)
        image_shape = m['image_shape'][0]
        window = norm_boxes_graph(m['window'], image_shape[:2])
        
        # Run detection refinement graph on each item in the batch
        detections_batch = utils.batch_slice(
            [rois, mrcnn_class, mrcnn_bbox, window],
            lambda x, y, w, z: refine_detections_graph(x, y, w, z, self.config),
            self.config.IMAGES_PER_GPU)

        # Reshape output
        # [batch, num_detections, (y1, x1, y2, x2, class_score)] in
        # normalized coordinates
        return tf.reshape(
            detections_batch,
            [self.config.BATCH_SIZE, self.config.DETECTION_MAX_INSTANCES, 6]) 

def mrcnn_bbox_loss_graph(target_bbox, target_class_ids, pred_bbox):
    """Loss for Mask R-CNN bounding box refinement.

    target_bbox: [batch, num_rois, (dy, dx, log(dh), log(dw))]
    target_class_ids: [batch, num_rois]. Integer class IDs.
    pred_bbox: [batch, num_rois, num_classes, (dy, dx, log(dh), log(dw))]
    """
    # Reshape to merge batch and roi dimensions for simplicity.
    target_class_ids = K.reshape(target_class_ids, (-1,))
    target_bbox = K.reshape(target_bbox, (-1, 4))
    pred_bbox = K.reshape(pred_bbox, (-1, K.int_shape(pred_bbox)[2], 4))

    # Only positive ROIs contribute to the loss. And only
    # the right class_id of each ROI. Get their indices.
    positive_roi_ix = tf.where(target_class_ids > 0)[:, 0]
    positive_roi_class_ids = tf.cast(
        tf.gather(target_class_ids, positive_roi_ix), tf.int64)
    indices = tf.stack([positive_roi_ix, positive_roi_class_ids], axis=1)

    # Gather the deltas (predicted and true) that contribute to loss
    target_bbox = tf.gather(target_bbox, positive_roi_ix)
    pred_bbox = tf.gather_nd(pred_bbox, indices)

    # Smooth-L1 Loss
    loss = K.switch(tf.size(target_bbox) > 0,
                    smooth_l1_loss(y_true=target_bbox, y_pred=pred_bbox),
                    tf.constant(0.0))
    loss = K.mean(loss)
    return loss 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_boxes_and_scores(feats, anchors, num_classes, input_shape, image_shape):
    '''Process Conv layer output'''
    box_xy, box_wh, box_confidence, box_class_probs = yolo_head(feats,
        anchors, num_classes, input_shape)
    boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)
    boxes = K.reshape(boxes, [-1, 4])
    box_scores = box_confidence * box_class_probs
    box_scores = K.reshape(box_scores, [-1, num_classes])
    return boxes, box_scores 

def relative_logits_1d(self, q, rel_k, H, W, transpose_mask):
        rel_logits = tf.einsum('bhxyd,md->bhxym', q, rel_k)
        rel_logits = K.reshape(rel_logits, [-1, self.num_heads * H, W, 2 * W - 1])
        rel_logits = self.rel_to_abs(rel_logits)
        rel_logits = K.reshape(rel_logits, [-1, self.num_heads, H, W, W])
        rel_logits = K.expand_dims(rel_logits, axis=3)
        rel_logits = K.tile(rel_logits, [1, 1, 1, H, 1, 1])
        rel_logits = K.permute_dimensions(rel_logits, transpose_mask)
        rel_logits = K.reshape(rel_logits, [-1, self.num_heads, H * W, H * W])
        return rel_logits 

def rel_to_abs(self, x):
        shape = K.shape(x)
        shape = [shape[i] for i in range(3)]
        B, Nh, L, = shape
        col_pad = K.zeros(K.stack([B, Nh, L, 1]))
        x = K.concatenate([x, col_pad], axis=3)
        flat_x = K.reshape(x, [B, Nh, L * 2 * L])
        flat_pad = K.zeros(K.stack([B, Nh, L - 1]))
        flat_x_padded = K.concatenate([flat_x, flat_pad], axis=2)
        final_x = K.reshape(flat_x_padded, [B, Nh, L + 1, 2 * L - 1])
        final_x = final_x[:, :, :L, L - 1:]
        return final_x 

def combine_heads_2d(self, inputs):
        # [batch, num_heads, height, width, depth_v // num_heads]
        transposed = K.permute_dimensions(inputs, [0, 2, 3, 1, 4])
        # [batch, height, width, num_heads, depth_v // num_heads]
        shape = K.shape(transposed)
        shape = [shape[i] for i in range(5)]

        a, b = shape[-2:]
        ret_shape = K.stack(shape[:-2] + [a * b])
        # [batch, height, width, depth_v]
        return K.reshape(transposed, ret_shape) 

def yolo_eval(yolo_outputs,
              image_shape,
              max_boxes=10,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input batch and return filtered boxes."""
    box_xy, box_wh, box_confidence, box_class_probs = yolo_outputs
    boxes = yolo_boxes_to_corners(box_xy, box_wh)
    boxes, scores, classes = yolo_filter_boxes(
        boxes, box_confidence, box_class_probs, threshold=score_threshold)

    # Scale boxes back to original image shape.
    height = image_shape[0]
    width = image_shape[1]
    image_dims = K.stack([height, width, height, width])
    image_dims = K.reshape(image_dims, [1, 4])
    boxes = boxes * image_dims

    # TODO: Something must be done about this ugly hack!
    max_boxes_tensor = K.variable(max_boxes, dtype='int32')
    K.get_session().run(tf.variables_initializer([max_boxes_tensor]))
    nms_index = tf.image.non_max_suppression(
        boxes, scores, max_boxes_tensor, iou_threshold=iou_threshold)
    boxes = K.gather(boxes, nms_index)
    scores = K.gather(scores, nms_index)
    classes = K.gather(classes, nms_index)
    return boxes, scores, classes 

def call(self, x, **kwargs):
        assert isinstance(x, list), 'SliceLayer input is not a list'
        return x[0] * K.reshape(x[1], (-1, 1)) 

def compute_kernel(x, y, kernel='rbf', **kwargs):
    """
        Computes RBF kernel between x and y.
        # Parameters
            x: Tensor
                Tensor with shape [batch_size, z_dim]
            y: Tensor
                Tensor with shape [batch_size, z_dim]
        # Returns
            returns the computed RBF kernel between x and y
    """
    scales = kwargs.get("scales", [])
    if kernel == "rbf":
        x_size = K.shape(x)[0]
        y_size = K.shape(y)[0]
        dim = K.shape(x)[1]
        tiled_x = K.tile(K.reshape(x, K.stack([x_size, 1, dim])), K.stack([1, y_size, 1]))
        tiled_y = K.tile(K.reshape(y, K.stack([1, y_size, dim])), K.stack([x_size, 1, 1]))
        return K.exp(-K.mean(K.square(tiled_x - tiled_y), axis=2) / K.cast(dim, tf.float32))
    elif kernel == 'raphy':
        scales = K.variable(value=np.asarray(scales))
        squared_dist = K.expand_dims(squared_distance(x, y), 0)
        scales = K.expand_dims(K.expand_dims(scales, -1), -1)
        weights = K.eval(K.shape(scales)[0])
        weights = K.variable(value=np.asarray(weights))
        weights = K.expand_dims(K.expand_dims(weights, -1), -1)
        return K.sum(weights * K.exp(-squared_dist / (K.pow(scales, 2))), 0)
    elif kernel == "multi-scale-rbf":
        sigmas = [1e-6, 1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1, 5, 10, 15, 20, 25, 30, 35, 100, 1e3, 1e4, 1e5, 1e6]

