183
improperly
243
9
274
213
floating
which
spaces
211
188
convolutions
159
identity
this
scripting
09
-
306
67
dilated
dropout
that
86
46
layers
of
221
computed
maxpool2d
piecewise
285
smoother
about
is
123
called
10
164
dimension
276
performs
for
addition
122
shrink
function
208
fractional
2
sigmoid
two
126
hardsigmoid
150
257
79
triggering
68
181
121
70
relu
39
applied
quantizable
batchnorm2d
elements
value
31
118
191
default
frozen
261
input
has
91
anchorgenerator
61
556
218
403
there
184
81
obscure
45
204
estimate
affine
187
413
limited
stride
115
stub
rectified
during
its
activation
06
18
66
246
features
242
drastically
245
anchor
95
ssd
feature
73
convolutional
03
62
363
convolved
includes
192
similarly
125
kernel
(
166
94
148
computationally
expensive
112
holder
level
75
in
top
30
7
into
343
32
262
92
calculates
311
47
_
21
module
237
uses
40
producing
124
additionally
operates
box
41
dequantization
gaussian
created
135
units
50
341
260
posenc
fixed
values
3
04
432
74
202
140
the
dynamically
34
output
37
2d
multi
270
82
409
)
71
relu6
across
1
65
instead
inserting
156
109
statistics
198
80
and
useful
network
to
point
197
281
155
22
transform
26
upsamplebilinear2d
55
251
training
.
where
weighted
174
rcnn
178
map
131
216
converts
generalized
same
422
normalization
adaptiveavgpool2d
196
swish
64
59
0
169
silu
195
05
pixel
standard
operation
approximation
49
20
76
scale
203
177
part
13
more
based
295
35
incoming
85
83
207
batch
19
36
layersthis
quantize
189
over
08
134
175
depth
54
87
structure
120
gelu
model
version
206
separable
17
25
395
from
5
avgpool2d
220
72
linear
287
96
290
average
contains
84
patch
detectors
143
neural
4
78
368
194
convolution
234
77
43
feeding
dil
modification
419
place
173
52
transformation
size
another
architecture
maximum
48
million
also
error
distance
42
158
116
299
non
165
attention
wider
layersin
108
93
while
277
fasterrcnn
106
172
fpn
200
classification
with
167
11
generator
15
489
last
between
hard
6
205
105
89
113
252
creates
or
313
280
303
exists
adaptive
101
239
57
parameters
before
405
60
44
unit
316
441
182
on
107
147
29
119
284
aims
calculating
69
23
single
210
350
replaces
without
297
114
applies
,
436
27
345
4d
63
232
102
28
209
146
nearest
logistic
339
quantized
sep
conv2d
zero
56
16
generalizedrcnn
24
322
hardswish
38
used
by
layernorm
486
142
classifier
overfitting
33
number
data
wise
new
51
88
97
07
are
divides
avoid
98
162
129
282
transposed
58
an
8
max
when
231
a
99
reduce
analogue
53
12
pooling
222
stochastic
four
137
141
327
chance
each
224
detection
132
piece
tensors
layer
482
14
it
roialign
totall
multiplied
180
