get_syn_pairs_with_skel_dists_and_apply_merge_model_parallel.py

import os
import sys

working_dir = os.path.dirname(__file__)
sys.path.insert(0, working_dir)
os.chdir(working_dir)

from numpy import mean
from scipy.spatial.distance import euclidean
from google.cloud import bigquery             
from google.oauth2 import service_account
from google.cloud import bigquery_storage            
from common_functions_h01 import get_skel_data_from_shard_dir, make_one_skel_graph_nx, get_skel_nodes_closest_to_synapses, get_nm_dist_along_skel_path
import time
import networkx as nx
from zipfile import ZipFile
import os
from multiprocessing import Pool
import pickle
import json
import numpy as np
import igraph as ig
import pandas as pd

model_name = 'synapse_merge_model_skel_only_20210412.pkl'
credentials_file = 'alexshapsoncoe.json' # or your credentials file
output_dir = f'{working_dir}/synapse_merging_output' # available from gs://h01_paper_public_files/synapse_merging_output
synapse_voxel_size = [8,8,33]
syn_db_name = 'goog14r0s5c3.synaptic_connections_ei_conserv_reorient_fix_ei'
skel_dir = '20200916c3_skeletons_6class_plus_myelin' # available from gs://h01_paper_public_files/20200916c3_skeletons_6class_plus_myelin
skel_voxel_size = [32,32,33]
skel_divisor = 42356404
cpu_num = 14

def do_one_skel_dir(site_pair_d, x):

    if f'{x}.json' in os.listdir(f'{output_dir}/skel_dists_temp/'):
        return

    start = time.time()
    print(f'Starting dir {x}')

    results = {}

    skel_path = skel_dir + '/' + str(x) + '.zip'

    if not os.path.exists(skel_path):

        for agglo_id in site_pair_d.keys():
            for syn_id1, centroid1, syn_id2, centroid2 in site_pair_d[agglo_id]:
                new_id = [syn_id1, syn_id2]
                new_id.sort()
                new_id = '_'.join(new_id)
                results[new_id] = 'no_skeleton'
        
        with open(f'{output_dir}/skel_dists_temp/{x}.json', 'w') as fp:
            json.dump(results, fp)

        return

    else:
        shard_dir = ZipFile(skel_path, 'r')

    this_batch_all_ids = set(site_pair_d.keys()) 

    if len(this_batch_all_ids) == 0:

        with open(f'{output_dir}/skel_dists_temp/{x}.json', 'w') as fp:
            json.dump(results, fp)

        return

    skel_data = get_skel_data_from_shard_dir(this_batch_all_ids, shard_dir)
    
    skel_not_found_count = 0
    skel_found_count = 0

    for neuron_id in site_pair_d.keys():

        if neuron_id not in skel_data:
            skel_not_found_count += 1
            for syn_id1, centroid1, syn_id2, centroid2 in site_pair_d[neuron_id]:
                new_id = [syn_id1, syn_id2]
                new_id.sort()
                new_id = '_'.join(new_id)
                results[new_id] = 'no_skeleton'
            continue
    
        skel_found_count += 1

        # Then make a graph of the segment in question:
        skel_g = make_one_skel_graph_nx(skel_data[neuron_id], skel_voxel_size, join_components = False)

        syn_locations = [] 
        syn_ids = []

        for syn_id1, centroid1, syn_id2, centroid2 in site_pair_d[neuron_id]:
            syn_ids.append(syn_id1)
            syn_locations.append(centroid1)
            syn_ids.append(syn_id2)
            syn_locations.append(centroid2)


        syn_node_lookup_batch_size = 1000

        chosen_nodes = []
        num_batches = int(len(syn_locations)/syn_node_lookup_batch_size)

        for batch in range(num_batches+1):

            syn_locs = syn_locations[batch*syn_node_lookup_batch_size:(batch+1)*syn_node_lookup_batch_size]
            if syn_locs == []: break
            chosen_node_batch = get_skel_nodes_closest_to_synapses(syn_locs, skel_g, list(skel_g.nodes()))
            chosen_nodes.extend(chosen_node_batch)

        assert len(chosen_nodes) == len(syn_ids)