        beta = 1. / (2. * (K.expand_dims(sigmas, 1)))
        distances = squared_distance(x, y)
        s = K.dot(beta, K.reshape(distances, (1, -1)))

        return K.reshape(tf.reduce_sum(tf.exp(-s), 0), K.shape(distances)) / len(sigmas) 

def call(self, inputs, **kwargs):
        def sqrt(x):
            return K.sqrt(K.maximum(x, K.epsilon()))

        # Both components and input given
        if isinstance(inputs, list) and len(inputs) > 1:
            signals, kernel = inputs
        else:
            signals = inputs
            kernel = self.components.astype(K.floatx())

        # move component_number to channel dimension
        kernel = K.permute_dimensions(kernel, (1, 2, 3, 0))
        # normalize kernel
        normed_kernel = kernel / sqrt(K.sum(K.square(kernel),
                                            (0, 1, 2),
                                            keepdims=True))

        # get norm of signals
        signals_norm = sqrt(K.conv2d(K.square(signals),
                                     np.ones(K.int_shape(kernel)[:3] + (1,),
                                             dtype=K.floatx()),
                                     strides=self.strides,
                                     padding=self.padding,
                                     data_format='channels_last',
                                     dilation_rate=self.dilation_rate))

        diss = K.conv2d(signals,
                        normed_kernel,
                        strides=self.strides,
                        padding=self.padding,
                        data_format='channels_last',
                        dilation_rate=self.dilation_rate) / signals_norm

        if self.n_replicas != 1:
            shape = K.int_shape(diss)
            diss = K.reshape(diss, (-1, shape[1], shape[2],
                                    shape[3] // self.n_replicas,
                                    self.n_replicas))

        return self.activation(diss) 

def call(self, inputs, **kwargs):
        # Both components and input given
        if isinstance(inputs, list) and len(inputs) > 1:
            signals, kernel = inputs
        else:
            signals = inputs
            kernel = self.components.astype(K.floatx())

        # move component_number to channel dimension
        kernel = K.permute_dimensions(kernel, (1, 2, 3, 0))
        # normalize kernel
        kernel_norm = K.sum(K.square(kernel), (0, 1, 2), keepdims=True)

        # get norm of signals
        signals_norm = K.conv2d(K.square(signals),
                                K.ones_like(kernel),
                                strides=self.strides,
                                padding=self.padding,
                                data_format='channels_last',
                                dilation_rate=self.dilation_rate)

        diss = kernel_norm \
               - 2 * K.conv2d(signals,
                              kernel,
                              strides=self.strides,
                              padding=self.padding,
                              data_format='channels_last',
                              dilation_rate=self.dilation_rate) \
               + signals_norm

        if self.n_replicas != 1:
            shape = K.int_shape(diss)
            diss = K.reshape(diss, (-1, shape[1], shape[2],
                                    shape[3] // self.n_replicas,
                                    self.n_replicas))

        return self.activation(diss) 

def call(self, inputs):

        input_shape = K.int_shape(inputs)
        if len(input_shape) != 4:
            raise ValueError('Inputs should have rank ' +
                             str(4) +
                             '; Received input shape:', str(input_shape))

        if self.data_format == 'channels_first':
            batch_size, c, h, w = input_shape
            if batch_size is None:
                batch_size = -1
            rh, rw = self.size
            oh, ow = h * rh, w * rw
            oc = c // (rh * rw)

            out = K.reshape(inputs, (batch_size, rh, rw, oc, h, w))
            out = K.permute_dimensions(out, (0, 3, 4, 1, 5, 2))
            out = K.reshape(out, (batch_size, oc, oh, ow))
            return out

        elif self.data_format == 'channels_last':
            batch_size, h, w, c = input_shape
            if batch_size is None:
                batch_size = -1
            rh, rw = self.size
            oh, ow = h * rh, w * rw
            oc = c // (rh * rw)

            out = K.reshape(inputs, (batch_size, h, w, rh, rw, oc))
            out = K.permute_dimensions(out, (0, 1, 3, 2, 4, 5))
            out = K.reshape(out, (batch_size, oh, ow, oc))
            return out 

def call(self, inputs):

        input_shape = K.int_shape(inputs)
        if len(input_shape) != 4:
            raise ValueError('Inputs should have rank ' +
                             str(4) +
                             '; Received input shape:', str(input_shape))

        if self.data_format == 'channels_first':
            batch_size, c, h, w = input_shape
            if batch_size is None:
                batch_size = -1
            rh, rw = self.size
            oh, ow = h * rh, w * rw
            oc = c // (rh * rw)

            out = K.reshape(inputs, (batch_size, rh, rw, oc, h, w))
            out = K.permute_dimensions(out, (0, 3, 4, 1, 5, 2))
            out = K.reshape(out, (batch_size, oc, oh, ow))
            return out

        elif self.data_format == 'channels_last':
            batch_size, h, w, c = input_shape
            if batch_size is None:
                batch_size = -1
            rh, rw = self.size
            oh, ow = h * rh, w * rw
            oc = c // (rh * rw)

            out = K.reshape(inputs, (batch_size, h, w, rh, rw, oc))
            out = K.permute_dimensions(out, (0, 1, 3, 2, 4, 5))
            out = K.reshape(out, (batch_size, oh, ow, oc))
            return out 

def repeat(x):
    return K.reshape(K.repeat(K.batch_flatten(x), nb_timestep), (b_s, nb_timestep, 512, shape_r_gt, shape_c_gt)) 

def call(self, x, mask=None):
        input_shape = self.input_spec[0].shape
        broadcast_shape = [1] * len(input_shape)
        broadcast_shape[self.axis] = input_shape[self.axis]

        out = K.reshape(self.gamma, broadcast_shape) * x + K.reshape(self.beta, broadcast_shape)
        return out 

def yolo_head(feats, anchors, num_classes, input_shape):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    box_xy = K.sigmoid(feats[..., :2])
    box_wh = K.exp(feats[..., 2:4])
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (box_xy + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = box_wh * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))

