-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathevaluate.py
173 lines (135 loc) · 6.94 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
import os, torch, math, tqdm, yaml
import pandas as pd
import numpy as np
from torch.utils.data import DataLoader
import torch.nn as nn
from easydict import EasyDict as edict
from datasets import PrecomputedInpaintingResultsDataset
from scores import SSIMScore, LPIPSScore, FIDScore
def move_to_device(obj, device):
if isinstance(obj, nn.Module):
return obj.to(device)
if torch.is_tensor(obj):
return obj.to(device)
if isinstance(obj, (tuple, list)):
return [move_to_device(el, device) for el in obj]
if isinstance(obj, dict):
return {name: move_to_device(val, device) for name, val in obj.items()}
raise ValueError(f'Unexpected type {type(obj)}')
def load_yaml(path):
with open(path, 'r') as f:
return edict(yaml.safe_load(f))
class InpaintingEvaluator():
def __init__(self, dataset, scores, area_grouping=True, bins=10, batch_size=32, device='cuda',
integral_func=None, integral_title=None, clamp_image_range=None):
"""
:param dataset: torch.utils.data.Dataset which contains images and masks
:param scores: dict {score_name: EvaluatorScore object}
:param area_grouping: in addition to the overall scores, allows to compute score for the groups of samples
which are defined by share of area occluded by mask
:param bins: number of groups, partition is generated by np.linspace(0., 1., bins + 1)
:param batch_size: batch_size for the dataloader
:param device: device to use
"""
self.scores = scores
self.dataset = dataset
self.area_grouping = area_grouping
self.bins = bins
self.device = torch.device(device)
self.dataloader = DataLoader(self.dataset, shuffle=False, batch_size=batch_size)
self.integral_func = integral_func
self.integral_title = integral_title
self.clamp_image_range = clamp_image_range
def _get_bin_edges(self):
bin_edges = np.linspace(0, 1, self.bins + 1)
num_digits = max(0, math.ceil(math.log10(self.bins)) - 1)
interval_names = []
for idx_bin in range(self.bins):
start_percent, end_percent = round(100 * bin_edges[idx_bin], num_digits), \
round(100 * bin_edges[idx_bin + 1], num_digits)
start_percent = '{:.{n}f}'.format(start_percent, n=num_digits)
end_percent = '{:.{n}f}'.format(end_percent, n=num_digits)
interval_names.append("{0}-{1}%".format(start_percent, end_percent))
groups = []
for batch in self.dataloader:
mask = batch['mask']
batch_size = mask.shape[0]
area = mask.to(self.device).reshape(batch_size, -1).mean(dim=-1)
bin_indices = np.searchsorted(bin_edges, area.detach().cpu().numpy(), side='right') - 1
# corner case: when area is equal to 1, bin_indices should return bins - 1, not bins for that element
bin_indices[bin_indices == self.bins] = self.bins - 1
groups.append(bin_indices)
groups = np.hstack(groups)
return groups, interval_names
def evaluate(self, model=None):
"""
:param model: callable with signature (image_batch, mask_batch); should return inpainted_batch
:return: dict with (score_name, group_type) as keys, where group_type can be either 'overall' or
name of the particular group arranged by area of mask (e.g. '10-20%')
and score statistics for the group as values.
"""
results = dict()
if self.area_grouping:
groups, interval_names = self._get_bin_edges()
else:
groups = None
for score_name, score in tqdm.auto.tqdm(self.scores.items(), desc='scores'):
score.to(self.device)
with torch.no_grad():
score.reset()
for batch in tqdm.auto.tqdm(self.dataloader, desc=score_name, leave=False):
batch = move_to_device(batch, self.device)
image_batch, mask_batch = batch['image'], batch['mask']
if self.clamp_image_range is not None:
image_batch = torch.clamp(image_batch,
min=self.clamp_image_range[0],
max=self.clamp_image_range[1])
if model is None:
assert 'inpainted' in batch, \
'Model is None, so we expected precomputed inpainting results at key "inpainted"'
inpainted_batch = batch['inpainted']
else:
inpainted_batch = model(image_batch, mask_batch)
score(inpainted_batch, image_batch, mask_batch)
total_results, group_results = score.get_value(groups=groups)
results[(score_name, 'total')] = total_results
if groups is not None:
for group_index, group_values in group_results.items():
group_name = interval_names[group_index]
results[(score_name, group_name)] = group_values
if self.integral_func is not None:
results[(self.integral_title, 'total')] = dict(mean=self.integral_func(results))
return results
def lpips_fid100_f1(metrics, fid_scale=100):
neg_lpips = 1 - metrics[('lpips', 'total')]['mean'] # invert, so bigger is better
fid = metrics[('fid', 'total')]['mean']
fid_rel = max(0, fid_scale - fid) / fid_scale
f1 = 2 * neg_lpips * fid_rel / (neg_lpips + fid_rel + 1e-3)
return f1
def main(args):
config = load_yaml("config.yaml")
dataset = PrecomputedInpaintingResultsDataset(args.datadir, args.predictdir, **config.dataset_kwargs)
metrics = {
'ssim': SSIMScore(),
'lpips': LPIPSScore(),
'fid': FIDScore()
}
enable_segm = config.get('segmentation', dict(enable=False)).get('enable', False)
evaluator = InpaintingEvaluator(dataset, scores=metrics,
integral_title='lpips_fid100_f1', integral_func=lpips_fid100_f1,
**config.evaluator_kwargs)
os.makedirs(os.path.dirname(args.outpath), exist_ok=True)
results = evaluator.evaluate()
results = pd.DataFrame(results).stack(1).unstack(0)
results.dropna(axis=1, how='all', inplace=True)
results.to_csv(args.outpath, sep='\t', float_format='%.4f')
print(results)
if __name__ == '__main__':
import argparse
aparser = argparse.ArgumentParser()
aparser.add_argument('datadir', type=str,
help='Path to folder with images and masks (output of gen_mask_dataset.py)')
aparser.add_argument('predictdir', type=str,
help='Path to folder with predicts (e.g. predict_hifill_baseline.py)')
aparser.add_argument('outpath', type=str, help='Where to put results')
main(aparser.parse_args())