-
Notifications
You must be signed in to change notification settings - Fork 7
/
evaluate.py
423 lines (346 loc) · 13.7 KB
/
evaluate.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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
import os
import json
import argparse
from tqdm import tqdm
import numpy as np
import pandas as pd
import scipy.stats as scstats
from skimage.io import imread
import zipfile
import warnings
warnings.filterwarnings('error')
from metrics.vqa import vqa_accuracy
from metrics.localization import loc_metric
from metrics.categorization import accuracy
from metrics.segmentation import seg_metric
from metrics.keypoint import kp_metric
from metrics.refexp import refexp_metric
from metrics.normal import sn_metric, get_mask_from_normals
from utils.io import load_json_object, dump_json_object, mkdir_if_not_exists, dumps_json_object
import utils.rle as rle
parser = argparse.ArgumentParser()
parser.add_argument('--pred_zip',type=str,default=None)
parser.add_argument('--pred_dir',type=str,default='/predictions')
parser.add_argument('--grit_samples_dir',type=str)
parser.add_argument('--grit_normals_dir',type=str)
parser.add_argument('--lemma2groups',type=str)
parser.add_argument('--subset',type=str)
parser.add_argument('--outdir',type=str)
TASKS_ABREV = dict(
categorization='cat',
localization='loc',
vqa='vqa',
refexp='ref',
segmentation='seg',
keypoint='kp',
normal='sn',
overall='all'
)
TASKS = [k for k in TASKS_ABREV if k!='overall']
NUM_TASKS = len(TASKS)
GROUPS = [
'people',
'body_parts',
'birds',
'animals',
'plants',
'furniture',
'structure',
'clothing',
'clothing_accessories',
'vehicles',
'transport_infrastructure',
'musical_instruments',
'food',
'beverages',
'technology',
'bathroom_objects',
'kitchen_objects',
'household_objects',
'stationery',
'sports_equipment',
'places',
'tools',
'brands',
'natural_landscape',
]
def compute_metric(pred,sample,task,args):
if task=='categorization':
return accuracy(pred['words'], sample['output']['words'][0])
elif task=='localization':
return loc_metric(pred['bboxes'], sample['output']['bboxes'])
elif task=='refexp':
return refexp_metric(pred['bboxes'], sample['output']['bboxes'])
elif task=='vqa':
return vqa_accuracy([pred['words']], sample['output']['words'], 1)
elif task=='segmentation':
stuff = "(Stuff)" in sample['input']['task_query']
pred_masks = rle.decode(pred['masks'])
gt_masks = rle.decode(sample['output']['masks'])
return seg_metric(pred_masks, gt_masks, stuff)
elif task=='keypoint':
return kp_metric(pred['points'], sample['output']['points'])
elif task=='normal':
source = sample['meta']['data_source']
gt_normals_path = os.path.join(
args.grit_normals_dir,
sample['output']['out_image_name'])
gt_normals_rgb = imread(gt_normals_path)[:,:,:3]
valid_mask = get_mask_from_normals(gt_normals_rgb)
pred_normals_path = os.path.join(
args.pred_dir,
args.subset,
'normals',
pred['normal'])
pred_normals_rgb = imread(pred_normals_path)[:,:,:3]
return sn_metric(pred_normals_rgb, gt_normals_rgb, valid_mask)
else:
raise NotImplementedError
def get_metric(df,metric,task,cgroup,partition):
df = df[(df.task==task) & (df.cgroup==cgroup)]
if partition in ['sameSrc','newSrc']:
df = df[(df.src==partition) & (df.dist=='undist')]
elif partition in ['sameCpt','newCpt']:
df = df[(df.cpt==partition) & (df.dist=='undist')]
elif partition in ['dist']:
df = df[(df.dist==partition)]
elif partition in ['undist']:
df = df[df.undistp==True]
elif partition=='deldist':
df_dist = df[df.dist=='dist']