    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_boxes_and_scores(feats, anchors, num_classes, input_shape, image_shape):
    '''Process Conv layer output'''
    box_xy, box_wh, box_confidence, box_class_probs = yolo_head(feats,
        anchors, num_classes, input_shape)
    boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)
    boxes = K.reshape(boxes, [-1, 4])
    box_scores = box_confidence * box_class_probs
    box_scores = K.reshape(box_scores, [-1, num_classes])
    return boxes, box_scores 

def call(self, inputs):
        if not isinstance(inputs, list):
            raise ValueError('Similarity layer expects a list '
                             'of tensors as inputs.')
        if len(inputs) != 2:
            raise ValueError('Similarity layer expects two tensors as '
                             'input, {} were given.'.format(len(inputs)))

        x = inputs[0]
        y = inputs[1]

        # Each line in X should have the form: dataX,       1s,      dataX.
        # Each line in Y should have the form: 1s,       dataY,      dataY.
        #
        #                               =>     dataX,    dataY,  dataX * dataY
        #
        x = K.concatenate([x, K.ones(K.shape(x)), x], axis=-1)
        y = K.concatenate([K.ones(K.shape(y)), y, y], axis=-1)

        # Pair each lines and take elementwise product (without summation).
        # x = K.reshape(x, (-1, x.shape[1], 1, x.shape[2]))
        # y = K.reshape(y, (-1, 1, y.shape[1], y.shape[2]))
        x = K.expand_dims(x, axis=2)
        y = K.expand_dims(y, axis=1)
        rez = x * y

        # Apply dot product with a vector.
        rez = rez * self.WS

        # return K.ones((1, 93)) * self.WS
        return K.sum(rez, axis=-1) 

def call(self, inputs):
        if not isinstance(inputs, list):
            raise ValueError('Similarity layer expects a list '
                             'of tensors as inputs.')
        if len(inputs) != 2:
            raise ValueError('Similarity layer expects two tensors as '
                             'input, {} were given.'.format(len(inputs)))

        x = inputs[0]
        y = inputs[1]

        # Each line in X should have the form: dataX,       1s,      dataX.
        # Each line in Y should have the form: 1s,       dataY,      dataY.
        #
        #                               =>     dataX,    dataY,  dataX * dataY
        #
        x = K.concatenate([x, K.ones(K.shape(x)), x], axis=-1)
        y = K.concatenate([K.ones(K.shape(y)), y, y], axis=-1)

        # Pair each lines and take elementwise product (without summation).
        # x = K.reshape(x, (-1, x.shape[1], 1, x.shape[2]))
        # y = K.reshape(y, (-1, 1, y.shape[1], y.shape[2]))
        x = K.expand_dims(x, axis=2)
        y = K.expand_dims(y, axis=1)
        rez = x * y

        # Apply dot product with a vector.
        rez = rez * self.WS

        # return K.ones((1, 93)) * self.WS
        return K.sum(rez, axis=-1) 

def call(self, inputs):
        if not isinstance(inputs, list):
            raise ValueError('Similarity layer expects a list '
                             'of tensors as inputs.')
        if len(inputs) != 2:
            raise ValueError('Similarity layer expects two tensors as '
                             'input, {} were given.'.format(len(inputs)))

        x = inputs[0]
        y = inputs[1]

        # Each line in X should have the form: dataX,       1s,      dataX.
        # Each line in Y should have the form: 1s,       dataY,      dataY.
        #
        #                               =>     dataX,    dataY,  dataX * dataY
        #
        x = K.concatenate([x, K.ones(K.shape(x)), x], axis=-1)
        y = K.concatenate([K.ones(K.shape(y)), y, y], axis=-1)

        # Pair each lines and take elementwise product (without summation).
        # x = K.reshape(x, (-1, x.shape[1], 1, x.shape[2]))
        # y = K.reshape(y, (-1, 1, y.shape[1], y.shape[2]))
        x = K.expand_dims(x, axis=2)
        y = K.expand_dims(y, axis=1)
        rez = x * y

        # Apply dot product with a vector.
        rez = rez * self.WS

        # return K.ones((1, 93)) * self.WS
        return K.sum(rez, axis=-1) 

def call(self, inputs):
        if not isinstance(inputs, list):
            raise ValueError('Similarity layer expects a list '
                             'of tensors as inputs.')
        if len(inputs) != 2:
            raise ValueError('Similarity layer expects two tensors as '
                             'input, {} were given.'.format(len(inputs)))

        x = inputs[0]
        y = inputs[1]

        # Each line in X should have the form: dataX,       1s,      dataX.
        # Each line in Y should have the form: 1s,       dataY,      dataY.
        #
        #                               =>     dataX,    dataY,  dataX * dataY
        #
        x = K.concatenate([x, K.ones(K.shape(x)), x], axis=-1)
        y = K.concatenate([K.ones(K.shape(y)), y, y], axis=-1)

        # Pair each lines and take elementwise product (without summation).
        # x = K.reshape(x, (-1, x.shape[1], 1, x.shape[2]))
        # y = K.reshape(y, (-1, 1, y.shape[1], y.shape[2]))
        x = K.expand_dims(x, axis=2)
        y = K.expand_dims(y, axis=1)
        rez = x * y

        # Apply dot product with a vector.
        rez = rez * self.WS

        # return K.ones((1, 93)) * self.WS
        return K.sum(rez, axis=-1) 

def path_energy0(y, x, U, mask=None):
    '''Path energy without boundary potential handling.'''
    n_classes = K.shape(x)[2]
    y_one_hot = K.one_hot(y, n_classes)

    # Tag path energy
    energy = K.sum(x * y_one_hot, 2)
    energy = K.sum(energy, 1)

    # Transition energy
    y_t = y[:, :-1]
    y_tp1 = y[:, 1:]
    U_flat = K.reshape(U, [-1])
    # Convert 2-dim indices (y_t, y_tp1) of U to 1-dim indices of U_flat:
    flat_indices = y_t * n_classes + y_tp1
    U_y_t_tp1 = K.gather(U_flat, flat_indices)

    if mask is not None:
        mask = K.cast(mask, K.floatx())
        y_t_mask = mask[:, :-1]
        y_tp1_mask = mask[:, 1:]
        U_y_t_tp1 *= y_t_mask * y_tp1_mask

    energy += K.sum(U_y_t_tp1, axis=1)

    return energy 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    #print(K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_boxes_and_scores(feats, anchors, num_classes, input_shape, image_shape):
    '''Process Conv layer output'''
    box_xy, box_wh, box_confidence, box_class_probs = yolo_head(feats,
        anchors, num_classes, input_shape)
    boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)
    boxes = K.reshape(boxes, [-1, 4])
    box_scores = box_confidence * box_class_probs
    #print(box_scores.shape)
    box_scores = K.reshape(box_scores, [-1, num_classes])
    return boxes, box_scores 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_boxes_and_scores(feats, anchors, num_classes, input_shape, image_shape):
    '''Process Conv layer output'''
    box_xy, box_wh, box_confidence, box_class_probs = yolo_head(feats,
        anchors, num_classes, input_shape)
    boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)
    boxes = K.reshape(boxes, [-1, 4])
    box_scores = box_confidence * box_class_probs
    box_scores = K.reshape(box_scores, [-1, num_classes])
    return boxes, box_scores 

def call(self, x, mask=None):
        features_dim = self.features_dim
        # 这里应该是 step_dim是我们指定的参数，它等于input_shape[1],也就是rnn的timesteps
        step_dim = self.step_dim

