-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
192 lines (188 loc) · 8.7 KB
/
utils.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
from datasets import *
from mindspore import nn
from model import resnet50
import mindspore.numpy as msnp
from loss import *
def get_preds_position_(unlabel_out, position, _postion, thres=0.001):
length = len(position)
r = []
un_idx = []
for idx in range(length):
pos = position[idx]
_pos = _postion[idx]
_out = unlabel_out[idx][pos]
out = ops.Softmax()(_out)
if len(pos)==1:
un_idx.append(idx)
continue
conf = ops.ArgMinWithValue()(out)[1]
if conf>thres:
un_idx.append(idx)
if len(_pos)==0:
r.append(ops.Argmin(axis=0)(out).asnumpy())
else:
r.append(_pos[-1])
continue
t, _ = get_preds(out)
a = pos[t]
_postion[idx].append(a)
position[idx].remove(a)
r.append(a)
return np.asarray(r), un_idx
def get_preds(out):
# out = F.softmax(out)
preds = ops.Argmin(axis=0)(out).asnumpy()
return preds, preds
def get_embedding(model, inputs, type=False):
batch_size = 64
inputs = Tensor(inputs)
embed = model(inputs)
assert embed.shape[0] == inputs.shape[0]
if type:return embed
return embed.numpy()
def train_loop(model, dataset, loss_fn, optimizer, inputs, targets):
# Define forward function
def forward_fn(data, label):
logits = model(data)
loss = loss_fn(logits, label) + entropy_loss(logits)
# loss = loss_fn(logits, label)
return loss, logits
# Get gradient function
grad_fn = ops.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)
# Define function of one-step training
def train_step(data, label):
(loss, logits), grads = grad_fn(data, label)
loss = ops.depend(loss, optimizer(grads))
return loss, logits
model.set_train()
inputs = Tensor(inputs)
loss, logits = train_step(inputs, targets)
return loss, logits
def train_loop_NL(model, dataset, loss_fn, optimizer, inputs, targets):
# Define forward function
def forward_fn(data, label):
logits = model(data)
loss = loss_fn(logits, label) + entropy_loss(logits)
# loss = loss_fn(logits, label)
return loss, logits
# Get gradient function
grad_fn = ops.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)
# Define function of one-step training
def train_step(data, label):
(loss, logits), grads = grad_fn(data, label)
loss = ops.depend(loss, optimizer(grads))
return loss, logits
model.set_train()
inputs = Tensor(inputs)
loss, logits = train_step(inputs, Tensor(targets))
return loss, logits
def meta_test_FC(args):
if args.dataset == 'cub':
num_classes = 100
elif args.dataset == 'tieredimagenet':
num_classes = 351
else:
num_classes = 64
model = resnet50(num_classes=num_classes, pretrained=False)
a = DataSet(data_root=args.folder)
sampler = ds.IterSampler(sampler=CategoriesSampler(a.label, args.num_batches,
args.num_test_ways, (args.num_shots, 15, args.unlabel)))
dataset = ds.GeneratorDataset(source=a, column_names=["image", "holder", "path"], sampler=sampler)
loader = dataset.create_dict_iterator()
k = args.num_shots * args.num_test_ways
for ia, check in enumerate(loader):
targets = msnp.tile(msnp.arange(args.num_test_ways), (args.num_shots + 15 + args.unlabel))
paths = check['path'].asnumpy()
data = []
c1 = transforms.Resize([84, 84], Inter.BICUBIC)
c2 = py_transforms.ToTensor()
for path in paths:
im = Image.open(path)
im = c1(im)
im = c2(im)
data.append(im)
train_inputs = data[:k]
train_targets = targets[:k]
test_inputs = data[k:k+15*args.num_test_ways]
test_targets = targets[k:k+15*args.num_test_ways].asnumpy()
unlabel_targets = targets[k + 15 * args.num_test_ways:].asnumpy()
train_embeddings = get_embedding(model=model, inputs=train_inputs, type=True)
test_embeddings = get_embedding(model, inputs=test_inputs, type=True)
