skeletonexport.py

import json
import logging
import networkx as nx
import pytz
from itertools import imap
from functools import partial
from collections import defaultdict
from math import sqrt
from datetime import datetime

from django.core.serializers.json import DjangoJSONEncoder
from django.db import connection
from django.http import HttpResponse

from rest_framework.decorators import api_view

from catmaid.models import UserRole, ClassInstance, Treenode, \
        TreenodeClassInstance, ConnectorClassInstance, Review
from catmaid.control import export_NeuroML_Level3
from catmaid.control.authentication import requires_user_role
from catmaid.control.common import get_relation_to_id_map
from catmaid.control.review import get_treenodes_to_reviews, \
        get_treenodes_to_reviews_with_time

from tree_util import edge_count_to_root, partition
try:
    from exportneuroml import neuroml_single_cell, neuroml_network
except ImportError:
    logging.getLogger(__name__).warn("NeuroML module could not be loaded.")


def get_treenodes_qs(project_id=None, skeleton_id=None, with_labels=True):
    treenode_qs = Treenode.objects.filter(skeleton_id=skeleton_id)
    if with_labels:
        labels_qs = TreenodeClassInstance.objects.filter(
            relation__relation_name='labeled_as',
            treenode__skeleton_id=skeleton_id).select_related('treenode', 'class_instance')
        labelconnector_qs = ConnectorClassInstance.objects.filter(
            relation__relation_name='labeled_as',
            connector__treenodeconnector__treenode__skeleton_id=skeleton_id).select_related('connector', 'class_instance')
    else:
        labels_qs = []
        labelconnector_qs = []
    return treenode_qs, labels_qs, labelconnector_qs


def get_swc_string(treenodes_qs):
    all_rows = []
    for tn in treenodes_qs:
        swc_row = [tn.id]
        swc_row.append(0)
        swc_row.append(tn.location_x)
        swc_row.append(tn.location_y)
        swc_row.append(tn.location_z)
        swc_row.append(max(tn.radius, 0))
        swc_row.append(-1 if tn.parent_id is None else tn.parent_id)
        all_rows.append(swc_row)
    result = ""
    for row in all_rows:
        result += " ".join(map(str, row)) + "\n"
    return result

def export_skeleton_response(request, project_id=None, skeleton_id=None, format=None):
    treenode_qs, labels_qs, labelconnector_qs = get_treenodes_qs(project_id, skeleton_id)

    if format == 'swc':
        return HttpResponse(get_swc_string(treenode_qs), content_type='text/plain')
    elif format == 'json':
        return HttpResponse(get_json_string(treenode_qs), content_type='application/json')
    else:
        raise Exception, "Unknown format ('%s') in export_skeleton_response" % (format,)


@requires_user_role(UserRole.Browse)
def compact_skeleton(request, project_id=None, skeleton_id=None, with_connectors=None, with_tags=None):
    """
        Performance-critical function. Do not edit unless to improve performance.

        Returns, in JSON, [[nodes], [connectors], {nodeID: [tags]}], with connectors and tags being empty when 0 == with_connectors and 0 == with_tags, respectively
    """

    # Sanitize
    project_id = int(project_id)
    skeleton_id = int(skeleton_id)
    with_connectors  = int(with_connectors)
    with_tags = int(with_tags)

    cursor = connection.cursor()

    cursor.execute('''
        SELECT id, parent_id, user_id,
               location_x, location_y, location_z,
               radius, confidence
        FROM treenode
        WHERE skeleton_id = %s
    ''' % skeleton_id)

    nodes = tuple(cursor.fetchall())

    if 0 == len(nodes):
        # Check if the skeleton exists
        if 0 == ClassInstance.objects.filter(pk=skeleton_id).count():
            raise Exception("Skeleton #%s doesn't exist" % skeleton_id)
        # Otherwise returns an empty list of nodes

    connectors = ()
    tags = defaultdict(list)

    if 0 != with_connectors or 0 != with_tags:
        # postgres is caching this query
        cursor.execute("SELECT relation_name, id FROM relation WHERE project_id=%s" % project_id)
        relations = dict(cursor.fetchall())

    if 0 != with_connectors:
        # Fetch all connectors with their partner treenode IDs
        pre = relations['presynaptic_to']
        post = relations['postsynaptic_to']
        gj = relations.get('gapjunction_with', -1)
        cursor.execute('''
            SELECT tc.treenode_id, tc.connector_id, tc.relation_id,
                   c.location_x, c.location_y, c.location_z
            FROM treenode_connector tc,
                 connector c
            WHERE tc.skeleton_id = %s
              AND tc.connector_id = c.id
              AND (tc.relation_id = %s OR tc.relation_id = %s OR tc.relation_id = %s)
        ''' % (skeleton_id, pre, post, gj))

