-
Notifications
You must be signed in to change notification settings - Fork 4
/
attention.py
390 lines (340 loc) · 14.1 KB
/
attention.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
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
import torch as T
from torch import nn
import torch.nn.functional as F
import numpy as np
from torch.nn import init
from dfn import DynamicConvFilter, DynamicConvFilterGenerator
from zoneout import ZoneoutLSTMCell
from util import *
def gaussian_masks(c, d, s, len_, glim_len):
'''
c, d, s: 2D Tensor (batch_size, n_glims)
len_, glim_len: int
returns: 4D Tensor (batch_size, n_glims, glim_len, len_)
each row is a 1D Gaussian
'''
batch_size, n_glims = c.size()
# The original HART code did not shift the coordinates by
# glim_len / 2. The generated Gaussian attention does not
# correspond to the actual crop of the bbox.
# Possibly a bug?
R = tovar(T.arange(0, glim_len)).view(1, 1, 1, -1) - glim_len / 2
C = tovar(T.arange(0, len_)).view(1, 1, -1, 1)
C = C.expand(batch_size, n_glims, len_, 1)
c = c[:, :, np.newaxis, np.newaxis]
d = d[:, :, np.newaxis, np.newaxis]
s = s[:, :, np.newaxis, np.newaxis]
cr = c + R * d
sr = tovar(T.ones(cr.size())) * s
mask = C - cr
mask = (-0.5 * (mask / sr) ** 2).exp()
mask = mask / (mask.sum(2, keepdim=True) + 1e-8)
return mask
def extract_gaussian_glims(x, a, glim_size):
'''
x: 4D Tensor (batch_size, nchannels, nrows, ncols)
a: 3D Tensor (batch_size, n_glims, att_params)
att_params: (cx, cy, dx, dy, sx, sy)
returns:
5D Tensor (batch_size, n_glims, nchannels, n_glim_rows, n_glim_cols)
'''
batch_size, n_glims, _ = a.size()
cx, cy, dx, dy, sx, sy = T.unbind(a, -1)
_, nchannels, nrows, ncols = x.size()
n_glim_rows, n_glim_cols = glim_size
# (batch_size, n_glims, nrows, n_glim_rows)
Fy = gaussian_masks(cy, dy, sy, nrows, n_glim_rows)
# (batch_size, n_glims, ncols, n_glim_cols)
Fx = gaussian_masks(cx, dx, sx, ncols, n_glim_cols)
# (batch_size, n_glims, 1, nrows, n_glim_rows)
Fy = Fy.unsqueeze(2)
# (batch_size, n_glims, 1, ncols, n_glim_cols)
Fx = Fx.unsqueeze(2)
# (batch_size, 1, nchannels, nrows, ncols)
x = x.unsqueeze(1)
# (batch_size, n_glims, nchannels, n_glim_rows, n_glim_cols)
g = Fy.transpose(-1, -2) @ x @ Fx
return g
class RATMAttention(nn.Module):
'''
[cx, cy, dx, dy, sx, sy]
0 <= cx, cy <= 1
0 <= dx, dy <= 1
'''
att_params = 6 # no. of parameters for attention parameterization
def __init__(self, x_size, glim_size):
'''
x_size: [n_image_rows, n_image_cols]
glim_size: [n_glim_rows, n_glim_cols]
'''
nn.Module.__init__(self)
self.x_size = x_size
self.glim_size = glim_size
def forward(self, x, spatial_att):
'''
x: 4D Tensor (batch_size, nchannels, n_image_rows, n_image_cols)
spatial_att: 3D Tensor (batch_size, n_glims, att_params) relative scales
'''
# (batch_size, n_glims, att_params)
absolute_att = self._to_absolute_attention(spatial_att)
glims = extract_gaussian_glims(x, absolute_att, self.glim_size)
return glims
def att_to_bbox(self, spatial_att):
'''
spatial_att: (..., 6) [cx, cy, dx, dy, sx, sy] relative scales ]0, 1[
return: (..., 4) [cx, cy, w, h] absolute scales
'''
cx = spatial_att[..., 0] * self.x_size[1]
cy = spatial_att[..., 1] * self.x_size[0]
w = T.abs(spatial_att[..., 2]) * (self.x_size[1] - 1)
h = T.abs(spatial_att[..., 3]) * (self.x_size[0] - 1)
bbox = T.stack([cx, cy, w, h], -1)
return bbox
def bbox_to_att(self, bbox):
'''
bbox: (..., 4) [cx, cy, w, h] absolute scales
return: (..., 6) [cx, cy, dx, dy, sx, sy] relative scales ]0, 1[
'''
cx = bbox[..., 0] / self.x_size[1]
cy = bbox[..., 1] / self.x_size[0]
dx = bbox[..., 2] / (self.x_size[1] - 1)
dy = bbox[..., 3] / (self.x_size[0] - 1)
sx = bbox[..., 2] * 0.5 / self.x_size[1]
sy = bbox[..., 3] * 0.5 / self.x_size[0]
spatial_att = T.stack([cx, cy, dx, dy, sx, sy], -1)
return spatial_att
def _to_axis_attention(self, image_len, glim_len, c, d, s):
c = c * image_len
d = d * (image_len - 1) / (glim_len - 1)
s = (s + 1e-5) * image_len / glim_len
return c, d, s
def _to_absolute_attention(self, params):
'''
params: 3D Tensor (batch_size, n_glims, att_params)
'''