if args.unlabel != 0:
unlabel_inputs = data[k + 15 * args.num_test_ways:]
unlabel_embeddings = get_embedding(model, unlabel_inputs, type=True)
if args.classifier =='linear':
ori_index = [x for x in range(250)]
_POSITION = [[] for _ in range(250)]
POSITION = [[0, 1, 2, 3, 4] for _ in range(250)]
criterion = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
clf = nn.Dense(in_channels=18432, out_channels=5)
optimizer = nn.SGD(clf.trainable_params(), learning_rate = 1e-3)
print('\n********************************************PL')
for epoch in range(100):
loss=train_loop(clf, None, criterion, optimizer, train_embeddings, train_targets)
print(f"Epoch: {epoch} Loss: {loss}")
unlabel_out = clf(unlabel_embeddings)
nl_pred, unselect_idx = get_preds_position_(unlabel_out, POSITION, _POSITION, thres = 0.2)
select_idx = [x for x in ori_index if x not in unselect_idx]
_unlabel_embeddings = unlabel_embeddings[select_idx]
_unlabel_t = unlabel_targets[select_idx]
nl_pred = nl_pred[select_idx]
print('********************************************SelectNL')
optimizer_NL = nn.SGD(clf.trainable_params(), learning_rate = 1e-3)
for epoch in range(10):
loss = train_loop_NL(clf, None, NL_loss, optimizer_NL, _unlabel_embeddings, nl_pred)
print(f"Epoch: {epoch} Loss: {loss}")
unlabel_out = clf(unlabel_embeddings)
nl_pred2, unselect_idx2 = get_preds_position_(unlabel_out, POSITION, _POSITION, thres=0.2)
# nl_pred2 = np.asarray(nl_pred2)
select_idx2 = [x for x in ori_index if x not in unselect_idx2]
_unlabel_embeddings = unlabel_embeddings[select_idx2]
_unlabel_t = unlabel_targets[select_idx2]
nl_pred2 = nl_pred2[select_idx2]
for epoch in range(10):
loss = train_loop_NL(clf, None, NL_loss, optimizer_NL, _unlabel_embeddings, nl_pred2)
print(f"Epoch: {epoch} Loss: {loss}")
unlabel_out = clf(unlabel_embeddings)
# nl_pred3, nl_conf = get_preds_third(unlabel_out)
nl_pred3, unselect_idx3 = get_preds_position_(unlabel_out, POSITION, _POSITION, thres=0.2)
nl_pred3 = np.asarray(nl_pred3)
select_idx3 = [x for x in ori_index if x not in unselect_idx3]
# select_idx3 = [x for x in ori_index if x not in _unselect_idx]
_unlabel_targets = unlabel_targets[select_idx3]
_unlabel_embeddings = unlabel_embeddings[select_idx3]
indexes_nl3 = [x for x in range(len(_unlabel_targets))]
nl_pred3 = nl_pred3[select_idx3]
for epoch in range(10):
loss = train_loop_NL(clf, None, NL_loss, optimizer_NL, _unlabel_embeddings, nl_pred3)
print(f"Epoch: {epoch} Loss: {loss}")
unlabel_out = clf(unlabel_embeddings)
nl_pred4, unselect_idx4 = get_preds_position_(unlabel_out, POSITION, _POSITION, thres=0.2)
# nl_pred4 = np.asarray(nl_pred4)
select_idx4 = [x for x in ori_index if x not in unselect_idx4]
# select_idx4 = [x for x in ori_index if x not in _unselect_idx]
_unlabel_targets = unlabel_targets[select_idx4]
_unlabel_embeddings = unlabel_embeddings[select_idx4]
nl_pred4 = nl_pred4[select_idx4]
indexes_nl4 = [x for x in range(len(_unlabel_targets))]
for epoch in range(10):
loss = train_loop_NL(clf, None, NL_loss, optimizer_NL, _unlabel_embeddings, nl_pred3)
print(f"Epoch: {epoch} Loss: {loss}")
class_num = [0 for _ in range(5)]
pseudo_label = []
index_pl = []
for idx in range(len(POSITION)):
item = POSITION[idx]
if len(item) == 1:
lab = item[0]
pseudo_label.append(item[-1])
class_num[lab] += 1
index_pl.append(idx)
class_num = [item + 8 for item in class_num]
max_ = max(class_num) * 1.0
pseudo_label = np.asarray(pseudo_label)
t1_ = unlabel_embeddings[index_pl]
t2_ = Tensor(pseudo_label, dtype=mstype.int64)
for epoch in range(100):
loss=train_loop(clf, None, criterion, optimizer, t1_, t2_)
print(f"Epoch: {epoch} Loss: {loss}")