        relation_index = {pre: 0, post: 1, gj: 2}

        connectors = tuple((row[0], row[1], relation_index.get(row[2], -1), row[3], row[4], row[5]) for row in cursor.fetchall())

    if 0 != with_tags:
        # Fetch all node tags
        cursor.execute('''
            SELECT c.name, tci.treenode_id
            FROM treenode t,
                 treenode_class_instance tci,
                 class_instance c
            WHERE t.skeleton_id = %s
              AND t.id = tci.treenode_id
              AND tci.relation_id = %s
              AND c.id = tci.class_instance_id
        ''' % (skeleton_id, relations['labeled_as']))

        for row in cursor.fetchall():
            tags[row[0]].append(row[1])

    return HttpResponse(json.dumps((nodes, connectors, tags), separators=(',', ':')))


@requires_user_role(UserRole.Browse)
def compact_arbor(request, project_id=None, skeleton_id=None, with_nodes=None, with_connectors=None, with_tags=None):
    """
    Performance-critical function. Do not edit unless to improve performance.
    Returns, in JSON, [[nodes], [connections], {nodeID: [tags]}],
    with connections being empty when 0 == with_connectors,
    and the dict of node tags being empty 0 == with_tags, respectively.

    The difference between this function and the compact_skeleton function is that
    the connections contain the whole chain from the skeleton of interest to the
    partner skeleton:
    [treenode_id, confidence,
     connector_id,
     confidence, treenode_id, skeleton_id,
     relation_id, relation_id]
    where the first 2 values are from the given skeleton_id,
    then the connector_id,
    then the next 3 values are from the partner skeleton,
    and finally the two relations: first for the given skeleton_id and then for the other skeleton.
    The relation_id is 0 for pre and 1 for post.
    """

    # Sanitize
    project_id = int(project_id)
    skeleton_id = int(skeleton_id)
    with_nodes = int(with_nodes)
    with_connectors  = int(with_connectors)
    with_tags = int(with_tags)

    cursor = connection.cursor()

    nodes = ()
    connectors = []
    tags = defaultdict(list)

    if 0 != with_nodes:
        cursor.execute('''
            SELECT id, parent_id, user_id,
                location_x, location_y, location_z,
                radius, confidence
            FROM treenode
            WHERE skeleton_id = %s
        ''' % skeleton_id)

        nodes = tuple(cursor.fetchall())

        if 0 == len(nodes):
            # Check if the skeleton exists
            if 0 == ClassInstance.objects.filter(pk=skeleton_id).count():
                raise Exception("Skeleton #%s doesn't exist" % skeleton_id)
            # Otherwise returns an empty list of nodes

    if 0 != with_connectors or 0 != with_tags:
        # postgres is caching this query
        cursor.execute("SELECT relation_name, id FROM relation WHERE project_id=%s" % project_id)
        relations = dict(cursor.fetchall())

    if 0 != with_connectors:
        # Fetch all inputs and outputs

        pre = relations['presynaptic_to']
        post = relations['postsynaptic_to']

        cursor.execute('''
            SELECT tc1.treenode_id, tc1.confidence,
                   tc1.connector_id,
                   tc2.confidence, tc2.treenode_id, tc2.skeleton_id,
                   tc1.relation_id, tc2.relation_id
            FROM treenode_connector tc1,
                 treenode_connector tc2
            WHERE tc1.skeleton_id = %s
              AND tc1.id != tc2.id
              AND tc1.connector_id = tc2.connector_id
              AND (tc1.relation_id = %s OR tc1.relation_id = %s)
        ''' % (skeleton_id, pre, post))

        for row in cursor.fetchall():
            # Ignore all other kinds of relation pairs (there shouldn't be any)
            if row[6] == pre and row[7] == post:
                connectors.append((row[0], row[1], row[2], row[3], row[4], row[5], 0, 1))
            elif row[6] == post and row[7] == pre:
                connectors.append((row[0], row[1], row[2], row[3], row[4], row[5], 1, 0))

    if 0 != with_tags:
        # Fetch all node tags
        cursor.execute('''
            SELECT c.name, tci.treenode_id
            FROM treenode t,
                 treenode_class_instance tci,
                 class_instance c
            WHERE t.skeleton_id = %s
              AND t.id = tci.treenode_id
              AND tci.relation_id = %s
              AND c.id = tci.class_instance_id
        ''' % (skeleton_id, relations['labeled_as']))

        for row in cursor.fetchall():
            tags[row[0]].append(row[1])

    return HttpResponse(json.dumps((nodes, connectors, tags), separators=(',', ':')))


@requires_user_role([UserRole.Browse])
def treenode_time_bins(request, project_id=None, skeleton_id=None):
    """ Return a map of time bins (minutes) vs. list of nodes. """
    minutes = defaultdict(list)
    epoch = datetime.utcfromtimestamp(0).replace(tzinfo=pytz.utc)

    for row in Treenode.objects.filter(skeleton_id=int(skeleton_id)).values_list('id', 'creation_time'):
        minutes[int((row[1] - epoch).total_seconds() / 60)].append(row[0])

    return HttpResponse(json.dumps(minutes, separators=(',', ':')))


