-
Notifications
You must be signed in to change notification settings - Fork 1
/
buffer.py
263 lines (204 loc) · 9.26 KB
/
buffer.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
import numpy as np
import math
import pdb
import torch
import torch.nn as nn
import torch.nn.functional as F
#import torchvision.transforms as transforms
import kornia
class Buffer(nn.Module):
def __init__(self, args, input_size=None):
super().__init__()
self.args = args
self.place_left = True
if input_size is None:
input_size = args.input_size
buffer_size = args.mem_size
print('buffer has %d slots' % buffer_size)
bx = torch.FloatTensor(buffer_size, *input_size).fill_(0)
by = torch.LongTensor(buffer_size).fill_(0)
bt = torch.LongTensor(buffer_size).fill_(0)
bidx = torch.LongTensor(buffer_size).fill_(0)
logits = torch.FloatTensor(buffer_size, args.n_classes).fill_(0)
self.current_index = 0
self.n_seen_so_far = 0
self.input_size = input_size
# registering as buffer allows us to save the object using `torch.save`
self.register_buffer('bx', bx)
self.register_buffer('by', by)
self.register_buffer('bt', bt)
self.register_buffer('logits', logits)
self.register_buffer('bidx', bidx)
self.to_one_hot = lambda x : x.new(x.size(0), args.n_classes).fill_(0).scatter_(1, x.unsqueeze(1), 1)
self.arange_like = lambda x : torch.arange(x.size(0)).to(x.device)
self.shuffle = lambda x : x[torch.randperm(x.size(0))]
@property
def x(self):
return self.bx[:self.current_index]
@property
def y(self):
return self.to_one_hot(self.by[:self.current_index])
@property
def t(self):
return self.bt[:self.current_index]
@property
def valid(self):
return self.is_valid[:self.current_index]
def add_reservoir(self, x, y, logits, t, idx, overwrite=True):
n_elem = x.size(0)
self.keep = None
# add whatever still fits in the buffer
place_left = max(0, self.bx.size(0) - self.current_index)
if place_left:
offset = min(place_left, n_elem)
self.bx[self.current_index: self.current_index + offset].data.copy_(x[:offset])
self.by[self.current_index: self.current_index + offset].data.copy_(y[:offset])
self.bt[self.current_index: self.current_index + offset].fill_(t)
if logits is not None:
self.logits[self.current_index: self.current_index + offset].data.copy_(logits[:offset])
if idx is not None:
self.bidx[self.current_index: self.current_index + offset].data.copy_(idx[:offset])
self.current_index += offset
self.n_seen_so_far += offset
# everything was added
if offset == x.size(0):
return
self.place_left = False
# remove what is already in the buffer
x, y, idx = x[place_left:], y[place_left:], idx[place_left:]
indices = torch.FloatTensor(x.size(0)).to(x.device).uniform_(0, self.n_seen_so_far).long()
valid_indices = (indices < self.bx.size(0))
idx_new_data = valid_indices.nonzero().squeeze(-1)
idx_buffer = indices[idx_new_data]
self.n_seen_so_far += x.size(0)
if idx_buffer.numel() == 0 and overwrite:
return
# perform overwrite op
if overwrite:
self.bx[idx_buffer] = x[idx_new_data]
self.by[idx_buffer] = y[idx_new_data]
self.bt[idx_buffer] = t
self.bidx[idx_buffer] = idx[idx_new_data]
if logits is not None:
self.logits[idx_buffer] = logits[idx_new_data]
else:
assert logits is None
del_new_data = (~valid_indices).nonzero().squeeze(-1) + self.bx.size(0)
""" instead we concatenate, and we will remove later! """
self.bx = torch.cat((self.bx, x))
self.by = torch.cat((self.by, y))
self.bt = torch.cat((self.bt, torch.zeros_like(y).fill_(t)))
self.bidx = torch.cat((self.bidx, idx))
assert self.by.size() == self.bt.size(), pdb.set_trace()
keep = torch.ones_like(self.by)
keep[idx_buffer] = 0
keep[del_new_data] = 0
# self.keep = keep.bool()
extra = keep.sum() - self.args.mem_size
if extra > 0:
# print(f'extra : {extra}')
probs = keep.float() / keep.sum()
also = torch.multinomial(probs, extra)
keep[also] = 0
