import numpy as np
""" Parsing options """
args = argparse.ArgumentParser()
parser = Parser(args)
opt = parser.make_options()
""" Load dataset """
data_path = '{}/{}'.format(opt.data_path, 'hico')
image_path = '{}/{}/{}'.format(opt.data_path, 'hico', 'images')
cand_dir = '{}/{}/{}'.format(opt.data_path, 'hico', 'detections')
dset = Hico(data_path, \
image_path, \
opt.test_split, \
cand_dir=cand_dir,\
thresh_file=opt.thresh_file, \
add_gt=False, \
train_mode=False, \
jittering=False, \
nms_thresh=opt.nms_thresh)
""" Load the test triplets """
target_triplets = dset.get_zeroshottriplets(
) # uncomment to eval zeroshot triplets
#target_triplets = dset.visualphrases.words() # uncomment to eval all triplets
""" Keys to analyze """
keys = ['s-sro-o', 's-r-o-sro']
""" Aggregate csv result files (from official HICO eval code) """
# Logger path
logger_path = osp.join(opt.logger_dir, opt.exp_name)
detection_path = parser.get_res_dir(opt, 'detections_' + opt.embedding_type)
import csv
from utils import Parser
""" Parsing options """
args = argparse.ArgumentParser()
parser = Parser(args)
opt = parser.make_options()
""" Load dataset """
data_path = '{}/{}'.format(opt.data_path, 'hico')
image_path = '{}/{}/{}'.format(opt.data_path, 'hico', 'images')
cand_dir = '{}/{}/{}'.format(opt.data_path, 'hico', 'detections')
dset = Hico(data_path, \
image_path, \
opt.test_split, \
cand_dir=cand_dir,\
thresh_file=opt.thresh_file, \
add_gt=False, \
train_mode=False, \
jittering=False, \
nms_thresh=opt.nms_thresh)
""" Key types """
keys = ['s-r-o', 's-sro-o', 's-r-o-sro']
""" Load the test triplets """
target_triplets = dset.get_zeroshottriplets()
subset = 'zeroshottriplet'
for key in keys:
""" Load ap results for all triplets """
filename_in = osp.join(opt.logger_dir, opt.exp_name, 'results_{}_{}_{}_{}_def.csv'.format(\
opt.cand_test,\
opt.test_split,\
store_ram = []
store_ram.append(
'objectscores') if opt.use_ram and opt.use_precompobjectscore else None
store_ram.append(
'appearance') if opt.use_ram and opt.use_precompappearance else None
data_path = '{}/{}'.format(opt.data_path, opt.data_name)
image_path = '{}/{}/{}'.format(opt.data_path, opt.data_name, 'images')
cand_dir = '{}/{}/{}'.format(opt.data_path, opt.data_name, 'detections')
dset = Hico(data_path, \
image_path, \
opt.test_split, \
cand_dir = cand_dir,\
thresh_file = opt.thresh_file,\
use_gt = False, \
train_mode = False, \
jittering = False,\
nms_thresh = opt.nms_thresh, \
store_ram = store_ram, \
l2norm_input = opt.l2norm_input)
dset_loader = TestSampler(dset, \
use_image = opt.use_image, \
use_precompappearance = opt.use_precompappearance, \
use_precompobjectscore = opt.use_precompobjectscore)
loader = torch.utils.data.DataLoader(dset_loader,
batch_size=8,
shuffle=False,
elif data_name == 'hicoforcocoa':
from datasets.hico_api import Hico as Dataset
splits = ['trainval', 'test']
elif data_name == 'cocoa':
from datasets.cocoa_api import Cocoa as Dataset
splits = ['all']
data_path = osp.join(DATA_PATH, data_name)
image_path = osp.join(data_path, 'images')
cand_dir = osp.join(data_path, 'detections')
proposals = {}
for split in splits:
dataset = Dataset(data_path, image_path, split, cand_dir=cand_dir,\
thresh_file='', use_gt=False, add_gt=True, train_mode=False, jittering=False, store_ram=[])
for im_id in dataset.image_ids:
cand_boxes = dataset.get_boxes(im_id)
obj_id = dataset.get_obj_id(im_id)
proposals[im_id] = np.hstack((obj_id[:, None], cand_boxes))
assert len(
np.unique(obj_id)) == len(obj_id), 'Careful duplicate obj_id'
pickle.dump(proposals,
open(osp.join(cand_dir, split + '_proposals.pkl'), 'wb'))
if opt.use_precompobjectscore:
store_ram.append('objectscores')
if opt.use_precompappearance:
store_ram.append('appearance')
data_path = '{}/{}'.format(opt.data_path, '%s')
image_path = '{}/{}/{}'.format(opt.data_path, '%s', 'images')
cand_dir = '{}/{}/{}'.format(opt.data_path, '%s', 'detections')
dset_train = Hico( data_path % 'hicoforcocoa',\
image_path % 'hicoforcocoa', \
opt.train_split, \
cand_dir = cand_dir % 'hicoforcocoa',\
thresh_file = opt.thresh_file, \
use_gt = opt.use_gt, \
add_gt = opt.add_gt, \
train_mode = False, \
jittering = False, \
nms_thresh = opt.nms_thresh,\
store_ram = [],\
l2norm_input = opt.l2norm_input,\
neg_GT = opt.neg_GT)
dset_test = Cocoa( data_path % 'cocoa',\
image_path % 'cocoa', \
'all', \
cand_dir = cand_dir % 'cocoa',\
thresh_file = opt.thresh_file, \
use_gt = False, \
add_gt = False, \
import os
import __init__
import numpy as np
import os.path as osp
from datasets.hico_api import Hico
import csv
import pickle
# Load vocabulary of triplets
root_path = './data'
data_path = '{}/{}'.format(root_path, 'hico')
image_path = '{}/{}/{}'.format(root_path, 'hico', 'images')
cand_dir = '{}/{}/{}'.format(root_path, 'hico', 'detections')
split = 'trainval' #'train','trainval'
dset = Hico(data_path, image_path, split, cand_dir)
# Get set of triplets
triplets_remove = dset.get_zeroshottriplets()
triplet_cat_remove = []
for l in range(len(triplets_remove)):
triplet_cat_remove.append(dset.visualphrases.word2idx[triplets_remove[l]])
# Build a new set of candidates excluding the triplet categories
cand_positives = pickle.load(open(osp.join(data_path, 'cand_positives_' + split + '.pkl'),'rb'))
idx_keep = []
for j in range(cand_positives.shape[0]):
if j%100000==0:
print('Done {}/{}'.format(j, cand_positives.shape[0]))