@requires_user_role([UserRole.Browse])
def compact_arbor_with_minutes(request, project_id=None, skeleton_id=None, with_nodes=None, with_connectors=None, with_tags=None):
    r = compact_arbor(request, project_id=project_id, skeleton_id=skeleton_id, with_nodes=with_nodes, with_connectors=with_connectors, with_tags=with_tags)
    r.content = "%s, %s]" % (r.content[:-1], treenode_time_bins(request, project_id=project_id, skeleton_id=skeleton_id).content)
    return r


# DEPRECATED. Will be removed.
def _skeleton_for_3d_viewer(skeleton_id, project_id, with_connectors=True, lean=0, all_field=False):
    """ with_connectors: when False, connectors are not returned
        lean: when not zero, both connectors and tags are returned as empty arrays. """
    skeleton_id = int(skeleton_id) # sanitize
    cursor = connection.cursor()

    # Fetch the neuron name
    cursor.execute(
        '''SELECT name
           FROM class_instance ci,
                class_instance_class_instance cici
           WHERE cici.class_instance_a = %s
             AND cici.class_instance_b = ci.id
        ''' % skeleton_id)
    row = cursor.fetchone()
    if not row:
        # Check that the skeleton exists
        cursor.execute('''SELECT id FROM class_instance WHERE id=%s''' % skeleton_id)
        if not cursor.fetchone():
            raise Exception("Skeleton #%s doesn't exist!" % skeleton_id)
        else:
            raise Exception("No neuron found for skeleton #%s" % skeleton_id)

    name = row[0]

    if all_field:
        added_fields = ', creation_time, edition_time'
    else:
        added_fields = ''

    # Fetch all nodes, with their tags if any
    cursor.execute(
        '''SELECT id, parent_id, user_id, location_x, location_y, location_z, radius, confidence %s
          FROM treenode
          WHERE skeleton_id = %s
        ''' % (added_fields, skeleton_id) )

    # array of properties: id, parent_id, user_id, x, y, z, radius, confidence
    nodes = tuple(cursor.fetchall())

    tags = defaultdict(list) # node ID vs list of tags
    connectors = []

    # Get all reviews for this skeleton
    if all_field:
        reviews = get_treenodes_to_reviews_with_time(skeleton_ids=[skeleton_id])
    else:
        reviews = get_treenodes_to_reviews(skeleton_ids=[skeleton_id])

    if 0 == lean: # meaning not lean
        # Text tags
        cursor.execute("SELECT id FROM relation WHERE project_id=%s AND relation_name='labeled_as'" % int(project_id))
        labeled_as = cursor.fetchall()[0][0]

        cursor.execute(
             ''' SELECT treenode_class_instance.treenode_id, class_instance.name
                 FROM treenode, class_instance, treenode_class_instance
                 WHERE treenode.skeleton_id = %s
                   AND treenode.id = treenode_class_instance.treenode_id
                   AND treenode_class_instance.class_instance_id = class_instance.id
                   AND treenode_class_instance.relation_id = %s
             ''' % (skeleton_id, labeled_as))

        for row in cursor.fetchall():
            tags[row[1]].append(row[0])

        if with_connectors:
            if all_field:
                added_fields = ', c.creation_time'
            else:
                added_fields = ''

            # Fetch all connectors with their partner treenode IDs
            cursor.execute(
                ''' SELECT tc.treenode_id, tc.connector_id, r.relation_name,
                           c.location_x, c.location_y, c.location_z %s
                    FROM treenode_connector tc,
                         connector c,
                         relation r
                    WHERE tc.skeleton_id = %s
                      AND tc.connector_id = c.id
                      AND tc.relation_id = r.id
                ''' % (added_fields, skeleton_id) )
            # Above, purposefully ignoring connector tags. Would require a left outer join on the inner join of connector_class_instance and class_instance, and frankly connector tags are pointless in the 3d viewer.