n_image_rows, n_image_cols = self.x_size
n_glim_rows, n_glim_cols = self.glim_size
cx, dx, sx = T.unbind(params[..., ::2], -1)
cy, dy, sy = T.unbind(params[..., 1::2], -1)
cx, dx, sx = self._to_axis_attention(
n_image_cols, n_glim_cols, cx, dx, sx)
cy, dy, sy = self._to_axis_attention(
n_image_rows, n_glim_rows, cy, dy, sy)
# ap is now the absolute coordinate/scale on image
# (batch_size, n_glims, att_params)
ap = T.stack([cx, cy, dx, dy, sx, sy], -1)
return ap
class AttentionCell(nn.Module):
def __init__(self,
state_size,
image_size,
glim_size,
app_size,
feature_extractor,
attender,
zoneout_prob,
n_glims=1,
n_dfn_channels=10,
att_scale_logit_init=-2.5,
normalize_glimpse=False,
mask_feat_size=10):
nn.Module.__init__(self)
self.feature_extractor = feature_extractor
self.attender = attender
self.normalize_glimpse = normalize_glimpse
self.state_size = state_size
self.n_glims = n_glims
self.app_size = app_size
self.att_params = attender.att_params
self.mask_feat_size = mask_feat_size
glim_flatsize = np.asscalar(np.prod(glim_size))
self.pre_dfngen = DynamicConvFilterGenerator(
app_size,
feature_extractor.n_out_channels,
n_dfn_channels,
(1, 1))
self.pre_dfn = DynamicConvFilter(
feature_extractor.n_out_channels,
n_dfn_channels,
(1, 1))
self.dfngen = DynamicConvFilterGenerator(
app_size,
n_dfn_channels,
n_dfn_channels,
(3, 3))
self.dfn = DynamicConvFilter(
n_dfn_channels,
n_dfn_channels,
(3, 3),
padding=1)
self.proj = nn.Sequential(
nn.Linear(
n_glims * (
n_dfn_channels *
feature_extractor.compute_output_flatsize(glim_size) +
self.attender.att_params),
state_size),
nn.ELU(),
)
self.app_predictor = nn.Sequential(
nn.Linear(state_size, n_glims * app_size),
nn.ELU(),
)
self.rnncell = ZoneoutLSTMCell(state_size, state_size, p=zoneout_prob)
self.masker = nn.Conv2d(n_dfn_channels, 1, (1, 1))
self.mask_feat_extractor = nn.Sequential(
nn.Linear(
feature_extractor.compute_output_flatsize(glim_size),
mask_feat_size
),
nn.ELU(),
)
att_pred_output_layer = nn.Linear(state_size, n_glims * self.att_params)
init.uniform(att_pred_output_layer.weight, -1e-3, 1e-3)
init.constant(att_pred_output_layer.bias, 0)
self.att_pred_layer = nn.Sequential(
nn.Linear(state_size + mask_feat_size * n_glims, state_size),
nn.ELU(),
att_pred_output_layer,
nn.Tanh(),
)
self.att_delta_scale_logit = nn.Parameter(
T.zeros(n_glims, self.att_params) +
att_scale_logit_init)
self._att_bias = nn.Parameter(T.zeros(self.att_params)) # ?
self._att_scale = .25 # ???
@property
def att_bias(self):
return self._att_scale * T.tanh(self._att_bias)
def zero_state(self, x, bbox, presence):
'''
x: 4D (batch_size, nchannels, nrows, ncols)
bbox: 3D (batch_size, nobjects, 4) [cx, cy, w, h]
presence: 2D (batch_size, nobjects)
'''
batch_size = x.size()[0]
# (batch_size, nobjects = nglims, att_params)
att = self.attender.bbox_to_att(bbox)
rnn_state = self.rnncell.zero_state(batch_size)
_, feats, _, _, _ = self.extract_features(x, att, None)
rnn_output, rnn_state = self.rnncell(feats, rnn_state)
att += self.att_bias.view(1, 1, -1)
app = self.app_predictor(rnn_output)
app = app.view(batch_size, self.n_glims, self.app_size)
return rnn_output, rnn_state, att, app
def extract_features(self, x, spatial_att, appearance):
'''
x: 4D Tensor (batch_size, nchannels, n_image_rows, n_image_cols)
spatial_att: 3D Tensor (batch_size, n_glims, att_params)
appearance: 3D Tensor (batch_size, n_glims, app_size)
or None (when first called during hidden state initialization)
returns:
glims: 5D (batch_size, n_glims, nchannels, n_glim_rows, n_glim_cols)
projected_feats: 2D
(batch_size, state_size)
mask_logit: 3D or None (if appearance is None)
(batch_size, n_glims, raw_feat_rows, raw_feat_cols)
dfn_norm: L2 norm of generated DFN parameters
prenorm_glims: as @glims, but not contrast-normalized (for viz)
'''
# Extract raw glimpse from spatial attention
# (batch_size, n_glims, nchannels, n_glim_rows, n_glim_cols)
prenorm_glims = glims = self.attender(x, spatial_att)
batch_size, n_glims, nchannels, n_glim_rows, n_glim_cols = glims.size()
if self.normalize_glimpse:
glims = util.normalize_contrast(glims)
