-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_hallucinator.py
160 lines (127 loc) · 5.47 KB
/
run_hallucinator.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
import paths
import torch
import base64
from pprint import pprint
from tqdm import tqdm
import os
import sys
import numpy as np
from refer360_utils import get_object_dictionaries
from refer360_utils import load_butd, load_vectors, DIR2IDX
import csv
DIRECTIONS = {
'navigation': ['ul', 'u', 'ur', 'l', 'r', 'dl', 'd', 'dr'],
'canonical': ['up', 'down', 'left', 'right'],
'cartesian': ['vertical', 'horizontal'],
'lup': ['lateral', 'up', 'down'],
'canonical_proximity': ['close_up', 'close_down', 'close_left', 'close_right',
'far_up', 'far_down', 'far_left', 'far_right']
}
file_path = os.path.dirname(__file__)
module_path = os.path.abspath(os.path.join(file_path))
sys.path.append(module_path)
module_path = os.path.abspath(os.path.join(
file_path, '..', '..', 'build_refer360'))
sys.path.append(module_path)
csv.field_size_limit(sys.maxsize)
DEVICE = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
def dump_predictions(model, nodes, features_prefix, fov_prefix, object_classes):
n_objects = len(object_classes)-1
all_preds = np.zeros((len(nodes), n_objects, len(DIR2IDX)))
for jj, n in enumerate(sorted(nodes.keys())):
idx = nodes[n]['idx']
fov_file = fov_prefix + '.{}.jpg'.format(idx)
img_feat_path = fov_file.replace('.jpg', '.feat.npy')
preds = np.zeros((n_objects, len(DIR2IDX)))
if os.path.exists(img_feat_path):
img_features = np.load(img_feat_path)
image = torch.FloatTensor(img_features).to(DEVICE)
img_batch = image.unsqueeze(0).repeat(n_objects, 1, 1, 1)
query = []
for obj_id, obj_name in enumerate(object_classes[:n_objects]):
query += [obj_id]
latitude = torch.FloatTensor([nodes[n]['lat']]).to(
DEVICE).unsqueeze(0).repeat(n_objects, 1)
longitude = torch.FloatTensor([nodes[n]['lng']]).to(
DEVICE).unsqueeze(0).repeat(n_objects, 1)
query = torch.LongTensor([query]).to(DEVICE).transpose(1, 0)
out = model({
'latitude': latitude,
'longitude': longitude,
'im_batch': img_batch,
'queries': query,
})
preds = out['action_logits'].cpu().detach().numpy() # 359 x 8
all_preds[idx, :, :] = preds
return all_preds
def run_hallucination(
cache_root='/projects2/refer360test/data/cached_data_30degrees',
butd_filename='./img_features/refer360_30degrees_obj36.tsv',
image_list_file='./refer360_data/imagelist.txt',
n_fovs=60,
word_embedding_path='./tasks/FAST/data/cc.en.300.vec',
obj_dict_file='./tasks/FAST/data/vg_object_dictionaries.all.json',
model_path='/projects2/refer360test/src/exp-PRIOR/default/model.best.pt',
outfile='img_features/refer360_30degrees_hallucinated_v4.tsv'):
image_list = [line.strip()
for line in open(image_list_file)]
print('\nloading model from', model_path)
model = torch.load(model_path)
model.eval()
meta_file = os.path.join(cache_root, 'meta.npy')
meta = np.load(meta_file, allow_pickle=True)[()]
nodes = meta['nodes']
vg2idx, idx2vg, obj_classes, name2vg, name2idx, vg2name = get_object_dictionaries(
obj_dict_file, return_all=True)
n_objects = len(vg2name)-1
print('# of objects (except unk)', n_objects)
print('loading w2v...', word_embedding_path)
w2v = load_vectors(word_embedding_path, name2vg)
print('loading BUTD boxes...', butd_filename)
fov2keys = load_butd(butd_filename,
vg2name=vg2name,
keys=['boxes', 'objects_id'])
print('loaded BUTD boxes!', image_list_file)
fov2keys = {'boxes': {}, 'objects_id': {}}
FIELDNAMES = ['pano_fov', 'features']
print('Generating hallucinated object directions')
pbar = tqdm(image_list)
with open(outfile, 'w') as tsvfile:
writer = csv.DictWriter(tsvfile, delimiter='\t', fieldnames=FIELDNAMES)
for fname in pbar:
pano = fname.split('/')[-1].split('.')[0]
features_prefix = os.path.join(
cache_root, 'features', '{}'.format(pano))
fov_prefix = os.path.join(
cache_root, 'fovs', '{}'.format(pano))
preds = dump_predictions(
model, nodes, features_prefix, fov_prefix, obj_classes) # n_fovs x n_obj x dir
for idx in range(n_fovs):
pano_fov = '{}_{}'.format(pano, idx)
features = np.zeros((9, 300), dtype=np.float32)
if pano_fov in fov2keys['boxes'] and pano_fov in fov2keys['objects_id']:
boxes = fov2keys['boxes'][pano_fov]
object_ids = fov2keys['objects_id'][pano_fov]
n_boxes = len(boxes)
emb_feats = np.zeros((n_boxes, 300), dtype=np.float32)
for ii, obj_id in enumerate(object_ids):
obj_name = vg2name.get(obj_id, '</s>')
emb_feats[ii, :] = w2v.get(obj_name, w2v['</s>'])
features[4, :] = np.sum(emb_feats, axis=0)
for direction in DIR2IDX:
feat_index = DIR2IDX[direction]
pred_index = DIRECTIONS['navigation'].index(direction)
dir_feats = np.zeros((1, 300), dtype=np.float32)
for co_idx in range(n_objects):
w = preds[idx, co_idx, pred_index]
if w > 0:
co_name = obj_classes[co_idx]
emb = w2v.get(co_name, w2v['</s>'])
dir_feats += emb * w
features[feat_index, :] = dir_feats
encoded = base64.b64encode(features).decode()
d = {'pano_fov': pano_fov,
'features': encoded}
writer.writerow(d)
if __name__ == '__main__':
run_hallucination()