            # List of (treenode_id, connector_id, relation_id, x, y, z)n with relation_id replaced by 0 (presynaptic) or 1 (postsynaptic)
            # 'presynaptic_to' has an 'r' at position 1:
            for row in cursor.fetchall():
                x, y, z = imap(float, (row[3], row[4], row[5]))
                connectors.append((row[0],
                                   row[1],
                                   0 if 'r' == row[2][1] else 1,
                                   x, y, z,
                                   row[6] if all_field else None))
            return name, nodes, tags, connectors, reviews

    return name, nodes, tags, connectors, reviews


# DEPRECATED. Will be removed.
@requires_user_role([UserRole.Annotate, UserRole.Browse])
def skeleton_for_3d_viewer(request, project_id=None, skeleton_id=None):
    return HttpResponse(json.dumps(_skeleton_for_3d_viewer(skeleton_id, project_id, with_connectors=request.POST.get('with_connectors', True), lean=int(request.POST.get('lean', 0)), all_field=request.POST.get('all_fields', False)), separators=(',', ':')))

# DEPRECATED. Will be removed.
@requires_user_role([UserRole.Annotate, UserRole.Browse])
def skeleton_with_metadata(request, project_id=None, skeleton_id=None):

    def default(obj):
        """Default JSON serializer."""
        import calendar, datetime

        if isinstance(obj, datetime.datetime):
            if obj.utcoffset() is not None:
                obj = obj - obj.utcoffset()
            millis = int(
                calendar.timegm(obj.timetuple()) * 1000 +
                obj.microsecond / 1000
            )
        return millis

    return HttpResponse(json.dumps(_skeleton_for_3d_viewer(skeleton_id, project_id, \
        with_connectors=True, lean=0, all_field=True), separators=(',', ':'), default=default))

def _measure_skeletons(skeleton_ids):
    if not skeleton_ids:
        raise Exception("Must provide the ID of at least one skeleton.")

    skids_string = ",".join(map(str, skeleton_ids))

    cursor = connection.cursor()
    cursor.execute('''
    SELECT id, parent_id, skeleton_id, location_x, location_y, location_z
    FROM treenode
    WHERE skeleton_id IN (%s)
    ''' % skids_string)

    # TODO should be all done with numpy,
    # TODO  by partitioning the skeleton into sequences of x,y,z representing the slabs
    # TODO  and then convolving them.

    class Skeleton():
        def __init__(self):
            self.nodes = {}
            self.raw_cable = 0
            self.smooth_cable = 0
            self.principal_branch_cable = 0
            self.n_ends = 0
            self.n_branch = 0
            self.n_pre = 0
            self.n_post = 0

    class Node():
        def __init__(self, parent_id, x, y, z):
            self.parent_id = parent_id
            self.x = x
            self.y = y
            self.z = z
            self.wx = x # weighted average of itself and neighbors
            self.wy = y
            self.wz = z
            self.children = {} # node ID vs distance

    skeletons = defaultdict(dict) # skeleton ID vs (node ID vs Node)
    for row in cursor.fetchall():
        skeleton = skeletons.get(row[2])
        if not skeleton:
            skeleton = Skeleton()
            skeletons[row[2]] = skeleton
        skeleton.nodes[row[0]] = Node(row[1], row[3], row[4], row[5])

    for skeleton in skeletons.itervalues():
        nodes = skeleton.nodes
        tree = nx.DiGraph()
        root = None
        # Accumulate children
        for nodeID, node in nodes.iteritems():
            if not node.parent_id:
                root = nodeID
                continue
            tree.add_edge(node.parent_id, nodeID)
            parent = nodes[node.parent_id]
            distance = sqrt(  pow(node.x - parent.x, 2)
                            + pow(node.y - parent.y, 2)
                            + pow(node.z - parent.z, 2))
            parent.children[nodeID] = distance
            # Measure raw cable, given that we have the parent already
            skeleton.raw_cable += distance
        # Utilize accumulated children and the distances to them
        for nodeID, node in nodes.iteritems():
            # Count end nodes and branch nodes
            n_children = len(node.children)
            if not node.parent_id:
                if 1 == n_children:
                    skeleton.n_ends += 1
                    continue
                if n_children > 2:
                    skeleton.n_branch += 1
                    continue
                # Else, if 2 == n_children, the root node is in the middle of the skeleton, being a slab node
            elif 0 == n_children:
                skeleton.n_ends += 1
                continue
            elif n_children > 1:
                skeleton.n_branch += 1
                continue
            # Compute weighted position for slab nodes only
            # (root, branch and end nodes do not move)
            oids = node.children.copy()
            if node.parent_id:
                oids[node.parent_id] = skeleton.nodes[node.parent_id].children[nodeID]
            sum_distances = sum(oids.itervalues())
            wx, wy, wz = 0, 0, 0
            for oid, distance in oids.iteritems():
                other = skeleton.nodes[oid]
                w = distance / sum_distances if sum_distances != 0 else 0
                wx += other.x * w
                wy += other.y * w
                wz += other.z * w
            node.wx = node.x * 0.4 + wx * 0.6
            node.wy = node.y * 0.4 + wy * 0.6
            node.wz = node.z * 0.4 + wz * 0.6
        # Find out nodes that belong to the principal branch
        principal_branch_nodes = set(sorted(partition(tree, root), key=len)[-1])
        # Compute smoothed cable length, also for principal branch
        for nodeID, node in nodes.iteritems():
            if not node.parent_id:
                # root node
                continue
            parent = nodes[node.parent_id]
            length = sqrt(  pow(node.wx - parent.wx, 2)
                          + pow(node.wy - parent.wy, 2)
                          + pow(node.wz - parent.wz, 2))
            skeleton.smooth_cable += length
            if nodeID in principal_branch_nodes:
                skeleton.principal_branch_cable += length