        # 输入和参数分别reshape再点乘后，tensor.shape变成了(batch_size*timesteps, 1),之后每个batch要分开进行归一化
		# 所以应该有 eij = K.reshape(..., (-1, timesteps))
        eij = K.reshape(K.dot(K.reshape(x, (-1, features_dim)),
                        K.reshape(self.W, (features_dim, 1))), (-1, step_dim))

        if self.bias:
            eij += self.b
        # RNN一般默认激活函数为tanh, 对attention来说激活函数差别不打，因为要做softmax
        eij = K.tanh(eij)

        a = K.exp(eij)

        if mask is not None:
        # 如果前面的层有mask，那么后面这些被mask掉的timestep肯定是不能参与计算输出的，也就是将他们的attention权重设为0
            a *= K.cast(mask, K.floatx())
        # cast是做类型转换，keras计算时会检查类型，可能是因为用gpu的原因
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        # a = K.expand_dims(a, axis=-1) , axis默认为-1， 表示在最后扩充一个维度。
		# 比如shape = (3,)变成 (3, 1)
        a = K.expand_dims(a)
        # 此时a.shape = (batch_size, timesteps, 1), x.shape = (batch_size, timesteps, units)
        weighted_input = x * a

        # weighted_input的shape为 (batch_size, timesteps, units), 每个timestep的输出向量已经乘上了该timestep的权重
		# weighted_input在axis=1上取和，返回值的shape为 (batch_size, 1, units)
        return K.sum(weighted_input, axis=1) 

def call(self, x, mask=None):
        features_dim = self.features_dim
        # 这里应该是 step_dim是我们指定的参数，它等于input_shape[1],也就是rnn的timesteps
        step_dim = self.step_dim

        # 输入和参数分别reshape再点乘后，tensor.shape变成了(batch_size*timesteps, 1),之后每个batch要分开进行归一化
		# 所以应该有 eij = K.reshape(..., (-1, timesteps))
        eij = K.reshape(K.dot(K.reshape(x, (-1, features_dim)),
                        K.reshape(self.W, (features_dim, 1))), (-1, step_dim))

        if self.bias:
            eij += self.b
        # RNN一般默认激活函数为tanh, 对attention来说激活函数差别不打，因为要做softmax
        eij = K.tanh(eij)

        a = K.exp(eij)

        if mask is not None:
        # 如果前面的层有mask，那么后面这些被mask掉的timestep肯定是不能参与计算输出的，也就是将他们的attention权重设为0
            a *= K.cast(mask, K.floatx())
        # cast是做类型转换，keras计算时会检查类型，可能是因为用gpu的原因
        a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())

        # a = K.expand_dims(a, axis=-1) , axis默认为-1， 表示在最后扩充一个维度。
		# 比如shape = (3,)变成 (3, 1)
        a = K.expand_dims(a)
        # 此时a.shape = (batch_size, timesteps, 1), x.shape = (batch_size, timesteps, units)
        weighted_input = x * a

        # weighted_input的shape为 (batch_size, timesteps, units), 每个timestep的输出向量已经乘上了该timestep的权重
		# weighted_input在axis=1上取和，返回值的shape为 (batch_size, 1, units)
        return K.sum(weighted_input, axis=1) 

def overlaps_graph(boxes1, boxes2):
    """Computes IoU overlaps between two sets of boxes.
    boxes1, boxes2: [N, (y1, x1, y2, x2)].
    """
    # 1. Tile boxes2 and repeate boxes1. This allows us to compare
    # every boxes1 against every boxes2 without loops.
    # TF doesn't have an equivalent to np.repeate() so simulate it
    # using tf.tile() and tf.reshape.
    b1 = tf.reshape(tf.tile(tf.expand_dims(boxes1, 1),
                            [1, 1, tf.shape(boxes2)[0]]), [-1, 4])
    b2 = tf.tile(boxes2, [tf.shape(boxes1)[0], 1])
    # 2. Compute intersections
    b1_y1, b1_x1, b1_y2, b1_x2 = tf.split(b1, 4, axis=1)
    b2_y1, b2_x1, b2_y2, b2_x2 = tf.split(b2, 4, axis=1)
    y1 = tf.maximum(b1_y1, b2_y1)
    x1 = tf.maximum(b1_x1, b2_x1)
    y2 = tf.minimum(b1_y2, b2_y2)
    x2 = tf.minimum(b1_x2, b2_x2)
    intersection = tf.maximum(x2 - x1, 0) * tf.maximum(y2 - y1, 0)
    # 3. Compute unions
    b1_area = (b1_y2 - b1_y1) * (b1_x2 - b1_x1)
    b2_area = (b2_y2 - b2_y1) * (b2_x2 - b2_x1)
    union = b1_area + b2_area - intersection
    # 4. Compute IoU and reshape to [boxes1, boxes2]
    iou = intersection / union
    overlaps = tf.reshape(iou, [tf.shape(boxes1)[0], tf.shape(boxes2)[0]])
    return overlaps 

def mrcnn_bbox_loss_graph(target_bbox, target_class_ids, pred_bbox):
    """Loss for Mask R-CNN bounding box refinement.

    target_bbox: [batch, num_rois, (dy, dx, log(dh), log(dw))]
    target_class_ids: [batch, num_rois]. Integer class IDs.
    pred_bbox: [batch, num_rois, num_classes, (dy, dx, log(dh), log(dw))]
    """
    # Reshape to merge batch and roi dimensions for simplicity.
    target_class_ids = K.reshape(target_class_ids, (-1,))
    target_bbox = K.reshape(target_bbox, (-1, 4))
    pred_bbox = K.reshape(pred_bbox, (-1, K.int_shape(pred_bbox)[2], 4))

    # Only positive ROIs contribute to the loss. And only
    # the right class_id of each ROI. Get their indicies.
    positive_roi_ix = tf.where(target_class_ids > 0)[:, 0]
    positive_roi_class_ids = tf.cast(
        tf.gather(target_class_ids, positive_roi_ix), tf.int64)
    indices = tf.stack([positive_roi_ix, positive_roi_class_ids], axis=1)

    # Gather the deltas (predicted and true) that contribute to loss
    target_bbox = tf.gather(target_bbox, positive_roi_ix)
    pred_bbox = tf.gather_nd(pred_bbox, indices)

