-
Notifications
You must be signed in to change notification settings - Fork 9
/
test_mlp.py
285 lines (215 loc) · 9.74 KB
/
test_mlp.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
import argparse
from tqdm import tqdm
from utils.utils import *
from loader.dataloader import dataloader
from torch.cuda.amp import autocast
from modules.segment_module import transform, untransform
from loader.netloader import network_loader, segment_mlp_loader, cluster_mlp_loader
def test(args, net, segment, cluster, nice, test_loader, cmap):
segment.eval()
prog_bar = tqdm(enumerate(test_loader), total=len(test_loader), leave=True)
with Pool(40) as pool:
for _, batch in prog_bar:
# image and label and self supervised feature
ind = batch["ind"].cuda()
img = batch["img"].cuda()
label = batch["label"].cuda()
with autocast():
# intermediate feature
feat = net(img)[:, 1:, :]
feat_flip = net(img.flip(dims=[3]))[:, 1:, :]
seg_feat = transform(segment.head_ema(feat))
seg_feat_flip = transform(segment.head_ema(feat_flip))
seg_feat = untransform((seg_feat + seg_feat_flip.flip(dims=[3])) / 2)
# interp feat
interp_seg_feat = F.interpolate(transform(seg_feat), label.shape[-2:], mode='bilinear', align_corners=False)
# cluster preds
cluster_preds = cluster.forward_centroid(untransform(interp_seg_feat), crf=True)
# crf
crf_preds = do_crf(pool, img, cluster_preds).argmax(1).cuda()
# nice evaluation
_, desc_nice = nice.eval(crf_preds, label)
# hungarian
hungarian_preds = nice.do_hungarian(crf_preds)
# save images
save_all(args, ind, img, label, cluster_preds.argmax(dim=1), crf_preds, hungarian_preds, cmap)
# real-time print
desc = f'{desc_nice}'
prog_bar.set_description(desc, refresh=True)
# evaludation metric reset
nice.reset()
def test_without_crf(args, net, segment, cluster, nice, test_loader):
segment.eval()
total_acc = 0
prog_bar = tqdm(enumerate(test_loader), total=len(test_loader), leave=True)
for idx, batch in prog_bar:
# image and label and self supervised feature
ind = batch["ind"].cuda()
img = batch["img"].cuda()
label = batch["label"].cuda()
# intermediate feature
with autocast():
feat = net(img)[:, 1:, :]
seg_feat_ema = segment.head_ema(feat)
# linear probe loss
linear_logits = segment.linear(seg_feat_ema)
linear_logits = F.interpolate(linear_logits, label.shape[-2:], mode='bilinear', align_corners=False)
flat_label = label.reshape(-1)
flat_label_mask = (flat_label >= 0) & (flat_label < args.n_classes)
# interp feat
interp_seg_feat = F.interpolate(transform(seg_feat_ema), label.shape[-2:], mode='bilinear', align_corners=False)
# cluster
cluster_preds = cluster.forward_centroid(untransform(interp_seg_feat), inference=True)
# nice evaluation
_, desc_nice = nice.eval(cluster_preds, label)
# linear probe acc check
pred_label = linear_logits.argmax(dim=1)
flat_pred_label = pred_label.reshape(-1)
acc = (flat_pred_label[flat_label_mask] == flat_label[flat_label_mask]).sum() / flat_label[
flat_label_mask].numel()
total_acc += acc.item()
# real-time print
desc = f'[TEST] Acc (Linear): {100. * total_acc / (idx + 1):.1f}% | {desc_nice}'
prog_bar.set_description(desc, refresh=True)
# evaludation metric reset
nice.reset()
def test_linear_without_crf(args, net, segment, nice, test_loader):
segment.eval()
prog_bar = tqdm(enumerate(test_loader), total=len(test_loader), leave=True)
with Pool(40) as pool:
for _, batch in prog_bar:
# image and label and self supervised feature
ind = batch["ind"].cuda()
img = batch["img"].cuda()
label = batch["label"].cuda()
with autocast():
# intermediate feature
feat = net(img)[:, 1:, :]
feat_flip = net(img.flip(dims=[3]))[:, 1:, :]
seg_feat = segment.transform(segment.head_ema(feat))
seg_feat_flip = segment.transform(segment.head_ema(feat_flip))
seg_feat = segment.untransform((seg_feat + seg_feat_flip.flip(dims=[3])) / 2)
# interp feat
interp_seg_feat = F.interpolate(segment.transform(seg_feat), label.shape[-2:], mode='bilinear', align_corners=False)
# linear probe interp feat
linear_logits = segment.linear(segment.untransform(interp_seg_feat))