    # Count inputs
    cursor.execute('''
    SELECT tc.skeleton_id, count(tc.skeleton_id)
    FROM treenode_connector tc,
         relation r
    WHERE tc.skeleton_id IN (%s)
      AND tc.relation_id = r.id
      AND r.relation_name = 'postsynaptic_to'
    GROUP BY tc.skeleton_id
    ''' % skids_string)

    for row in cursor.fetchall():
        skeletons[row[0]].n_pre = row[1]

    # Count outputs
    cursor.execute('''
    SELECT tc1.skeleton_id, count(tc1.skeleton_id)
    FROM treenode_connector tc1,
         treenode_connector tc2,
         relation r1,
         relation r2
    WHERE tc1.skeleton_id IN (%s)
      AND tc1.connector_id = tc2.connector_id
      AND tc1.relation_id = r1.id
      AND r1.relation_name = 'presynaptic_to'
      AND tc2.relation_id = r2.id
      AND r2.relation_name = 'postsynaptic_to'
      GROUP BY tc1.skeleton_id
    ''' % skids_string)

    for row in cursor.fetchall():
        skeletons[row[0]].n_post = row[1]

    return skeletons


@requires_user_role([UserRole.Annotate, UserRole.Browse])
def measure_skeletons(request, project_id=None):
    skeleton_ids = tuple(int(v) for k,v in request.POST.iteritems() if k.startswith('skeleton_ids['))
    def asRow(skid, sk):
        return (skid, int(sk.raw_cable), int(sk.smooth_cable), sk.n_pre, sk.n_post, len(sk.nodes), sk.n_branch, sk.n_ends, sk.principal_branch_cable)
    return HttpResponse(json.dumps([asRow(skid, sk) for skid, sk in _measure_skeletons(skeleton_ids).iteritems()]))


def _skeleton_neuroml_cell(skeleton_id, preID, postID):
    skeleton_id = int(skeleton_id) # sanitize
    cursor = connection.cursor()

    cursor.execute('''
    SELECT id, parent_id, location_x, location_y, location_z, radius
    FROM treenode
    WHERE skeleton_id = %s
    ''' % skeleton_id)
    nodes = {row[0]: (row[1], (row[2], row[3], row[4]), row[5]) for row in cursor.fetchall()}

    cursor.execute('''
    SELECT tc.treenode_id, tc.connector_id, tc.relation_id
    FROM treenode_connector tc
    WHERE tc.skeleton_id = %s
      AND (tc.relation_id = %s OR tc.relation_id = %s)
    ''' % (skeleton_id, preID, postID))
    pre = defaultdict(list) # treenode ID vs list of connector ID
    post = defaultdict(list)
    for row in cursor.fetchall():
        if row[2] == preID:
            pre[row[0]].append(row[1])
        else:
            post[row[0]].append(row[1])

    return neuroml_single_cell(skeleton_id, nodes, pre, post)


@requires_user_role(UserRole.Browse)
def skeletons_neuroml(request, project_id=None):
    """ Export a list of skeletons each as a Cell in NeuroML. """
    project_id = int(project_id) # sanitize
    skeleton_ids = tuple(int(v) for k,v in request.POST.iteritems() if k.startswith('skids['))

    cursor = connection.cursor()

    relations = get_relation_to_id_map(project_id, ('presynaptic_to', 'postsynaptic_to'), cursor)
    preID = relations['presynaptic_to']
    postID = relations['postsynaptic_to']

    # TODO could certainly fetch all nodes and synapses in one single query and then split them up.
    cells = (_skeleton_neuroml_cell(skeleton_id, preID, postID) for skeleton_id in skeleton_ids)

    response = HttpResponse(content_type='text/txt')
    response['Content-Disposition'] = 'attachment; filename="data.neuroml"'

    neuroml_network(cells, response)

    return response


@requires_user_role(UserRole.Browse)
def export_neuroml_level3_v181(request, project_id=None):
    """Export the NeuroML Level 3 version 1.8.1 representation of one or more skeletons.
    Considers synapses among the requested skeletons only. """
    skeleton_ids = tuple(int(v) for v in request.POST.getlist('skids[]'))
    mode = int(request.POST.get('mode'))
    skeleton_strings = ",".join(map(str, skeleton_ids))
    cursor = connection.cursor()

    relations = get_relation_to_id_map(project_id, ('presynaptic_to', 'postsynaptic_to'), cursor)
    presynaptic_to = relations['presynaptic_to']
    postsynaptic_to = relations['postsynaptic_to']

