import pickle
import numpy as np
dataset = 'vrd'
if dataset == 'vrd':
from open_relation.dataset.vrd.label_hier.obj_hier import objnet
else:
from open_relation.dataset.vg.label_hier.obj_hier import objnet
# prepare label maps
org2wn = objnet.raw2wn()
label2index = objnet.label2index()
index2label = objnet.index2label()
# evaluation result
e_p_path = 'e_p.bin'
e_p_list = pickle.load(open(e_p_path, 'rb'))
# cal acc for each class
e_freq = dict()
e_acc = dict()
# prediction matrix : expected -> prediction
mat2org = [label2index[i] for i in org2wn.keys()]
org2mat = dict()
for i, org in enumerate(mat2org):
org2mat[org] = i
e_p_mat = np.zeros((len(mat2org), len(mat2org)))
for e_p in e_p_list:
if e_p[0] not in e_acc:
def eval4(label_vecs, label2index, label1, label2):
ind1 = label2index[label1]
ind2= label2index[label2]
vec1 = label_vecs[ind1]
vec2 = label_vecs[ind2]
sub = vec1 - vec2
sub[sub < 0] = 0
sub_square = sub * sub
E = np.sqrt(np.sum(sub_square))
print('P( %s , %s ) = %.2f' % (label1, label2, E))
if __name__ == '__main__':
# label vectors
weight_path = data_config.extra_config[target].config['label_vec_path']
label_vec_file = h5py.File(weight_path, 'r')
label_vecs = np.array(label_vec_file['label_vec'])
# label mapping
label2index = classnet.label2index()
index2label = classnet.index2label()
raw2wn = classnet.raw2wn()
eval2(label_vecs, index2label, label2index, raw2wn)
# eval3(label_vecs, index2label, label2index, raw2wn, 'jeans')
# eval4(label_vecs, label2index, 'shirt', 'garment.n.01')
# objnet.get_node_by_name('person').show_hyper_paths()