-
Notifications
You must be signed in to change notification settings - Fork 4
Expand file tree
/
Copy pathmaterial_proposal.py
More file actions
163 lines (125 loc) · 5.64 KB
/
Copy pathmaterial_proposal.py
File metadata and controls
163 lines (125 loc) · 5.64 KB
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
161
162
163
import os
import time
import json
import openai
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
from gpt_inference import gpt_candidate_materials, gpt_thickness, parse_material_list, \
parse_material_hardness, gpt4v_candidate_materials, parse_material_json
from utils import load_images, get_scenes_list
from arguments import get_args
from my_api_key import OPENAI_API_KEY
BASE_SEED = 100
def gpt_wrapper(gpt_fn, parse_fn, max_tries=10, sleep_time=3):
"""Wrap gpt_fn with error handling and retrying."""
tries = 0
# sleep to avoid overloading openai api
time.sleep(sleep_time)
try:
gpt_response = gpt_fn(BASE_SEED + tries)
result = parse_fn(gpt_response)
except Exception as error:
print('error:', error)
result = None
while result is None and tries < max_tries:
tries += 1
time.sleep(sleep_time)
print('retrying...')
try:
gpt_response = gpt_fn(BASE_SEED + tries)
result = parse_fn(gpt_response)
except:
result = None
return gpt_response
def show_img_to_caption(scene_dir, idx_to_caption):
img_dir = os.path.join(scene_dir, 'images')
imgs = load_images(img_dir, bg_change=None, return_masks=False)
img_to_caption = imgs[idx_to_caption]
plt.imshow(img_to_caption)
plt.show()
plt.close()
return
def predict_candidate_materials(args, scene_dir, show=False):
# load caption info
with open(os.path.join(scene_dir, '%s.json' % args.caption_load_name), 'r') as f:
info = json.load(f)
caption = info['caption']
gpt_fn = lambda seed: gpt_candidate_materials(caption, property_name=args.property_name,
model_name=args.gpt_model_name, seed=seed)
parse_fn = parse_material_hardness if args.property_name == 'hardness' else parse_material_list
candidate_materials = gpt_wrapper(gpt_fn, parse_fn)
info['candidate_materials_%s' % args.property_name] = candidate_materials
print('-' * 50)
print('scene: %s, info:' % os.path.basename(scene_dir), info)
print('candidate materials (%s):' % args.property_name)
mat_names, mat_vals = parse_fn(candidate_materials)
for mat_i, mat_name in enumerate(mat_names):
print('%16s: %8.1f -%8.1f' % (mat_name, mat_vals[mat_i][0], mat_vals[mat_i][1]))
if show:
show_img_to_caption(scene_dir, int(info['idx_to_caption']))
# save info to json
with open(os.path.join(scene_dir, '%s.json' % args.mats_save_name), 'w') as f:
json.dump(info, f, indent=4)
return info
def predict_object_info_gpt4v(args, scene_dir, show=False):
"""(EXPERIMENTAL) Predict materials directly from image with GPT-4V."""
img_dir = os.path.join(scene_dir, 'images')
imgs, masks = load_images(img_dir, return_masks=True)
mask_areas = [np.mean(mask) for mask in masks]
idx_to_caption = np.argsort(mask_areas)[int(len(mask_areas) * args.mask_area_percentile)]
img_to_caption = imgs[idx_to_caption]
# save img_to_caption in img_dir
img_to_caption = Image.fromarray(img_to_caption)
img_path = os.path.join(scene_dir, 'img_to_caption.png')
img_to_caption.save(img_path)
gpt_fn = lambda seed: gpt4v_candidate_materials(img_path, property_name=args.property_name, seed=seed)
candidate_materials = gpt_wrapper(gpt_fn, parse_material_json)
info = {'idx_to_caption': str(idx_to_caption),
'candidate_materials_%s' % args.property_name: candidate_materials}
print('-' * 50)
print('scene: %s, info:' % os.path.basename(scene_dir), info)
print('candidate materials (%s):' % args.property_name)
mat_names, mat_vals = parse_material_list(candidate_materials)
for mat_i, mat_name in enumerate(mat_names):
print('%16s: %8.1f -%8.1f' % (mat_name, mat_vals[mat_i][0], mat_vals[mat_i][1]))
if show:
show_img_to_caption(scene_dir, int(info['idx_to_caption']))
# save info to json
with open(os.path.join(scene_dir, '%s.json' % args.mats_save_name), 'w') as f:
json.dump(info, f, indent=4)
return info
def predict_thickness(args, scene_dir, mode='list', show=False):
# load info
with open(os.path.join(scene_dir, '%s.json' % args.mats_save_name), 'r') as f:
info = json.load(f)
if mode == 'list':
caption = info['caption']
elif mode == 'json': # json contains caption inside
caption = None
else:
raise NotImplementedError
candidate_materials = info['candidate_materials_density']
gpt_fn = lambda seed: gpt_thickness(caption, candidate_materials,
model_name=args.gpt_model_name, mode=mode, seed=seed)
thickness = gpt_wrapper(gpt_fn, parse_material_list)
info['thickness'] = thickness
print('thickness (cm):')
mat_names, mat_vals = parse_material_list(thickness)
for mat_i, mat_name in enumerate(mat_names):
print('%16s: %8.1f -%8.1f' % (mat_name, mat_vals[mat_i][0], mat_vals[mat_i][1]))
if show:
show_img_to_caption(scene_dir, int(info['idx_to_caption']))
# save info to json
with open(os.path.join(scene_dir, '%s.json' % args.mats_save_name), 'w') as f:
json.dump(info, f, indent=4)
return info
if __name__ == '__main__':
args = get_args()
scenes_dir = os.path.join(args.data_dir, 'scenes')
scenes = get_scenes_list(args)
openai.api_key = OPENAI_API_KEY
for j, scene in enumerate(scenes):
mats_info = predict_candidate_materials(args, os.path.join(scenes_dir, scene))
if args.include_thickness:
mats_info = predict_thickness(args, os.path.join(scenes_dir, scene))