    cursor.execute('''
    SELECT cici.class_instance_a, ci.name
    FROM class_instance_class_instance cici,
         class_instance ci,
         relation r
    WHERE cici.class_instance_a IN (%s)
      AND cici.class_instance_b = ci.id
      AND cici.relation_id = r.id
      AND r.relation_name = 'model_of'
    ''' % skeleton_strings)

    neuron_names = dict(cursor.fetchall())

    skeleton_query = '''
        SELECT id, parent_id, location_x, location_y, location_z,
               radius, skeleton_id
        FROM treenode
        WHERE skeleton_id IN (%s)
        ORDER BY skeleton_id
        ''' % skeleton_strings

    if 0 == mode:
        cursor.execute('''
        SELECT treenode_id, connector_id, relation_id, skeleton_id
        FROM treenode_connector
        WHERE skeleton_id IN (%s)
          AND (relation_id = %s OR relation_id = %s)
        ''' % (skeleton_strings, presynaptic_to, postsynaptic_to))

        # Dictionary of connector ID vs map of relation_id vs list of treenode IDs
        connectors = defaultdict(partial(defaultdict, list))

        for row in cursor.fetchall():
            connectors[row[1]][row[2]].append((row[0], row[3]))

        # Dictionary of presynaptic skeleton ID vs map of postsynaptic skeleton ID vs list of tuples with presynaptic treenode ID and postsynaptic treenode ID.
        connections = defaultdict(partial(defaultdict, list))

        for connectorID, m in connectors.iteritems():
            for pre_treenodeID, skID1 in m[presynaptic_to]:
                for post_treenodeID, skID2 in m[postsynaptic_to]:
                    connections[skID1][skID2].append((pre_treenodeID, post_treenodeID))

        cursor.execute(skeleton_query)

        generator = export_NeuroML_Level3.exportMutual(neuron_names, cursor.fetchall(), connections)

    else:
        if len(skeleton_ids) > 1:
            raise Exception("Expected a single skeleton for mode %s!" % mode)
        input_ids = tuple(int(v) for v in request.POST.getlist('inputs[]', []))
        input_strings = ",".join(map(str, input_ids))
        if 2 == mode:
            constraint = "AND tc2.skeleton_id IN (%s)" % input_strings
        elif 1 == mode:
            constraint = ""
        else:
            raise Exception("Unknown mode %s" % mode)

        cursor.execute('''
        SELECT tc2.skeleton_id, tc1.treenode_id
        FROM treenode_connector tc1,
             treenode_connector tc2
        WHERE tc1.skeleton_id = %s
          AND tc1.connector_id = tc2.connector_id
          AND tc1.treenode_id != tc2.treenode_id
          AND tc1.relation_id = %s
          AND tc2.relation_id = %s
          %s
        ''' % (skeleton_strings, postsynaptic_to, presynaptic_to, constraint))

        # Dictionary of skeleton ID vs list of treenode IDs at which the neuron receives inputs
        inputs = defaultdict(list)
        for row in cursor.fetchall():
            inputs[row[0]].append(row[1])

        cursor.execute(skeleton_query)

        generator = export_NeuroML_Level3.exportSingle(neuron_names, cursor.fetchall(), inputs)

    response = HttpResponse(generator, content_type='text/plain')
    response['Content-Disposition'] = 'attachment; filename=neuronal-circuit.neuroml'

    return response


@requires_user_role(UserRole.Browse)
def skeleton_swc(*args, **kwargs):
    kwargs['format'] = 'swc'
    return export_skeleton_response(*args, **kwargs)


def _export_review_skeleton(project_id=None, skeleton_id=None,
                            subarbor_node_id=None):
    """ Returns a list of segments for the requested skeleton. Each segment
    contains information about the review status of this part of the skeleton.
    If a valid subarbor_node_id is given, only data for the sub-arbor is
    returned that starts at this node.
    """
    # Get all treenodes of the requested skeleton
    cursor = connection.cursor()
    cursor.execute("""
            SELECT
                t.id,
                t.parent_id,
                t.location_x,
                t.location_y,
                t.location_z,
                ARRAY_AGG(svt.orientation),
                ARRAY_AGG(svt.location_coordinate)
            FROM treenode t
            LEFT OUTER JOIN suppressed_virtual_treenode svt
              ON (t.id = svt.child_id)
            WHERE t.skeleton_id = %s
            GROUP BY t.id;
            """, (skeleton_id,))
    treenodes = cursor.fetchall()
    # Get all reviews for the requested skeleton
    reviews = get_treenodes_to_reviews_with_time(skeleton_ids=[skeleton_id])

    if 0 == len(treenodes):
        return []