assert keep.sum() == self.args.mem_size, pdb.set_trace()
self.keep = keep.bool()
def balance_memory(self):
if self.keep is None:
return
self.bx = self.bx[self.keep]
self.by = self.by[self.keep]
self.bt = self.bt[self.keep]
self.bidx = self.bidx[self.keep]
self.keep = None
def fetch_pos_neg_samples(self, label, task, idx, data=None, task_free=True):
# a sample is uniquely identifiable using `task` and `idx`
if type(task) == int:
task = torch.LongTensor(label.size(0)).to(label.device).fill_(task)
same_label = label.view(1, -1) == self.by.view(-1, 1) # buf_size x label_size
same_task = task.view(1, -1) == self.bt.view(-1, 1) # buf_size x label_size
same_idx = idx.view(1, -1) == self.bidx.view(-1, 1) # buf_size x label_size
same_ex = same_task & same_idx
task_labels = label.unique()
real_same_task = same_task
# TASK FREE METHOD : instead of using the task ID, we'll use labels in
# the current batch to mimic task
if task_free:
same_task = torch.zeros_like(same_task)
for label_ in task_labels:
label_exp = label_.view(1, -1).expand_as(same_task)
same_task = same_task | (label_exp == self.by.view(-1, 1))
valid_pos = same_label & ~same_ex
valid_neg_same_t = ~same_label & same_task & ~same_ex
valid_neg_diff_t = ~same_label & ~same_task & ~same_ex
# remove points which don't have pos, neg from same and diff t
has_valid_pos = valid_pos.sum(0) > 0
has_valid_neg = (valid_neg_same_t.sum(0) > 0) & (valid_neg_diff_t.sum(0) > 0)
invalid_idx = (~has_valid_pos | ~has_valid_neg).nonzero().squeeze()
if invalid_idx.numel() > 0:
# so the fetching operation won't fail
valid_pos[:, invalid_idx] = 1
valid_neg_same_t[:, invalid_idx] = 1
valid_neg_diff_t[:, invalid_idx] = 1
# easier if invalid_idx is a binary tensor
is_invalid = torch.zeros_like(label).bool()
is_invalid[invalid_idx] = 1
# fetch positive samples
pos_idx = torch.multinomial(valid_pos.float().T, 1).squeeze(1)
neg_idx_same_t = torch.multinomial(valid_neg_same_t.float().T, 1).squeeze(1)
neg_idx_diff_t = torch.multinomial(valid_neg_diff_t.float().T, 1).squeeze(1)
return self.bx[pos_idx], \
self.bx[neg_idx_same_t], \
self.bx[neg_idx_diff_t], \
is_invalid,\
self.by[pos_idx], \
self.by[neg_idx_same_t], \
self.by[neg_idx_diff_t]
def sample(self, amt, exclude_task=None, exclude_labels=None, return_logits=False, aug=False):
assert not (exclude_task is not None and exclude_labels is not None)
if exclude_task is not None:
valid_indices = (self.t != exclude_task)
valid_indices = valid_indices.nonzero().squeeze()
bx = self.bx[valid_indices]
by = self.by[valid_indices]
bt = self.bt[valid_indices]
logits = self.logits[valid_indices]
elif exclude_labels is not None:
# all true tensor
valid_indices = self.bt > -1
for label in exclude_labels:
valid_indices = valid_indices & (self.by != label)
valid_indices = valid_indices.nonzero().squeeze()
bx = self.bx[valid_indices]
by = self.by[valid_indices]
bt = self.bt[valid_indices]
logits = self.logits[valid_indices]
else:
bx = self.bx[:self.current_index]
by = self.by[:self.current_index]
bt = self.bt[:self.current_index]
bidx = self.bidx[:self.current_index]
logits = self.logits[:self.current_index]
if bx.size(0) < amt:
if return_logits:
return bx, by, bt, logits
else:
return bx, by, bt
else:
indices = torch.from_numpy(np.random.choice(bx.size(0), amt, replace=False))
indices = indices.to(self.bx.device)
if aug:
# A la G-Dumb oui
transform = nn.Sequential(
kornia.augmentation.RandomCrop(size=self.input_size[1:],padding=4),
kornia.augmentation.RandomHorizontalFlip()
)
ret = transform(bx[indices])
else:
ret = bx[indices]
if return_logits:
return ret, by[indices], bt[indices], logits[indices]
else:
return ret, by[indices], bt[indices]
def split(self, amt):
indices = torch.randperm(self.current_index).to(self.args.device)
return indices[:amt], indices[amt:]