        synid2node = {k: v for (k, v) in zip(syn_ids, chosen_nodes)}

        for syn_id1, centroid1, syn_id2, centroid2 in site_pair_d[neuron_id]:
            
            new_id = [syn_id1, syn_id2]
            new_id.sort()
            new_id = '_'.join(new_id)

            skel_node1 = synid2node[syn_id1]
            skel_node2 = synid2node[syn_id2]

            if skel_node1 == skel_node2: 
                results[new_id] = 0
                continue

            try:
                sp = nx.shortest_path(skel_g, source=skel_node1, target=skel_node2)

            except nx.exception.NetworkXNoPath:

                results[new_id] = 'no_skeleton'
                continue


            results[new_id] = get_nm_dist_along_skel_path(skel_g, sp)

    with open(f'{output_dir}/skel_dists_temp/{x}.json', 'w') as fp:
        json.dump(results, fp)

    print(f'Dir {x} took {time.time()-start}, found skeletons for {skel_found_count} / {skel_found_count+skel_not_found_count} agglo IDs')

def get_same_agglo_pairs(upper_threshold):

    print('Retrieving synapse pairs')
    credentials = service_account.Credentials.from_service_account_file(credentials_file)
    client = bigquery.Client(project=credentials.project_id, credentials=credentials)
    bqstorageclient = bigquery_storage.BigQueryReadClient(credentials=credentials)

    query = f"""WITH 
            all_edges AS (
                SELECT 
                    pre_synaptic_site.neuron_id AS pre_seg_id, 
                    post_synaptic_partner.neuron_id AS post_seg_id, 
                    COUNT(*) AS pair_count
                FROM {syn_db_name}
                GROUP BY pre_synaptic_site.neuron_id, post_synaptic_partner.neuron_id
                HAVING count(*) >= 2
                )

            SELECT 
                CAST(pre_synaptic_site.neuron_id as STRING) AS pre_agglo_id, 
                CAST(post_synaptic_partner.neuron_id as STRING) AS post_agglo_id, 
                CAST(pre_synaptic_site.id as STRING) AS pre_syn_id, 
                CAST(post_synaptic_partner.id as STRING) AS post_syn_id,
                CAST(post_synaptic_partner.centroid.x as STRING) AS post_centroid_x,
                CAST(post_synaptic_partner.centroid.y as STRING) AS post_centroid_y,
                CAST(post_synaptic_partner.centroid.z as STRING) AS post_centroid_z,
                CAST(pre_synaptic_site.centroid.x as STRING) AS pre_centroid_x,
                CAST(pre_synaptic_site.centroid.y as STRING) AS pre_centroid_y,
                CAST(pre_synaptic_site.centroid.z as STRING) AS pre_centroid_z,
                FROM {syn_db_name} AS all_syn
                INNER JOIN all_edges AS edge_info 
                    ON all_syn.pre_synaptic_site.neuron_id = edge_info.pre_seg_id AND all_syn.post_synaptic_partner.neuron_id = edge_info.post_seg_id
                WHERE pre_synaptic_site.neuron_id IS NOT NULL AND post_synaptic_partner.neuron_id IS NOT NULL
                """
    c = bigquery.job.QueryJobConfig(allow_large_results = True)
    df = client.query(query, job_config=c).result().to_dataframe(bqstorage_client=bqstorageclient) #, progress_bar_type='tqdm_gui')
    print('Retrieved synapse pairs')
        
    all_syn_seg_data = {}

    for x in df.index:
                
        pre_agglo_id = df.at[x, 'pre_agglo_id']
        post_agglo_id = df.at[x, 'post_agglo_id']

        pre_syn_id = df.at[x, 'pre_syn_id']
        post_syn_id = df.at[x, 'post_syn_id']
        
        post_centroid = tuple([
            int(df.at[x, 'post_centroid_x'])*synapse_voxel_size[0],
            int(df.at[x, 'post_centroid_y'])*synapse_voxel_size[1],
            int(df.at[x, 'post_centroid_z'])*synapse_voxel_size[2],
        ])

        pre_centroid = tuple([
            int(df.at[x, 'pre_centroid_x'])*synapse_voxel_size[0],
            int(df.at[x, 'pre_centroid_y'])*synapse_voxel_size[1],
            int(df.at[x, 'pre_centroid_z'])*synapse_voxel_size[2],
        ])

        common_info = {
            'pre_centroid': pre_centroid, 
            'post_centroid': post_centroid, 
            'pre_syn_id': pre_syn_id, 
            'post_syn_id': post_syn_id,
        }
        
        if pre_agglo_id not in all_syn_seg_data:
            all_syn_seg_data[pre_agglo_id] = {}
        
        if post_agglo_id not in all_syn_seg_data[pre_agglo_id]:
            all_syn_seg_data[pre_agglo_id][post_agglo_id] = []

        all_syn_seg_data[pre_agglo_id][post_agglo_id].append(common_info) 