    # The root node will be assigned below, depending on retrieved nodes and
    # sub-arbor requests
    root_id = None

    # Add each treenode to a networkx graph and attach reviewer information to
    # it.
    g = nx.DiGraph()
    reviewed = set()
    for t in treenodes:
        # While at it, send the reviewer IDs, which is useful to iterate fwd
        # to the first unreviewed node in the segment.
        g.add_node(t[0], {'id': t[0],
                          'x': t[2],
                          'y': t[3],
                          'z': t[4],
                          'rids': reviews[t[0]],
                          'sup': [[o, l] for [o, l] in zip(t[5], t[6]) if o is not None]})
        if reviews[t[0]]:
            reviewed.add(t[0])
        if t[1]: # if parent
            g.add_edge(t[1], t[0]) # edge from parent to child
        else:
            root_id = t[0]

    if subarbor_node_id and subarbor_node_id != root_id:
        # Make sure the subarbor node ID (if any) is part of this skeleton
        if subarbor_node_id not in g:
            raise ValueError("Supplied subarbor node ID (%s) is not part of "
                             "provided skeleton (%s)" % (subarbor_node_id, skeleton_id))

        # Remove connection to parent
        parent = g.predecessors(subarbor_node_id)[0]
        g.remove_edge(parent, subarbor_node_id)
        # Remove all nodes that are upstream from the subarbor node
        to_delete = set()
        to_lookat = [root_id]
        while to_lookat:
            n = to_lookat.pop()
            to_lookat.extend(g.successors(n))
            to_delete.add(n)
        g.remove_nodes_from(to_delete)
        # Replace root id with sub-arbor ID
        root_id=subarbor_node_id

    if not root_id:
        if subarbor_node_id:
            raise ValueError("Couldn't find a reference root node in provided "
                             "skeleton (%s)" % (skeleton_id,))
        else:
            raise ValueError("Couldn't find a reference root node for provided "
                             "subarbor (%s) in provided skeleton (%s)" % (subarbor_node_id, skeleton_id))

    # Create all sequences, as long as possible and always from end towards root
    distances = edge_count_to_root(g, root_node=root_id) # distance in number of edges from root
    seen = set()
    sequences = []
    # Iterate end nodes sorted from highest to lowest distance to root
    endNodeIDs = (nID for nID in g.nodes() if 0 == len(g.successors(nID)))
    for nodeID in sorted(endNodeIDs, key=distances.get, reverse=True):
        sequence = [g.node[nodeID]]
        parents = g.predecessors(nodeID)
        while parents:
            parentID = parents[0]
            sequence.append(g.node[parentID])
            if parentID in seen:
                break
            seen.add(parentID)
            parents = g.predecessors(parentID)

        if len(sequence) > 1:
            sequences.append(sequence)

    # Calculate status

    segments = []
    for sequence in sorted(sequences, key=len, reverse=True):
        segments.append({
            'id': len(segments),
            'sequence': sequence,
            'status': '%.2f' % (100.0 * sum(1 for node in sequence if node['id'] in reviewed) / len(sequence)),
            'nr_nodes': len(sequence)
        })
    return segments

@api_view(['POST'])
@requires_user_role(UserRole.Browse)
def export_review_skeleton(request, project_id=None, skeleton_id=None):
    """Export skeleton as a set of segments with per-node review information.

    Export the skeleton as a list of segments of non-branching node paths,
    with detailed information on reviewers and review times for each node.
    ---
    parameters:
    - name: subarbor_node_id
      description: |
        If provided, only the subarbor starting at this treenode is returned.
      required: false
      type: integer
      paramType: form
    models:
      export_review_skeleton_segment:
        id: export_review_skeleton_segment
        properties:
          status:
            description: |
              Percentage of nodes in this segment reviewed by the request user
            type: number
            format: double
            required: true
          id:
            description: |
              Index of this segment in the list (order by descending segment
              node count)
            type: integer
            required: true
          nr_nodes:
            description: Number of nodes in this segment
            type: integer
            required: true
          sequence:
            description: Detail for nodes in this segment
            type: array
            items:
              type: export_review_skeleton_segment_node
            required: true
      export_review_skeleton_segment_node:
        id: export_review_skeleton_segment_node
        properties:
          id:
            description: ID of this treenode
            type: integer
            required: true
          x:
            type: double
            required: true
          y:
            type: double
            required: true
          z:
            type: double
            required: true
          rids:
            type: array
            items:
              type: export_review_skeleton_segment_node_review
            required: true
          sup:
            type: array
            items:
              type: export_review_skeleton_segment_node_sup
            required: true
      export_review_skeleton_segment_node_review:
        id: export_review_skeleton_segment_node_review
        properties:
        - description: Reviewer ID
          type: integer
          required: true
        - description: Review timestamp
          type: string
          format: date-time
          required: true
      export_review_skeleton_segment_node_sup:
        id: export_review_skeleton_segment_node_sup
        properties:
        - description: |
            Stack orientation to determine which axis is the coordinate of the
            plane where virtual nodes are suppressed. 0 for z, 1 for y, 2 for x.
          required: true
          type: integer
        - description: |
            Coordinate along the edge from this node to its parent where
            virtual nodes are suppressed.
          required: true
          type: number
          format: double
    type:
    - type: array
      items:
        type: export_review_skeleton_segment
      required: true
    """
    try:
        subarbor_node_id = int(request.POST.get('subarbor_node_id', ''))
    except ValueError:
        subarbor_node_id = None

    segments = _export_review_skeleton(project_id, skeleton_id, subarbor_node_id)
    return HttpResponse(json.dumps(segments, cls=DjangoJSONEncoder),
            content_type='application/json')

