forked from hirl-team/HCSC
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy patheval_knn.py
198 lines (173 loc) · 7.82 KB
/
eval_knn.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
import os
import sys
import argparse
import torch
from torch import nn
import torch.distributed as dist
import torch.backends.cudnn as cudnn
from torchvision import datasets
from torchvision import transforms
from torchvision import models
from utils.options import parse_args_knn
import utils.utils as utils
def extract_feature_pipeline(args):
# ============ preparing data ... ============
transform = transforms.Compose([
transforms.Resize(256, interpolation=3),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
if args.debug:
dataset_train = ReturnIndexDataset(os.path.join(args.data, "val"), transform=transform)
else:
dataset_train = ReturnIndexDataset(os.path.join(args.data, "train"), transform=transform)
dataset_val = ReturnIndexDataset(os.path.join(args.data, "val"), transform=transform)
sampler = torch.utils.data.DistributedSampler(dataset_train, shuffle=False)
data_loader_train = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
drop_last=False,
)
data_loader_val = torch.utils.data.DataLoader(
dataset_val,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
drop_last=False,
)
print(f"Data loaded with {len(dataset_train)} train and {len(dataset_val)} val imgs.")
if args.arch in models.__dict__.keys():
model = models.__dict__[args.arch](num_classes=0)
model.fc = nn.Identity()
else:
print(f"Architecture {args.arch} non supported")
sys.exit(1)
model.cuda()
utils.load_pretrained_weights(model, args.pretrained, args.checkpoint_key, args.arch)
model.eval()
# ============ extract features ... ============
print("Extracting features for train set...")
train_features = extract_features(model, data_loader_train, args.use_cuda)
print("Extracting features for val set...")
test_features = extract_features(model, data_loader_val, args.use_cuda)
# by default, l2 normalization would be applied
if utils.get_rank() == 0:
train_features = nn.functional.normalize(train_features, dim=1, p=2)
test_features = nn.functional.normalize(test_features, dim=1, p=2)
train_labels = torch.tensor([s[-1] for s in dataset_train.samples]).long()
test_labels = torch.tensor([s[-1] for s in dataset_val.samples]).long()
print(train_labels)
# save features and labels
if args.dump_features and dist.get_rank() == 0:
torch.save(train_features.cpu(), os.path.join(args.dump_features, "trainfeat.pth"))
torch.save(test_features.cpu(), os.path.join(args.dump_features, "testfeat.pth"))
torch.save(train_labels.cpu(), os.path.join(args.dump_features, "trainlabels.pth"))
torch.save(test_labels.cpu(), os.path.join(args.dump_features, "testlabels.pth"))
return train_features, test_features, train_labels, test_labels
@torch.no_grad()
def extract_features(model, data_loader, use_cuda=True, multiscale=False):
metric_logger = utils.MetricLogger(delimiter=" ")
features = None
for samples, index in metric_logger.log_every(data_loader, 10):
samples = samples.cuda(non_blocking=True)
index = index.cuda(non_blocking=True)
if multiscale:
feats = utils.multi_scale(samples, model)
else:
feats = model(samples).clone()
# init storage feature matrix
if dist.get_rank() == 0 and features is None:
features = torch.zeros(len(data_loader.dataset), feats.shape[-1])
if use_cuda:
features = features.cuda(non_blocking=True)
print(f"Storing features into tensor of shape {features.shape}")
# get indexes from all processes
y_all = torch.empty(dist.get_world_size(), index.size(0), dtype=index.dtype, device=index.device)
y_l = list(y_all.unbind(0))
y_all_reduce = torch.distributed.all_gather(y_l, index, async_op=True)
y_all_reduce.wait()
index_all = torch.cat(y_l)
# share features between processes
feats_all = torch.empty(
dist.get_world_size(),
feats.size(0),
feats.size(1),
dtype=feats.dtype,
device=feats.device,
)
output_l = list(feats_all.unbind(0))
output_all_reduce = torch.distributed.all_gather(output_l, feats, async_op=True)
output_all_reduce.wait()
# update storage feature matrix
if dist.get_rank() == 0:
if use_cuda:
features.index_copy_(0, index_all, torch.cat(output_l))
else:
features.index_copy_(0, index_all.cpu(), torch.cat(output_l).cpu())
return features
@torch.no_grad()
def knn_classifier(train_features, train_labels, test_features,
test_labels, k, T, num_classes=1000,):
top1, top5, total = 0.0, 0.0, 0
train_features = train_features.t()
num_test_images, num_chunks = test_labels.shape[0], 100
imgs_per_chunk = num_test_images // num_chunks
retrieval_one_hot = torch.zeros(k, num_classes).cuda()
for idx in range(0, num_test_images, imgs_per_chunk):
# get the features for test images
features = test_features[
idx : min((idx + imgs_per_chunk), num_test_images), :
]
targets = test_labels[idx : min((idx + imgs_per_chunk), num_test_images)]
batch_size = targets.shape[0]
# calculate the dot product and compute top-k neighbors
similarity = torch.mm(features, train_features)
distances, indices = similarity.topk(k, largest=True, sorted=True)
candidates = train_labels.view(1, -1).expand(batch_size, -1)
retrieved_neighbors = torch.gather(candidates, 1, indices)
retrieval_one_hot.resize_(batch_size * k, num_classes).zero_()
retrieval_one_hot.scatter_(1, retrieved_neighbors.view(-1, 1), 1)
distances_transform = distances.clone().div_(T).exp_()
probs = torch.sum(
torch.mul(
retrieval_one_hot.view(batch_size, -1, num_classes),
distances_transform.view(batch_size, -1, 1),
),
1,
)
_, predictions = probs.sort(1, True)
# find the predictions that match the target
correct = predictions.eq(targets.data.view(-1, 1))
top1 = top1 + correct.narrow(1, 0, 1).sum().item()
top5 = top5 + correct.narrow(1, 0, min(5, k)).sum().item() # top5 does not make sense if k < 5
total += targets.size(0)
top1 = top1 * 100.0 / total
top5 = top5 * 100.0 / total
return top1, top5
class ReturnIndexDataset(datasets.ImageFolder):
def __getitem__(self, idx):
img, label = super(ReturnIndexDataset, self).__getitem__(idx)
return img, idx
if __name__ == '__main__':
args = parse_args_knn()
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
cudnn.benchmark = True
train_features, test_features, train_labels, test_labels = extract_feature_pipeline(args)
if utils.get_rank() == 0:
if args.use_cuda:
train_features = train_features.cuda()
test_features = test_features.cuda()
train_labels = train_labels.cuda()
test_labels = test_labels.cuda()
print("Features are ready!\nStart the k-NN classification.")
for k in args.nb_knn:
top1, top5 = knn_classifier(train_features, train_labels,
test_features, test_labels, k, args.temperature)
print(f"{k}-NN classifier result: Top1: {top1}, Top5: {top5}")
dist.barrier()