# cluster preds
cluster_preds = linear_logits.argmax(dim=1)
# nice evaluation
_, desc_nice = nice.eval(cluster_preds, label)
# real-time print
desc = f'{desc_nice}'
prog_bar.set_description(desc, refresh=True)
# evaludation metric reset
nice.reset()
def test_linear(args, net, segment, nice, test_loader):
segment.eval()
prog_bar = tqdm(enumerate(test_loader), total=len(test_loader), leave=True)
with Pool(40) as pool:
for _, batch in prog_bar:
# image and label and self supervised feature
ind = batch["ind"].cuda()
img = batch["img"].cuda()
label = batch["label"].cuda()
with autocast():
# intermediate feature
feat = net(img)[:, 1:, :]
feat_flip = net(img.flip(dims=[3]))[:, 1:, :]
seg_feat = segment.transform(segment.head_ema(feat))
seg_feat_flip = segment.transform(segment.head_ema(feat_flip))
seg_feat = segment.untransform((seg_feat + seg_feat_flip.flip(dims=[3])) / 2)
# interp feat
interp_seg_feat = F.interpolate(segment.transform(seg_feat), label.shape[-2:], mode='bilinear', align_corners=False)
# linear probe interp feat
linear_logits = segment.linear(segment.untransform(interp_seg_feat))
# cluster preds
cluster_preds = torch.log_softmax(linear_logits * 10, dim=1).argmax(dim=1)
# crf
onehot = F.one_hot(cluster_preds.to(torch.int64), args.n_classes).to(torch.float32)
crf_preds = do_crf(pool, img, onehot.permute(0, 3, 1, 2)).argmax(1).cuda()
# nice evaluation
_, desc_nice = nice.eval(crf_preds, label)
# real-time print
desc = f'{desc_nice}'
prog_bar.set_description(desc, refresh=True)
# evaludation metric reset
nice.reset()
def main(rank, args):
# setting gpu id of this process
torch.cuda.set_device(rank)
# print argparse
print_argparse(args, rank=0)
# dataset loader
_, test_loader, _ = dataloader(args, False)
# network loader
net = network_loader(args, rank)
segment = segment_mlp_loader(args, rank)
cluster = cluster_mlp_loader(args, rank)
# evaluation
nice = NiceTool(args.n_classes)
# color map
cmap = create_cityscapes_colormap() if args.dataset == 'cityscapes' else create_pascal_label_colormap()
# param size
print(f'# of Parameters: {num_param(segment)/10**6:.2f}(M)')
# post-processing with crf and hungarian matching
test_without_crf(
args,
net,
segment,
cluster,
nice,
test_loader)
# post-processing with crf and hungarian matching
test(
args,
net,
segment,
cluster,
nice,
test_loader,
cmap)
# post-processing with crf and hungarian matching
# test_linear_without_crf(
# args,
# net,
# segment,
# nice,
# test_loader)
# test_linear(
# args,
# net,
# segment,
# nice,
# test_loader)
if __name__ == "__main__":
# fetch args
parser = argparse.ArgumentParser()
# model parameter
parser.add_argument('--NAME-TAG', default='CAUSE-MLP', type=str)
parser.add_argument('--data_dir', default='/mnt/hard2/lbk-iccv/datasets', type=str)
parser.add_argument('--dataset', default='coco171', type=str)
parser.add_argument('--port', default='12355', type=str)
parser.add_argument('--load_segment', default=True, type=str2bool)
parser.add_argument('--load_cluster', default=True, type=str2bool)
parser.add_argument('--ckpt', default='checkpoint/dino_vit_small_8.pth', type=str)
parser.add_argument('--distributed', default=False, type=str2bool)
parser.add_argument('--train_resolution', default=224, type=int)
parser.add_argument('--test_resolution', default=320, type=int)
parser.add_argument('--batch_size', default=32, type=int)
parser.add_argument('--num_workers', default=int(os.cpu_count() / 8), type=int)
parser.add_argument('--gpu', default='4', type=str)
parser.add_argument('--num_codebook', default=2048, type=int)
# model parameter
parser.add_argument('--reduced_dim', default=90, type=int)
parser.add_argument('--projection_dim', default=2048, type=int)
args = parser.parse_args()
if 'dinov2' in args.ckpt: args.test_resolution=322
if 'small' in args.ckpt:
args.dim = 384
elif 'base' in args.ckpt:
args.dim = 768
# the number of gpus for multi-process
gpu_list = list(map(int, args.gpu.split(',')))
ngpus_per_node = len(gpu_list)
# first gpu index is activated once there are several gpu in args.gpu
main(rank=gpu_list[0], args=args)