    # Get all pairs and their distances:
    skel_site_pairs = [{} for x in range(10000)]

    same_agglo_pairs = []

    for pre_agglo_id in all_syn_seg_data.keys():

        for post_agglo_id in all_syn_seg_data[pre_agglo_id].keys():

            pair_synapses = all_syn_seg_data[pre_agglo_id][post_agglo_id]

            assert len(pair_synapses) > 1

            # Identify skeletons to obtain distances for:

            for syn1_pos, syn1 in enumerate(pair_synapses):

                for syn2 in pair_synapses[syn1_pos+1:]:

                    syn1_centre = mean([syn1['pre_centroid'], syn1['post_centroid']], axis=0)
                    syn2_centre = mean([syn2['pre_centroid'], syn2['post_centroid']], axis=0)
                    dist_nm = float(euclidean(syn1_centre, syn2_centre))

                    this_p = {  
                        'pre_agglo_id': pre_agglo_id,
                        'post_agglo_id': post_agglo_id,
                        'synapse_1': syn1,
                        'synapse_2': syn2,
                        'dist_nm':  dist_nm,
                            }

                    same_agglo_pairs.append(this_p)

                    if dist_nm < upper_threshold and dist_nm > lower_threshold:

                        for dtype in ['pre', 'post']:

                            agglo_id = this_p[f'{dtype}_agglo_id']

                            idx = int(int(agglo_id)/skel_divisor)
                            centroid1 = syn1[f'{dtype}_centroid']
                            syn_id1 = syn1[f'{dtype}_syn_id']
                            centroid2 = syn2[f'{dtype}_centroid']
                            syn_id2 = syn2[f'{dtype}_syn_id']

                            if agglo_id not in skel_site_pairs[idx]:
                                skel_site_pairs[idx][agglo_id] = []

                            skel_site_pairs[idx][agglo_id].append((syn_id1, centroid1, syn_id2, centroid2))

    print('Found '+ str(len(same_agglo_pairs)) + 'same seg pairs')

    with open(f'{output_dir}/same_agglo_pairs.json', 'w') as fp:
        json.dump(same_agglo_pairs, fp)

    with open(f'{output_dir}/skel_site_pairs.json', 'w') as fp:
        json.dump(skel_site_pairs, fp)


if __name__ == '__main__':

    if not os.path.exists(output_dir):
        os.mkdir(output_dir)

    if 'skel_dists_temp' not in os.listdir(output_dir):
        os.mkdir(f'{output_dir}/skel_dists_temp')

    if 'final_join_decisions.json' not in os.listdir(output_dir):

        with open(model_name, 'rb') as fp:
            merge_model = pickle.load(fp)

        lower_threshold = merge_model.lower_threshold
        upper_threshold = merge_model.upper_threshold

        if not (('same_agglo_pairs.json' in os.listdir(output_dir)) and ('skel_site_pairs.json' in os.listdir(output_dir))):

            get_same_agglo_pairs(upper_threshold)

        if len(os.listdir(f'{output_dir}/skel_dists_temp/')) < 10000:

            print('Loading skel_site_pairs')
            with open(f'{output_dir}/skel_site_pairs.json', 'r') as fp:
                skel_site_pairs = json.load(fp)
            print('Loaded skel_site_pairs')
                