    # Smooth-L1 Loss
    loss = K.switch(tf.size(target_bbox) > 0,
                    smooth_l1_loss(y_true=target_bbox, y_pred=pred_bbox),
                    tf.constant(0.0))
    loss = K.mean(loss)
    loss = K.reshape(loss, [1, 1])
    return loss 

def call(self, x, mask=None):
        h1 = x[0]
        h2 = x[1]    
        dif = K.sum(K.abs(h1-h2),axis=1)  
        h = K.exp(-dif)
        #print h.shape
        h=K.clip(h,1e-7,1.0-1e-7)
        h = K.reshape(h, (h.shape[0],1))
        return h 

def call(self, x, mask=None):
        W = tf.transpose(self.tied_to.W, (1, 0, 2, 3))
        output = K.conv2d(x, W, strides=self.subsample,
                          border_mode=self.border_mode,
                          dim_ordering=self.dim_ordering,
                          filter_shape=self.W_shape)
        if self.bias:
            if self.dim_ordering == 'th':
                output += K.reshape(self.b, (1, self.nb_filter, 1, 1))
            elif self.dim_ordering == 'tf':
                output += K.reshape(self.b, (1, 1, 1, self.nb_filter))
            else:
                raise Exception('Invalid dim_ordering: ' + self.dim_ordering)
        output = self.activation(output)
        return output 

def _time_distributed_dense(x, w, b=None, dropout=None,
                            input_dim=None, output_dim=None,
                            timesteps=None, training=None):
    """Apply `y . w + b` for every temporal slice y of x.

    # Arguments
        x: input tensor.
        w: weight matrix.
        b: optional bias vector.
        dropout: wether to apply dropout (same dropout mask
            for every temporal slice of the input).
        input_dim: integer; optional dimensionality of the input.
        output_dim: integer; optional dimensionality of the output.
        timesteps: integer; optional number of timesteps.
        training: training phase tensor or boolean.

    # Returns
        Output tensor.
    """
    if not input_dim:
        input_dim = K.shape(x)[2]
    if not timesteps:
        timesteps = K.shape(x)[1]
    if not output_dim:
        output_dim = K.int_shape(w)[1]

    if dropout is not None and 0. < dropout < 1.:
        # apply the same dropout pattern at every timestep
        ones = K.ones_like(K.reshape(x[:, 0, :], (-1, input_dim)))
        dropout_matrix = K.dropout(ones, dropout)
        expanded_dropout_matrix = K.repeat(dropout_matrix, timesteps)
        x = K.in_train_phase(x * expanded_dropout_matrix, x, training=training)

    # collapse time dimension and batch dimension together
    x = K.reshape(x, (-1, input_dim))
    x = K.dot(x, w)
    if b is not None:
        x = K.bias_add(x, b)
    # reshape to 3D tensor
    if K.backend() == 'tensorflow':
        x = K.reshape(x, K.stack([-1, timesteps, output_dim]))
        x.set_shape([None, None, output_dim])
    else:
        x = K.reshape(x, (-1, timesteps, output_dim))
    return x 

def _time_distributed_dense(x, w, b=None, dropout=None,
                            input_dim=None, output_dim=None,
                            timesteps=None, training=None):
    """Apply `y . w + b` for every temporal slice y of x.

    # Arguments
        x: input tensor.
        w: weight matrix.
        b: optional bias vector.
        dropout: wether to apply dropout (same dropout mask
            for every temporal slice of the input).
        input_dim: integer; optional dimensionality of the input.
        output_dim: integer; optional dimensionality of the output.
        timesteps: integer; optional number of timesteps.
        training: training phase tensor or boolean.

    # Returns
        Output tensor.
    """
    if not input_dim:
        input_dim = K.shape(x)[2]
    if not timesteps:
        timesteps = K.shape(x)[1]
    if not output_dim:
        output_dim = K.int_shape(w)[1]

    if dropout is not None and 0. < dropout < 1.:
        # apply the same dropout pattern at every timestep
        ones = K.ones_like(K.reshape(x[:, 0, :], (-1, input_dim)))
        dropout_matrix = K.dropout(ones, dropout)
        expanded_dropout_matrix = K.repeat(dropout_matrix, timesteps)
        x = K.in_train_phase(x * expanded_dropout_matrix, x, training=training)

    # collapse time dimension and batch dimension together
    x = K.reshape(x, (-1, input_dim))
    x = K.dot(x, w)
    if b is not None:
        x = K.bias_add(x, b)
    # reshape to 3D tensor
    if K.backend() == 'tensorflow':
        x = K.reshape(x, K.stack([-1, timesteps, output_dim]))
        x.set_shape([None, None, output_dim])
    else:
        x = K.reshape(x, (-1, timesteps, output_dim))
    return x 

def yolo_predict(yolo_outputs,
                 anchors,
                 num_classes,
                 image_shape,
                 max_boxes=20,
                 score_threshold=.6,
                 iou_threshold=.5):
    """
    Get prediction from YOLO model on given input and return filtered boxes.
    """
    num_layers = len(yolo_outputs)
    anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]

    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        # Get the boxes and the confidences for each box
        box_xy, box_wh, box_confidence, box_class_probs = yolo_head(yolo_outputs[l],
                                                                    anchors[anchor_mask[l]],
                                                                    num_classes,
                                                                    input_shape)
        _boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)
        _boxes = K.reshape(_boxes, [-1, 4])
        _box_scores = box_confidence * box_class_probs
        _box_scores = K.reshape(_box_scores, [-1, num_classes])

        boxes.append(_boxes)
        box_scores.append(_box_scores)

    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        # TensorFlow function for non max suppression of detection candidates
        # This is for filtering overlapping bounding boxes
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def call(self, inputs):
        # Box Scores. Use the foreground class confidence. [Batch, num_rois, 1]
        scores = inputs[0][:, :, 1]
        # Box deltas [batch, num_rois, 4]
        deltas = inputs[1]
        deltas = deltas * np.reshape(self.config.RPN_BBOX_STD_DEV, [1, 1, 4])
        # Anchors
        anchors = inputs[2]

        # Improve performance by trimming to top anchors by score
        # and doing the rest on the smaller subset.
        pre_nms_limit = tf.minimum(6000, tf.shape(anchors)[1])
        ix = tf.nn.top_k(scores, pre_nms_limit, sorted=True,
                         name="top_anchors").indices
        scores = utils.batch_slice([scores, ix], lambda x, y: tf.gather(x, y),
                                   self.config.IMAGES_PER_GPU)
        deltas = utils.batch_slice([deltas, ix], lambda x, y: tf.gather(x, y),
                                   self.config.IMAGES_PER_GPU)
        pre_nms_anchors = utils.batch_slice([anchors, ix], lambda a, x: tf.gather(a, x),
                                    self.config.IMAGES_PER_GPU,
                                    names=["pre_nms_anchors"])

        # Apply deltas to anchors to get refined anchors.
        # [batch, N, (y1, x1, y2, x2)]
        boxes = utils.batch_slice([pre_nms_anchors, deltas],
                                  lambda x, y: apply_box_deltas_graph(x, y),
                                  self.config.IMAGES_PER_GPU,
                                  names=["refined_anchors"])

        # Clip to image boundaries. Since we're in normalized coordinates,
        # clip to 0..1 range. [batch, N, (y1, x1, y2, x2)]
        window = np.array([0, 0, 1, 1], dtype=np.float32)
        boxes = utils.batch_slice(boxes,
                                  lambda x: clip_boxes_graph(x, window),
                                  self.config.IMAGES_PER_GPU,
                                  names=["refined_anchors_clipped"])

        # Filter out small boxes
        # According to Xinlei Chen's paper, this reduces detection accuracy
        # for small objects, so we're skipping it.

