def do_eval(evaluator, model, vcoco_set, save_dir, postfix=""):
results = evaluator.evaluate_model(model, "vcoco_val", ld.COCO_IMGDIR)
outfile = os.path.join(save_dir, "evaluation%s.pkl" % postfix)
with open(outfile, "w") as f:
pik.dump(results, f)
vcocoeval = VCOCOeval(ld.get_vsrl_labels(vcoco_set), ld.COCO_VCOCO_ANN,
ld.get_ids(vcoco_set))
vcocoeval._do_eval(outfile, ovr_thresh=0.5)
folder_name = '{}'.format(first_word)
from vsrl_eval import VCOCOeval
#vsrl_annot_file='/media/data/iftekhar/v-coco/data/vcoco/vcoco_val.json'
#split_file='/media/data/iftekhar/v-coco/data/splits/vcoco_val.ids'
if flag == 'train':
vsrl_annot_file_s = path + '/data/vcoco/vcoco_train.json'
split_file_s = path + '/data/splits/vcoco_train.ids'
elif flag == 'test':
vsrl_annot_file_s = path + '/data/vcoco/vcoco_test.json'
split_file_s = path + '/data/splits/vcoco_test.ids'
elif flag == 'val':
vsrl_annot_file_s = path + '/data/vcoco/vcoco_val.json'
split_file_s = path + '/data/splits/vcoco_val.ids'
coco_file_s = path + '/data/instances_vcoco_all_2014.json'
vcocoeval = VCOCOeval(vsrl_annot_file_s, coco_file_s, split_file_s)
#try:
# file_name='/home/iftekhar/object_interact/codes_object_interact/'+folder_name+'/'+'{}{}.pickle'.format(flag,saving_epoch)
#except:
file_name = '../' + folder_name + '/' + '{}{}.pickle'.format(
flag, saving_epoch)
print(file_name)
with open(file_name, 'rb') as handle:
b = pickle.load(handle)
print(len(b))
vcocoeval._do_eval(b, ovr_thresh=0.5)
#print(b[1]['person_box'])
human_thres = 0.3
object_thres = 0.1
action_thres = 0.05
np.random.seed(3)
detection = []
DATA_DIR = '/home/wangtiancai/data/vcoco'
with open(DATA_DIR + '/' + 'prior_mask.pkl', 'rb') as f:
prior_mask = pickle.load(f, encoding='latin1')
with open(DATA_DIR + '/' + 'Test_Faster_RCNN_R-50-PFN_2x_VCOCO.pkl',
'rb') as f:
Test_RCNN = pickle.load(f, encoding='latin1')
Action_dic = json.load(open(DATA_DIR + '/' + 'action_index.json'))
Action_dic_inv = {y: x for x, y in Action_dic.items()}
ROOT_DIR = '/home/wangtiancai/data/vcoco/'
output_file = ROOT_DIR + '/Results/' + 'SS' + '_' + 'HOI' + '.pkl'
vcocoeval = VCOCOeval(DATA_DIR + '/' + 'vcoco_test.json',
DATA_DIR + '/' + 'instances_vcoco_all_2014.json',
DATA_DIR + '/' + 'vcoco_test.ids')
test(opt, Test_RCNN, prior_mask, Action_dic_inv, output_file, human_thres,
object_thres, action_thres, detection)
vcocoeval._do_eval(output_file, ovr_thresh=0.5)
model = LIGHTEN_image(spatial_subnet=cfg.SPATIAL_SUBNET, drop=0.0)
res50 = torch.hub.load('pytorch/vision:v0.5.0', 'resnet50', pretrained=True)
res50 = torch.nn.Sequential(*(list(res50.children())[:-1]))
res50.cuda()
res50.eval()
sigmoid = nn.Sigmoid()
model.float().cuda()
model.eval()
# set to true if test scores are already computed
do_eval_only = False
if do_eval_only:
vcocoeval = VCOCOeval(
os.path.join(cfg.VCOCO_EVAL_DIR, 'vcoco/vcoco_test.json'),
os.path.join(cfg.VCOCO_EVAL_DIR, 'instances_vcoco_all_2014.json'),
os.path.join(cfg.VCOCO_EVAL_DIR, 'splits/vcoco_test.ids'))
vcocoeval._do_eval(save_path, ovr_thresh=0.5)
print(save_path)
exit(1)
experiment_name = cfg.SPATIAL_SUBNET
if cfg.TRAIN_VAL:
experiment_name += '_trainVal'
checkpoint_file = os.path.join(cfg.CHECKPOINT_DIR,
'checkpoint_' + experiment_name + '.pth')
if os.path.exists(checkpoint_file):
checkpoint = torch.load(checkpoint_file)
model.load_state_dict(checkpoint['state_dict'])
print('Loaded model from checkpoint')
action_kind][rn]
if new_pred_dict[keytwo][4] > score:
#print new_pred_dict[keytwo], obj_box
continue
new_pred_dict[keytwo] = obj_box
elif score > 0.1:
print action_dict[action_kind], score
new_pred_dict['person_box'] = person_box
new_pred_dict[keyone] = score
new_pred_dict['image_id'] = imgid
if not found:
prediction_list.append(new_pred_dict)
pickle_out = open("detections.pkl", "wb")
pickle.dump(prediction_list, pickle_out)
pickle_out.close()
from vsrl_eval import VCOCOeval
vsrl_annot_file = 'data/vcoco/vcoco_val.json'
coco_file = 'data/instances_vcoco_all_2017.json'
split_file = 'data/splits/vcoco_val.ids'
det_file = 'detections.pkl'
vcocoeval = VCOCOeval(vsrl_annot_file, coco_file, split_file)
fp, fn = vcocoeval._do_eval(det_file, ovr_thresh=0.5)
pickle_out = open("fp_fn_samples.pkl", "wb")
pickle.dump([fp, fn], pickle_out)
pickle_out.close()