# Extract glimpse features
# (batch_size * n_glims, ...)
glims_reshaped = glims.view(-1, nchannels, n_glim_rows, n_glim_cols)
raw_feats, readout, feats = self.feature_extractor(glims_reshaped)
if appearance is not None:
# If we have appearance vector (i.e. not computing the initial RNN
# state), construct a 2-layer dynamic filter according to the
# appearance vector.
# The mask is computed by applying the dynamic filters on the
# features. The features will then be masked.
appearance = appearance.view(-1, self.app_size)
_, _, raw_feat_rows, raw_feat_cols = raw_feats.size()
# (batch_size * n_glims, n_dfn_channels,
# raw_feat_rows, raw_feat_cols)
pre_dfn_w, pre_dfn_b = self.pre_dfngen(appearance)
pre_dfn_feats = F.elu(self.pre_dfn(raw_feats, pre_dfn_w, pre_dfn_b))
dfn_w, dfn_b = self.dfngen(appearance)
dfn_feats = F.elu(self.dfn(pre_dfn_feats, dfn_w, dfn_b))
dfn_l2 = T.norm(pre_dfn_w) ** 2 + T.norm(pre_dfn_b) ** 2
dfn_l2 += T.norm(dfn_w) ** 2 + T.norm(dfn_b) ** 2
mask_logit = self.masker(dfn_feats)
mask = F.sigmoid(mask_logit)
masked_feats = feats * mask
mask_logit = mask_logit.squeeze(1).view(
batch_size, n_glims, raw_feat_rows, raw_feat_cols)
else:
masked_feats = feats
mask_logit = None
dfn_l2 = 0
# Now I'm collapsing glimpse features for every object into the
# same vector. Not sure if I need to make a separate one for each
# glimpse.
projected_feats = self.proj(
T.cat([
masked_feats.view(batch_size, -1),
spatial_att.view(batch_size, -1),
], 1)
)
return glims, projected_feats, mask_logit, dfn_l2, prenorm_glims
def forward(self,
x,
spatial_att,
appearance,
presence,
hidden_state):
'''
x: 4D (batch_size, nchannels, nrows, ncols)
spatial_att: 3D (batch_size, n_glims, att_params)
[cx, cy, dx, dy, sx, sy]
presence: (batch_size, n_glims)
appearance: (batch_size, n_glims, app_size)
hidden_state: something shouldn't care
returns:
rnn_output: (batch_size, state_size)
new_att: (batch_size, n_glims, att_params)
new_app: (batch_size, n_glims, app_size)
presence: (batch_size, n_glims)
glims: (batch_size, n_glims, nchannels, n_glim_rows, n_glim_cols)
mask_logit: (batch_size, n_glims, raw_feat_rows, raw_feat_cols)
mask_feats: (batch_size, n_glims, mask_feat_size)
next_state: something shouldn't care
'''
batch_size = x.size()[0]
glims, feats, mask_logit, dfn_l2, raw_glims = self.extract_features(
x, spatial_att, appearance)
rnn_output, next_state = self.rnncell(feats, hidden_state)
# Combine mask with RNN state
mask_feats = self.mask_feat_extractor(
mask_logit.view(batch_size * self.n_glims, -1))
att_input = T.cat(
[rnn_output, mask_feats.view(batch_size, -1)],
1)
mask_feats = mask_feats.view(
batch_size, self.n_glims, self.mask_feat_size)
# Compute the predicted
att_pred = self.att_pred_layer(att_input)
att_pred = att_pred.view(batch_size, self.n_glims, self.att_params)
att_delta_scale = F.sigmoid(self.att_delta_scale_logit)
new_att = spatial_att + att_delta_scale.unsqueeze(0) * att_pred
# The original HART paper uses RNN state itself as appearance feature
# vector... Here I'm projecting it once.
new_app = self.app_predictor(rnn_output)
new_app = new_app.view(batch_size, self.n_glims, self.app_size)
return (
new_att,
new_app,
presence,
next_state,
rnn_output,
glims,
mask_logit,
mask_feats,
dfn_l2,
raw_glims,
)