            # Obtain skeleton distances:
            pool = Pool(cpu_num)
            pool.starmap(do_one_skel_dir, zip(skel_site_pairs, range(10000)))
            pool.close()
            pool.join()
                    

            del skel_site_pairs

        synpair2dist = {}

        for x in range(10000):

            with open(f'{output_dir}/skel_dists_temp/{x}.json', 'r') as fp:
                shard_dir = json.load(fp)
            
            synpair2dist.update(shard_dir)


        all_syn_combined_ids = {'pre': [], 'post': []}

        with open(f'{output_dir}/same_agglo_pairs.json', 'r') as fp:
            same_agglo_pairs = json.load(fp)

        for p in same_agglo_pairs:

            for dtype in ['pre', 'post']:

                combo_id = [p['synapse_1'][f'{dtype}_syn_id'], p['synapse_2'][f'{dtype}_syn_id']]
                combo_id.sort()
                combo_id = '_'.join(combo_id)

                all_syn_combined_ids[dtype].append(combo_id)
        
        if len(synpair2dist.keys()) > 1:
            ave_pre_skel_dist = np.mean([synpair2dist[x] for x in all_syn_combined_ids['pre'] & synpair2dist.keys() if synpair2dist[x]!= 'no_skeleton'])
            ave_post_skel_dist = np.mean([synpair2dist[x] for x in all_syn_combined_ids['post'] & synpair2dist.keys() if synpair2dist[x]!= 'no_skeleton'])

        # Use model to make final decison:
        if 'final_decisions_not_organized.json' not in os.listdir(output_dir):
            print('making final decisons')
            to_merge = []

            no_skel_data_count = {'pre': 0, 'post': 0}

            for pos, pair in enumerate(same_agglo_pairs):

                if pos%10000 == 0:
                    print(f'Completed {pos} pairs out of {len(same_agglo_pairs)} in total, {len(to_merge)} merged so far')

                syn1_id = pair['synapse_1']['pre_syn_id'] + '_' + pair['synapse_1']['post_syn_id']
                syn2_id = pair['synapse_2']['pre_syn_id'] + '_' + pair['synapse_2']['post_syn_id']
                pair_id = [syn1_id, syn2_id]
                pair_id.sort()
                pair_id = tuple(pair_id)

                dist = pair['dist_nm']

                if dist >= upper_threshold:
                    # final_decisions.append((pair_id[0]))
                    # final_decisions.append((pair_id[1]))
                    continue

                if dist <= lower_threshold:
                    to_merge.append(pair_id)
                    continue

                # If not decided by distance thresholds, use skeleton distance:
                skel_dists = {'pre': None, 'post': None}

                for dtype in ['pre', 'post']:

                    combo_id = [pair['synapse_1'][f'{dtype}_syn_id'], pair['synapse_2'][f'{dtype}_syn_id']]
                    combo_id.sort()
                    combo_id = '_'.join(combo_id)

                    if combo_id in synpair2dist:
                        skel_dist = synpair2dist[combo_id]
                    else:
                        skel_dist = 'no_skeleton'

                    if skel_dist == 'no_skeleton':
                        no_skel_data_count[dtype] +=1
                        skel_dists[dtype] = ave_pre_skel_dist
                    else:
                        skel_dists[dtype] = skel_dist

                predictors = [[max(skel_dists['pre'], skel_dists['post'])]]
            
                decision = int(merge_model.predict(predictors))
        
                if decision == 1:
                    to_merge.append(pair_id)

            with open(f'{output_dir}/final_join_decisions.json', 'w') as fp:
                json.dump(to_merge, fp)
            
            print('Made all decisions')
            del same_agglo_pairs

    with open(f'{output_dir}/final_join_decisions.json', 'r') as fp:
        all_edges = json.load(fp)

    # Get tables of synapses to discard:
    g = ig.Graph()

    #all_edges = [(f"{x['pre1']}_{x['post1']}", f"{x['pre2']}_{x['post2']}") for x in final_decisions]
    all_vertices = [a for b in all_edges for a in b]

    print(len(all_vertices), 'synapses in pairs')
    
    g.add_vertices(all_vertices)
    g.add_edges(all_edges)

    discarded_synapses = []

    for synapse_collection in g.components(mode='WEAK'):

        for synapse in synapse_collection[1:]:
            discarded_synapses.append(g.vs[synapse]['name'].split('_'))
    
    print(len(discarded_synapses), 'synapses discarded')
    
    df = pd.DataFrame(discarded_synapses, columns=['pre', 'post'])
        
    df.to_csv(f'{output_dir}/synapses_to_discard.csv', index=0)