        # Non-max suppression
        def nms(boxes, scores):
            indices = tf.image.non_max_suppression(
                boxes, scores, self.proposal_count,
                self.nms_threshold, name="rpn_non_max_suppression")
            proposals = tf.gather(boxes, indices)
            # Pad if needed
            padding = tf.maximum(self.proposal_count - tf.shape(proposals)[0], 0)
            proposals = tf.pad(proposals, [(0, padding), (0, 0)])
            return proposals
        proposals = utils.batch_slice([boxes, scores], nms,
                                      self.config.IMAGES_PER_GPU)
        return proposals 

def rpn_graph(feature_map, anchors_per_location, anchor_stride):
    """Builds the computation graph of Region Proposal Network.

    feature_map: backbone features [batch, height, width, depth]
    anchors_per_location: number of anchors per pixel in the feature map
    anchor_stride: Controls the density of anchors. Typically 1 (anchors for
                   every pixel in the feature map), or 2 (every other pixel).

    Returns:
        rpn_logits: [batch, H, W, 2] Anchor classifier logits (before softmax)
        rpn_probs: [batch, H, W, 2] Anchor classifier probabilities.
        rpn_bbox: [batch, H, W, (dy, dx, log(dh), log(dw))] Deltas to be
                  applied to anchors.
    """
    # TODO: check if stride of 2 causes alignment issues if the feature map
    # is not even.
    # Shared convolutional base of the RPN
    shared = KL.Conv2D(512, (3, 3), padding='same', activation='relu',
                       strides=anchor_stride,
                       name='rpn_conv_shared')(feature_map)

    # Anchor Score. [batch, height, width, anchors per location * 2].
    x = KL.Conv2D(2 * anchors_per_location, (1, 1), padding='valid',
                  activation='linear', name='rpn_class_raw')(shared)

    # Reshape to [batch, anchors, 2]
    rpn_class_logits = KL.Lambda(
        lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 2]))(x)

    # Softmax on last dimension of BG/FG.
    rpn_probs = KL.Activation(
        "softmax", name="rpn_class_xxx")(rpn_class_logits)

    # Bounding box refinement. [batch, H, W, anchors per location, depth]
    # where depth is [x, y, log(w), log(h)]
    x = KL.Conv2D(anchors_per_location * 4, (1, 1), padding="valid",
                  activation='linear', name='rpn_bbox_pred')(shared)

    # Reshape to [batch, anchors, 4]
    rpn_bbox = KL.Lambda(lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 4]))(x)

    return [rpn_class_logits, rpn_probs, rpn_bbox] 

def call(self, inputs, **kwargs):
        if self.axis == 1:
            # If channels first, force it to be channels last for these ops
            inputs = K.permute_dimensions(inputs, [0, 2, 3, 1])

        q, k, v = tf.split(inputs, [self.depth_k, self.depth_k, self.depth_v], axis=-1)

        q = self.split_heads_2d(q)
        k = self.split_heads_2d(k)
        v = self.split_heads_2d(v)

        # scale query
        depth_k_heads = self.depth_k / self.num_heads
        q *= (depth_k_heads ** -0.5)

        # [Batch, num_heads, height * width, depth_k or depth_v] if axis == -1
        qk_shape = [self._batch, self.num_heads, self._height * self._width, self.depth_k // self.num_heads]
        v_shape = [self._batch, self.num_heads, self._height * self._width, self.depth_v // self.num_heads]
        flat_q = K.reshape(q, K.stack(qk_shape))
        flat_k = K.reshape(k, K.stack(qk_shape))
        flat_v = K.reshape(v, K.stack(v_shape))

        # [Batch, num_heads, HW, HW]
        logits = tf.matmul(flat_q, flat_k, transpose_b=True)

        # Apply relative encodings
        if self.relative:
            h_rel_logits, w_rel_logits = self.relative_logits(q)
            logits += h_rel_logits
            logits += w_rel_logits

        weights = K.softmax(logits, axis=-1)
        attn_out = tf.matmul(weights, flat_v)

        attn_out_shape = [self._batch, self.num_heads, self._height, self._width, self.depth_v // self.num_heads]
        attn_out_shape = K.stack(attn_out_shape)
        attn_out = K.reshape(attn_out, attn_out_shape)
        attn_out = self.combine_heads_2d(attn_out)
        # [batch, height, width, depth_v]

        if self.axis == 1:
            # return to [batch, depth_v, height, width] for channels first
            attn_out = K.permute_dimensions(attn_out, [0, 3, 1, 2])

        return attn_out 

def call(self, inputs, **kwargs):
        # decode the reasoning probabilities
        positive_kernel = self.reasoning_probabilities[0]
        negative_kernel = (1 - positive_kernel) * \
                          self.reasoning_probabilities[1]

        if self.use_component_probabilities:
            # squash component probabilities
            components_probabilities = softmax(self.component_probabilities,
                                               axis=2)

            positive_kernel = positive_kernel * components_probabilities
            negative_kernel = negative_kernel * components_probabilities

        # get normalization tensor
        # stabilize the division with a small epsilon
        normalization = K.sum(positive_kernel + negative_kernel,
                              axis=2,
                              keepdims=True) + K.epsilon()

        # get sliding kernel and bias
        if self.use_pixel_probabilities:
            pixel_probabilities = softmax(self.pixel_probabilities,
                                          axis=(0, 1))
            # scale kernel with priors
            kernel = (positive_kernel - negative_kernel) / normalization \
                     * pixel_probabilities
            bias = K.sum(negative_kernel / normalization
                         * pixel_probabilities,
                         axis=(0, 1, 2),
                         keepdims=True)
        else:
            kernel = (positive_kernel - negative_kernel) / normalization
            bias = K.sum(negative_kernel / normalization,
                         axis=(0, 1, 2),
                         keepdims=True)

        # compute probabilities by a sliding operation
        probs = K.conv2d(inputs, kernel,
                         strides=self.strides,
                         padding=self.padding,
                         data_format='channels_last',
                         dilation_rate=self.dilation_rate) + bias

        if not self.use_pixel_probabilities:
            # divide by number of kernel_size
            probs = probs / np.prod(self.kernel_size)