@requires_user_role(UserRole.Browse)
def skeleton_connectors_by_partner(request, project_id):
    """ Return a dict of requested skeleton vs relation vs partner skeleton vs list of connectors.
    Connectors lacking a skeleton partner will of course not be included. """
    skeleton_ids = set(int(v) for k,v in request.POST.iteritems() if k.startswith('skids['))
    cursor = connection.cursor()

    relations = get_relation_to_id_map(project_id, ('presynaptic_to', 'postsynaptic_to'), cursor)
    pre = relations['presynaptic_to']
    post = relations['postsynaptic_to']

    cursor.execute('''
    SELECT tc1.skeleton_id, tc1.relation_id,
           tc2.skeleton_id, tc1.connector_id
    FROM treenode_connector tc1,
         treenode_connector tc2
    WHERE tc1.skeleton_id IN (%s)
      AND tc1.connector_id = tc2.connector_id
      AND tc1.skeleton_id != tc2.skeleton_id
      AND tc1.relation_id != tc2.relation_id
      AND (tc1.relation_id = %s OR tc1.relation_id = %s)
      AND (tc2.relation_id = %s OR tc2.relation_id = %s)
    ''' % (','.join(map(str, skeleton_ids)), pre, post, pre, post))

    # Dict of skeleton vs relation vs skeleton vs list of connectors
    partners = defaultdict(partial(defaultdict, partial(defaultdict, list)))

    for row in cursor.fetchall():
        relation_name = 'presynaptic_to' if row[1] == pre else 'postsynaptic_to'
        partners[row[0]][relation_name][row[2]].append(row[3])

    return HttpResponse(json.dumps(partners))


@requires_user_role(UserRole.Browse)
def export_skeleton_reviews(request, project_id=None, skeleton_id=None):
    """ Return a map of treenode ID vs list of reviewer IDs,
    without including any unreviewed treenode. """
    m = defaultdict(list)
    for row in Review.objects.filter(skeleton_id=int(skeleton_id)).values_list('treenode_id', 'reviewer_id', 'review_time').iterator():
        m[row[0]].append(row[1:3])

    return HttpResponse(json.dumps(m, separators=(',', ':'), cls=DjangoJSONEncoder))

@requires_user_role(UserRole.Browse)
def partners_by_connector(request, project_id=None):
    """ Return a list of skeleton IDs related to the given list of connector IDs of the given skeleton ID.
    Will optionally filter for only presynaptic (relation=0) or only postsynaptic (relation=1). """
    skid = request.POST.get('skid', None)
    if not skid:
        raise Exception("Need a reference skeleton ID!")
    skid = int(skid)
    connectors = tuple(int(v) for k,v in request.POST.iteritems() if k.startswith('connectors['))
    rel_type = int(request.POST.get("relation", 0))
    size_mode = int(request.POST.get("size_mode", 0))

    query = '''
SELECT DISTINCT tc2.skeleton_id
FROM treenode_connector tc1,
     treenode_connector tc2
WHERE tc1.project_id = %s
  AND tc1.skeleton_id = %s
  AND tc1.connector_id = tc2.connector_id
  AND tc1.skeleton_id != tc2.skeleton_id
  AND tc1.relation_id != tc2.relation_id
  AND tc1.connector_id IN (%s)
''' % (project_id, skid, ",".join(str(x) for x in connectors))

    # Constrain the relation of the second part
    if 0 == rel_type or 1 == rel_type:
        query += "AND tc2.relation_id = (SELECT id FROM relation WHERE project_id = %s AND relation_name = '%s')" % (project_id, 'presynaptic_to' if 1 == rel_type else 'postsynaptic_to')

    cursor = connection.cursor()
    cursor.execute(query)

    if 0 == size_mode or 1 == size_mode:
        # Filter by size: only those with more than one treenode or with exactly one
        cursor.execute('''
SELECT skeleton_id
FROM treenode
WHERE skeleton_id IN (%s)
GROUP BY skeleton_id
HAVING count(*) %s 1
''' % (",".join(str(row[0]) for row in cursor.fetchall()), ">" if 0 == size_mode else "="))

    return HttpResponse(json.dumps(tuple(row[0] for row in cursor.fetchall())))