r_dist = {r['example_id'][:-5]:r[metric] for r in df_dist.to_records()}
df_undistp = df[df.undistp==True]
r_undistp = {r['example_id']:r[metric] for r in df_undistp.to_records()}
records = []
for idx,undistp_metric in r_undistp.items():
record = {'example_id': idx}
record[metric] = undistp_metric - r_dist[idx]
records.append(record)
df = pd.DataFrame.from_records(records,columns=['example_id',metric])
else:
raise NotImplementedError()
samples = df[metric]
N = len(samples)
if N <= 1:
if N==1:
mean = samples.mean()
else:
mean = None
return mean,None,None,None,N
mean = samples.mean()
var = samples.var()
lower,upper = scstats.t.interval(0.95,N-1,mean,samples.sem()+1e-6)
return mean,lower,upper,var,N
def get_parallel_undist_example_ids(samples):
N = len('_dist')
return {s['input']['example_id'][:-N] \
for s in samples if s['meta']['is_distorted']}
def compute_avg_bounds(mus,vs,ns):
d = np.sqrt(np.sum([v/n for v,n in zip(vs,ns)])) / len(ns)
mean = np.mean(mus)
upper = mean + 1.96*d
lower = mean - 1.96*d
return mean, upper, lower
def compute_sample_metrics(args):
if args.pred_zip is not None:
with zipfile.ZipFile(args.pred_zip,'r') as f:
f.extractall(args.pred_dir)
params = load_json_object(
os.path.join(
args.pred_dir,
args.subset,
'params.json'))['params_in_millions']
records = []
missing_tasks = set()
for task in TASKS:
samples_json = os.path.join(
args.grit_samples_dir,
args.subset,
f'{task}.json')
if not os.path.exists(samples_json):
raise FileNotFoundError(f'{samples_json} does not exist')
pred_json = os.path.join(
args.pred_dir,
args.subset,
f'{task}.json')
if not os.path.exists(pred_json):
print(f'{pred_json} does not exist')
missing_tasks.add(task)
continue
lemma2groups = load_json_object(args.lemma2groups)
samples = load_json_object(samples_json)
undistp_example_ids = get_parallel_undist_example_ids(samples)
preds = load_json_object(pred_json)
preds = {pred['example_id']:pred for pred in preds}
task_metric = []
same_src_metric = []
for sample in tqdm(samples):
pred = preds[sample['output']['example_id']]
metric = compute_metric(pred,sample,task,args)
task_metric.append(metric)
if sample['meta']['is_new_source'] is False:
same_src_metric.append(metric)
concepts = set()
groups = set()
for c in sample['meta']['concepts']:
concepts.update(c['lemma'])
groups.update(lemma2groups.get(c['lemma'],['_ungrouped_']))
conf = pred['confidence']
src = 'sameSrc'
if sample['meta']['is_new_source']:
src = 'newSrc'
cpt = 'sameCpt'
if sample['meta']['has_new_concept']:
cpt = 'newCpt'
dist = 'undist'
if sample['meta']['is_distorted']:
dist = 'dist'
undistp = False
if sample['input']['example_id'] in undistp_example_ids:
undistp = True
records.append(dict(
example_id=sample['input']['example_id'],
task=task,
src=src,
cpt=cpt,
dist=dist,
undistp=undistp,
cgroup='any',
acc=100*metric,
inf=100*conf*metric,
misinf=100*conf*(1-metric),
conf=100*(conf),
sa=100*(conf*metric + (1-conf)*(1-metric)),
rmse=(100*(conf-metric))**2
))
for cg in groups:
if cg=='_ungrouped_':
continue
records.append(dict(
example_id=sample['input']['example_id'],
task=task,
src=src,
cpt=cpt,
dist=dist,
undistp=undistp,
cgroup=cg,
acc=100*metric,
inf=100*conf*metric,
misinf=100*conf*(1-metric),
conf=100*(conf),
sa=100*(conf*metric + (1-conf)*(1-metric)),
rmse=(100*(conf-metric))**2
))
df = pd.DataFrame.from_records(records)
metrics = dict()
metric_vars = dict()
for metric in ['acc','inf','misinf','conf','sa','rmse']:
for cgroup in ['any',*GROUPS]:
for partition in ['sameSrc','newSrc','agg','sameCpt','newCpt','dist','undist','deldist']:
if cgroup!='any':
# sameSrc and newSrc are needed for computing agg
if metric=='acc' and partition in ['sameSrc','newSrc','agg']:
pass
else:
continue
if metric=='rmse' and partition=='deldist':
continue
for task in TASKS:
task_abrev = TASKS_ABREV[task]
metric_name = f'{metric}.{cgroup}.{partition}.{task_abrev}'
if partition=='agg':
m1 = f'{metric}.{cgroup}.sameSrc.{task_abrev}'
m2 = f'{metric}.{cgroup}.newSrc.{task_abrev}'
if f'{m1}.mean' not in metrics or \
f'{m2}.mean' not in metrics:
continue
mus = [metrics[f'{m1}.mean'],metrics[f'{m2}.mean']]
vs = [metric_vars[m1],metric_vars[m2]]
ns = [metrics[f'{m1}.cnt'],metrics[f'{m2}.cnt']]
mean, upper, lower = compute_avg_bounds(mus, vs, ns)
else:
mean,lower,upper,var,cnt = get_metric(df, metric, task, cgroup, partition)
if None in [mean,lower,upper,var]:
continue
metrics[f'{metric_name}.cnt'] = cnt
metric_vars[metric_name] = var
metrics[f'{metric_name}.mean'] = mean
metrics[f'{metric_name}.upper'] = upper
metrics[f'{metric_name}.lower'] = lower
for metric in ['acc','inf','misinf','conf','sa','rmse']:
for partition in ['agg','sameSrc','newSrc','dist','undist','deldist']:
if metric=='rmse' and partition=='deldist':
continue
mus = []
vs = []
ns = []
cnts = dict()
for task in TASKS:
task_abrev = TASKS_ABREV[task]
if task in missing_tasks and partition=='agg':
mus.extend([0,0])
vs.extend([0,0])
ns.extend([100,100]) # doesn't matter what number you put here since variance is 0
elif task in missing_tasks:
mus.append(0)
vs.append(0)
ns.append(100) # doesn't matter what number you put here since variance is 0
elif partition=='agg':
for p in ['sameSrc','newSrc']:
mus.append(metrics[f'{metric}.any.{p}.{task_abrev}.mean'])
vs.append(metric_vars[f'{metric}.any.{p}.{task_abrev}'])
ns.append(metrics[f'{metric}.any.{p}.{task_abrev}.cnt'])
else:
p = partition
mus.append(metrics[f'{metric}.any.{p}.{task_abrev}.mean'])
vs.append(metric_vars[f'{metric}.any.{p}.{task_abrev}'])
ns.append(metrics[f'{metric}.any.{p}.{task_abrev}.cnt'])
mean, upper, lower = compute_avg_bounds(mus, vs, ns)
metrics[f'{metric}.any.{partition}.all.mean'] = mean
metrics[f'{metric}.any.{partition}.all.upper'] = upper
metrics[f'{metric}.any.{partition}.all.lower'] = lower
overall_metric = 'acc.any.agg.all'
metrics['overall.mean'] = metrics[f'{overall_metric}.mean']
metrics['overall.upper'] = metrics[f'{overall_metric}.upper']
metrics['overall.lower'] = metrics[f'{overall_metric}.lower']
metrics['params'] = params
for cgroup in GROUPS:
for partition in ['sameSrc','newSrc']:
for task in TASKS:
task_abrev = TASKS_ABREV[task]
for stat in ['mean','upper','lower','cnt','var']:
metrics.pop(f'acc.{cgroup}.{partition}.{task_abrev}.{stat}',None)
for k,v in list(metrics.items()):
if k=='params':
metrics[k] = int(v)
print(k,metrics[k])
continue
if 'overall' in k:
_,stat = k.split('.')
partition = 'overall'
else:
metric,cgroup,partition,task,stat = k.split('.')
if metric=='rmse' and stat in ['mean','upper','lower']:
if stat=='lower':
if partition!='deldist':
v = max(0,v)
v = v ** 0.5
if partition=='deldist':
metrics[k] = round(v,4)
continue
if stat == 'upper':
metrics[k] = round(min(100,v),4)
elif stat == 'lower':
metrics[k] = round(max(0,v),4)
elif stat == 'mean':
metrics[k] = round(v,4)
elif stat == 'var':
metrics.pop(k,None)
print(dumps_json_object(metrics,indent=4))
mkdir_if_not_exists(args.outdir)
dump_json_object(
metrics,
os.path.join(args.outdir,f'{args.subset}_metrics.json'))
if __name__=='__main__':
args = parser.parse_args()
compute_sample_metrics(args)