        # reshape to m x n x #classes x #replicas
        probs = K.reshape(probs,
                          (-1,) + K.int_shape(probs)[1:3]
                          + (self.n_classes, self.n_replicas))

        # squeeze replica dimension if one.
        if self.n_replicas == 1:
            probs = K.squeeze(probs, axis=-1)

        return probs 

def get_training_data( images,landmarks,batch_size):
  while 1:
    indices = numpy.random.choice(range(0,images.shape[0]),size=batch_size,replace=True)
    for i,index in enumerate(indices):
      image = images[index]
      seed  = int(time.time())
      image = random_transform( image, seed, **random_transform_args )
      closest = ( numpy.mean(numpy.square(landmarks[index]-landmarks),axis=(1,2)) ).argsort()[1:20]
      closest = numpy.random.choice(closest, 6, replace=False)
      closestMerged = numpy.dstack([  
                                      cv2.resize( random_transform( images[closest[0]][:,:,:3] ,seed, **random_transform_args) , (64,64)), 
                                      cv2.resize( random_transform( images[closest[1]][:,:,:3] ,seed, **random_transform_args) , (64,64)), 
                                      cv2.resize( random_transform( images[closest[2]][:,:,:3] ,seed, **random_transform_args) , (64,64)),
                                      cv2.resize( random_transform( images[closest[3]][:,:,:3] ,seed, **random_transform_args) , (64,64)),
                                      cv2.resize( random_transform( images[closest[4]][:,:,:3] ,seed, **random_transform_args) , (64,64)),
                                      cv2.resize( random_transform( images[closest[5]][:,:,:3] ,seed, **random_transform_args) , (64,64)),
                                    ])

      if i == 0:
          warped_images  = numpy.empty( (batch_size,)  + (64,64,3),   image.dtype )
          example_images = numpy.empty( (batch_size,)  + (64,64,18),  image.dtype )
          target_images  = numpy.empty( (batch_size,)  + (128,128,3), image.dtype )
          mask_images    = numpy.empty( (batch_size,)  + (128,128,1), image.dtype )

      warped_image =  random_warp( image[:,:,:3] )

      warped_image =  cv2.GaussianBlur( warped_image,(91,91),0 )

      image_mask = image[:,:,3].reshape((image.shape[0],image.shape[1],1)) * numpy.ones((image.shape[0],image.shape[1],3)).astype(float)


      foreground = cv2.multiply(image_mask, warped_image.astype(float))
      background = cv2.multiply(1.0 - image_mask, image[:,:,:3].astype(float))

      warped_image = numpy.add(background,foreground)

      warped_image = cv2.resize(warped_image,(64,64))

      warped_images[i]  = warped_image
      example_images[i] = closestMerged
      target_images[i]  = cv2.resize( image[:,:,:3], (128,128) )
      mask_images[i]    = cv2.resize( image[:,:,3], (128,128) ).reshape((128,128,1))
    yield warped_images,example_images,target_images,mask_images 

def call(self, inputs, training=None):
        # inputs.shape=[None, input_num_capsule, input_dim_vector]
        # Expand dims to [None, input_num_capsule, 1, 1, input_dim_vector]
        inputs_expand = K.expand_dims(K.expand_dims(inputs, 2), 2)

        # Replicate num_capsule dimension to prepare being multiplied by W
        # Now it has shape = [None, input_num_capsule, num_capsule, 1, input_dim_vector]
        inputs_tiled = K.tile(inputs_expand, [1, 1, self.num_capsule, 1, 1])

        """
        # Begin: inputs_hat computation V1 ---------------------------------------------------------------------#
        # Compute `inputs * W` by expanding the first dim of W. More time-consuming and need batch_size.
        # w_tiled.shape = [batch_size, input_num_capsule, num_capsule, input_dim_vector, dim_vector]
        w_tiled = K.tile(K.expand_dims(self.W, 0), [self.batch_size, 1, 1, 1, 1])

        # Transformed vectors, inputs_hat.shape = [None, input_num_capsule, num_capsule, 1, dim_vector]
        inputs_hat = K.batch_dot(inputs_tiled, w_tiled, [4, 3])
        # End: inputs_hat computation V1 ---------------------------------------------------------------------#
        """

        # Begin: inputs_hat computation V2 ---------------------------------------------------------------------#
        # Compute `inputs * W` by scanning inputs_tiled on dimension 0. This is faster but requires Tensorflow.
        # inputs_hat.shape = [None, input_num_capsule, num_capsule, 1, dim_vector]
        inputs_hat = tf.scan(lambda ac, x: K.batch_dot(x, self.W, [3, 2]),
                             elems=inputs_tiled,
                             initializer=K.zeros([self.input_num_capsule, self.num_capsule, 1, self.dim_vector]))
        # End: inputs_hat computation V2 ---------------------------------------------------------------------#
        """
        # Begin: routing algorithm V1, dynamic ------------------------------------------------------------#
        def body(i, b, outputs):
            c = tf.nn.softmax(b, dim=2)  # dim=2 is the num_capsule dimension
            outputs = squash(K.sum(c * inputs_hat, 1, keepdims=True))
            if i != 1:
                b = b + K.sum(inputs_hat * outputs, -1, keepdims=True)
            return [i-1, b, outputs]
        cond = lambda i, b, inputs_hat: i > 0
        loop_vars = [K.constant(self.num_routing), self.bias, K.sum(inputs_hat, 1, keepdims=True)]
        shape_invariants = [tf.TensorShape([]),
                            tf.TensorShape([None, self.input_num_capsule, self.num_capsule, 1, 1]),
                            tf.TensorShape([None, 1, self.num_capsule, 1, self.dim_vector])]
        _, _, outputs = tf.while_loop(cond, body, loop_vars, shape_invariants)
        # End: routing algorithm V1, dynamic ------------------------------------------------------------#
        """

        # Begin: routing algorithm V2, static -----------------------------------------------------------#
        # Routing algorithm V2. Use iteration. V2 and V1 both work without much difference on performance
        assert self.num_routing > 0, 'The num_routing should be > 0.'
        for i in range(self.num_routing):
            c = tf.nn.softmax(self.bias, dim=2)  # dim=2 is the num_capsule dimension
            # outputs.shape=[None, 1, num_capsule, 1, dim_vector]
            outputs = squash(K.sum(c * inputs_hat, 1, keepdims=True))

