Skip to content
Snippets Groups Projects
pascalpart.py 6.25 KiB
Newer Older
Ivan Donadello's avatar
Ivan Donadello committed
import logictensornetworks as ltn
import tensorflow as tf
import numpy as np
import csv, pdb
import timeit

ltn.default_layers = 2
ltn.default_smooth_factor = 1e-15
ltn.default_tnorm = "luk"
ltn.default_aggregator = "hmean"
ltn.default_positive_fact_penality = 0.
ltn.default_clauses_aggregator = "hmean"

data_training_dir = "data/training/"
data_testing_dir = "data/testing/"
zero_distance_threshold = 6
number_of_features = 65

types = np.genfromtxt("data/classes.csv", dtype="S", delimiter=",")

# uncomment this line for training the vehicle object types
#selected_types = np.array(['aeroplane','artifact_wing','body','engine','stern','wheel','bicycle','chain_wheel','handlebar','headlight','saddle','bus','bodywork','door','license_plate','mirror','window','car','motorbike','train','coach','locomotive','boat'])

# uncomment this line for training the indoor object types
selected_types = np.array(['bottle','body','cap','pottedplant','plant','pot','tvmonitor','screen','chair','sofa','diningtable'])

# uncomment this line for training the animal object types
#selected_types = np.array(['person','arm','ear','ebrow','foot','hair','hand','mouth','nose','eye','head','leg','neck','torso','cat','tail','bird','animal_wing','beak','sheep','horn','muzzle','cow','dog','horse','hoof'])

# uncomment this line for training all the object types
#selected_types = types[1:]

objects = ltn.Domain(number_of_features-1,label="a_bounding_box")

pairs_of_objects = ltn.Domain(2*(number_of_features-1)+2,label="a_pair_of_bounding_boxes")

isOfType = {}
for t in selected_types:
    isOfType[t] = ltn.Predicate("is_of_type_"+t,objects,layers=5)
isPartOf = ltn.Predicate("is_part_of",pairs_of_objects)

objects_of_type = {}
objects_of_type_not = {}
for t in selected_types:
    objects_of_type[t] = ltn.Domain(number_of_features-1,label="objects_of_type_"+t)
    objects_of_type_not[t] = ltn.Domain(number_of_features-1,label="objects_of_type_not_" + t)

object_pairs_in_partOf = ltn.Domain((number_of_features-1) * 2 + 2,
                                    label="object_pairs_in_partof_relation")
object_pairs_not_in_partOf = ltn.Domain((number_of_features-1) * 2 + 2,
                                        label="object_pairs_not_in_partof_relation")

def containment_ratios_between_two_bbxes(bb1, bb2):
    bb1_area = (bb1[-2] - bb1[-4]) * (bb1[-1] - bb1[-3])
    bb2_area = (bb2[-2] - bb2[-4]) * (bb2[-1] - bb2[-3])
    w_intersec = max(0,min([bb1[-2], bb2[-2]]) - max([bb1[-4], bb2[-4]]))
    h_intersec = max(0,min([bb1[-1], bb2[-1]]) - max([bb1[-3], bb2[-3]]))
    bb_area_intersection = w_intersec * h_intersec
    return [float(bb_area_intersection)/bb1_area, float(bb_area_intersection)/bb2_area]

def get_data(train_or_test_swritch,max_rows=10000000):
    assert train_or_test_swritch == "train" or train_or_test_swritch == "test"

    # Fetching the data from the file system

    if train_or_test_swritch == "train":
        data_dir = data_training_dir
    if train_or_test_swritch == "test":
        data_dir = data_testing_dir
    data = np.genfromtxt(data_dir+"features.csv",delimiter=",",max_rows=max_rows)
    types_of_data = types[np.genfromtxt(data_dir + "types.csv", dtype="i", max_rows=max_rows)]
    idx_whole_for_data = np.genfromtxt(data_dir+ "partOf.csv",dtype="i",max_rows=max_rows)
    idx_of_cleaned_data = np.where(np.logical_and(
        np.all(data[:, -2:] - data[:, -4:-2] >= zero_distance_threshold, axis=1),
        np.in1d(types_of_data,selected_types)))[0]
    print "deleting", len(data) - len(idx_of_cleaned_data), "small bb out of", data.shape[0], "bb"
    data = data[idx_of_cleaned_data]
    data[:, -4:] /= 500

    # Cleaning data by removing small bounding boxes and recomputing indexes of partof data

    types_of_data = types_of_data[idx_of_cleaned_data]
    idx_whole_for_data = idx_whole_for_data[idx_of_cleaned_data]
    for i in range(len(idx_whole_for_data)):
        if idx_whole_for_data[i] != -1 and idx_whole_for_data[i] in idx_of_cleaned_data:
            idx_whole_for_data[i] = np.where(idx_whole_for_data[i] == idx_of_cleaned_data)[0]
        else:
            idx_whole_for_data[i] = -1

    # Grouping bbs that belong to the same picture

    pics = {}
    for i in range(len(data)):
        if data[i][0] in pics:
            pics[data[i][0]].append(i)
        else:
            pics[data[i][0]] = [i]

    pairs_of_data = np.array(
        [np.concatenate((data[i][1:], data[j][1:], containment_ratios_between_two_bbxes(data[i], data[j]))) for p in
         pics for i in pics[p] for j in pics[p]])

    pairs_of_bb_idxs = np.array([(i,j) for p in pics for i in pics[p] for j in pics[p]])

    partOf_of_pair_of_data = np.array([idx_whole_for_data[i] == j for p in pics for i in pics[p] for j in pics[p]])

    return data, pairs_of_data, types_of_data, partOf_of_pair_of_data, pairs_of_bb_idxs, pics

def get_part_whole_ontology():
    with open('data/pascalPartOntology.csv') as f:
        ontologyReader = csv.reader(f)
        parts_of_whole = {}
        wholes_of_part = {}
        for row in ontologyReader:
            parts_of_whole[row[0]] = row[1:]
            for t in row[1:]:
                if t in wholes_of_part:
                    wholes_of_part[t].append(row[0])
                else:
                    wholes_of_part[t] = [row[0]]
        for whole in parts_of_whole:
            wholes_of_part[whole] = []
        for part in wholes_of_part:
            if part not in parts_of_whole:
                parts_of_whole[part] = []
    selected_parts_of_whole = {}
    selected_wholes_of_part = {}
    for t in selected_types:
        selected_parts_of_whole[t] = [p for p in parts_of_whole[t] if p in selected_types]
        selected_wholes_of_part[t] = [w for w in wholes_of_part[t] if w in selected_types]
    return selected_parts_of_whole, selected_wholes_of_part

# reporting measures

def precision(conf_matrix, prediction_array=None):
    if prediction_array is not None:
        return conf_matrix.diagonal()/prediction_array
    else:
        return conf_matrix.diagonal() / conf_matrix.sum(1).T

def recall(conf_matrix,gold_array=None):
    if gold_array is not None:
        return conf_matrix.diagonal() / gold_array
    else:
        return conf_matrix.diagonal() / conf_matrix.sum(0)

def f1(precision,recall):
    return np.multiply(2*precision,recall)/(precision + recall)

print "end of new pascalpart.py"