            # last iteration needs not compute bias which will not be passed to the graph any more anyway.
            if i != self.num_routing - 1:
                # self.bias = K.update_add(self.bias, K.sum(inputs_hat * outputs, [0, -1], keepdims=True))
                self.bias += K.sum(inputs_hat * outputs, -1, keepdims=True)
            # tf.summary.histogram('BigBee', self.bias)  # for debugging
        # End: routing algorithm V2, static ------------------------------------------------------------#

        return K.reshape(outputs, [-1, self.num_capsule, self.dim_vector]) 

def call(self, inputs):
        # Box Scores. Use the foreground class confidence. [Batch, num_rois, 1]
        scores = inputs[0][:, :, 1]
        # Box deltas [batch, num_rois, 4]
        deltas = inputs[1]
        deltas = deltas * np.reshape(self.config.RPN_BBOX_STD_DEV, [1, 1, 4])
        # Base anchors
        anchors = self.anchors

        # Improve performance by trimming to top anchors by score
        # and doing the rest on the smaller subset.
        pre_nms_limit = min(6000, self.anchors.shape[0])
        ix = tf.nn.top_k(scores, pre_nms_limit, sorted=True,
                         name="top_anchors").indices
        scores = utils.batch_slice([scores, ix], lambda x, y: tf.gather(x, y),
                                   self.config.IMAGES_PER_GPU)
        deltas = utils.batch_slice([deltas, ix], lambda x, y: tf.gather(x, y),
                                   self.config.IMAGES_PER_GPU)
        anchors = utils.batch_slice(ix, lambda x: tf.gather(anchors, x),
                                    self.config.IMAGES_PER_GPU,
                                    names=["pre_nms_anchors"])

        # Apply deltas to anchors to get refined anchors.
        # [batch, N, (y1, x1, y2, x2)]
        boxes = utils.batch_slice([anchors, deltas],
                                  lambda x, y: apply_box_deltas_graph(x, y),
                                  self.config.IMAGES_PER_GPU,
                                  names=["refined_anchors"])

        # Clip to image boundaries. [batch, N, (y1, x1, y2, x2)]
        height, width = self.config.IMAGE_SHAPE[:2]
        window = np.array([0, 0, height, width]).astype(np.float32)
        boxes = utils.batch_slice(boxes,
                                  lambda x: clip_boxes_graph(x, window),
                                  self.config.IMAGES_PER_GPU,
                                  names=["refined_anchors_clipped"])

        # Filter out small boxes
        # According to Xinlei Chen's paper, this reduces detection accuracy
        # for small objects, so we're skipping it.

        # Normalize dimensions to range of 0 to 1.
        normalized_boxes = boxes / np.array([[height, width, height, width]])

        # Non-max suppression
        def nms(normalized_boxes, scores):
            indices = tf.image.non_max_suppression(
                normalized_boxes, scores, self.proposal_count,
                self.nms_threshold, name="rpn_non_max_suppression")
            proposals = tf.gather(normalized_boxes, indices)
            # Pad if needed
            padding = tf.maximum(self.proposal_count - tf.shape(proposals)[0], 0)
            proposals = tf.pad(proposals, [(0, padding), (0, 0)])
            return proposals
        proposals = utils.batch_slice([normalized_boxes, scores], nms,
                                      self.config.IMAGES_PER_GPU)
        return proposals 

def rpn_graph(feature_map, anchors_per_location, anchor_stride):
    """Builds the computation graph of Region Proposal Network.

    feature_map: backbone features [batch, height, width, depth]
    anchors_per_location: number of anchors per pixel in the feature map
    anchor_stride: Controls the density of anchors. Typically 1 (anchors for
                   every pixel in the feature map), or 2 (every other pixel).

    Returns:
        rpn_logits: [batch, H, W, 2] Anchor classifier logits (before softmax)
        rpn_probs: [batch, W, W, 2] Anchor classifier probabilities.
        rpn_bbox: [batch, H, W, (dy, dx, log(dh), log(dw))] Deltas to be
                  applied to anchors.
    """
    # TODO: check if stride of 2 causes alignment issues if the featuremap
    #       is not even.
    # Shared convolutional base of the RPN
    shared = KL.Conv2D(512, (3, 3), padding='same', activation='relu',
                       strides=anchor_stride,
                       name='rpn_conv_shared')(feature_map)

    # Anchor Score. [batch, height, width, anchors per location * 2].
    x = KL.Conv2D(2 * anchors_per_location, (1, 1), padding='valid',
                  activation='linear', name='rpn_class_raw')(shared)

    # Reshape to [batch, anchors, 2]
    rpn_class_logits = KL.Lambda(
        lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 2]))(x)

    # Softmax on last dimension of BG/FG.
    rpn_probs = KL.Activation(
        "softmax", name="rpn_class_xxx")(rpn_class_logits)

    # Bounding box refinement. [batch, H, W, anchors per location, depth]
    # where depth is [x, y, log(w), log(h)]
    x = KL.Conv2D(anchors_per_location * 4, (1, 1), padding="valid",
                  activation='linear', name='rpn_bbox_pred')(shared)

    # Reshape to [batch, anchors, 4]
    rpn_bbox = KL.Lambda(lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 4]))(x)

    return [rpn_class_logits, rpn_probs, rpn_bbox] 

def make_parallel(self):
        """Creates a new wrapper model that consists of multiple replicas of
        the original model placed on different GPUs.
        """
        # Slice inputs. Slice inputs on the CPU to avoid sending a copy
        # of the full inputs to all GPUs. Saves on bandwidth and memory.
        input_slices = {name: tf.split(x, self.gpu_count)
                        for name, x in zip(self.inner_model.input_names,
                                           self.inner_model.inputs)}

        output_names = self.inner_model.output_names
        outputs_all = []
        for i in range(len(self.inner_model.outputs)):
            outputs_all.append([])

        # Run the model call() on each GPU to place the ops there
        for i in range(self.gpu_count):
            with tf.device('/gpu:%d' % i):
                with tf.name_scope('tower_%d' % i):
                    # Run a slice of inputs through this replica
                    zipped_inputs = zip(self.inner_model.input_names,
                                        self.inner_model.inputs)
                    inputs = [
                        KL.Lambda(lambda s: input_slices[name][i],
                                  output_shape=lambda s: (None,) + s[1:])(tensor)
                        for name, tensor in zipped_inputs]
                    # Create the model replica and get the outputs
                    outputs = self.inner_model(inputs)
                    if not isinstance(outputs, list):
                        outputs = [outputs]
                    # Save the outputs for merging back together later
                    for l, o in enumerate(outputs):
                        outputs_all[l].append(o)

        # Merge outputs on CPU
        with tf.device('/cpu:0'):
            merged = []
            for outputs, name in zip(outputs_all, output_names):
                # If outputs are numbers without dimensions, add a batch dim.
                def add_dim(tensor):
                    """Add a dimension to tensors that don't have any."""
                    if K.int_shape(tensor) == ():
                        return KL.Lambda(lambda t: K.reshape(t, [1, 1]))(tensor)
                    return tensor
                outputs = list(map(add_dim, outputs))

                # Concatenate
                merged.append(KL.Concatenate(axis=0, name=name)(outputs))
        return merged