Hugging Face's logo

Hawkynt
/
mediadroid-models

LiteRT
ONNX
video-frame-interpolation
rife
dain
ncnn
xvfi
flavr
cain
realsr
android
Model card Files
xet
 Community
Hawkynt commited on
Commit
24ef5fe
·
verified ·
1 Parent(s): 3198814

* FLAVR: its 3D layers as 2D ones over the four frames — 31 times faster, on the GPU too

Browse files
Files changed (4) hide show
  1. README.md +7 -4
  2. flavr/body.bin +1 -1
  3. flavr/body.param +443 -111
  4. tools/flavr/export.py +74 -15
README.md CHANGED
@@ -116,11 +116,13 @@ the weights loaded as tensors only. About 84 MB, weights as half floats:
116
  | `body` | 4D: 3 channels x 4 frames x H x W, in 0..1, each channel's mean over the frames taken away | 64 x 4 x H x W |
117
  | `head` | 256 x H x W: body's four times folded into channels, time major | 3 x H x W, the mean still away |
118
 
119
- H and W multiples of 8. The in-between is that of the middle two frames. On the self-test's frames it is off by 5.23
120
- per value from the frame truly halfway (the PyTorch network 5.25, a blend 29.95). ncnn's 3D layers have no optimised or
121
- GPU path, so it is slow: 46 s at 128 x 128 on a desktop CPU.
 
 
122
 
123
- Changes (Apache-2.0, 4b): converted to ncnn as above, weights as half floats. Licence: `LICENSE-FLAVR`.
124
 
125
  ## Other models used by MediaDroid
126
 
@@ -136,3 +138,4 @@ Bao), `LICENSE-dain-ncnn-vulkan` (Copyright (c) 2020 nihui). XVFI for research a
136
 
137
 
138
 
 
 
116
  | `body` | 4D: 3 channels x 4 frames x H x W, in 0..1, each channel's mean over the frames taken away | 64 x 4 x H x W |
117
  | `head` | 256 x H x W: body's four times folded into channels, time major | 3 x H x W, the mean still away |
118
 
119
+ H and W multiples of 8. The in-between is that of the middle two frames. The 3D layers are written as 2D ones over the
120
+ four frames — the frames either side of each stacked into its channels, the four frames one below the other with zero
121
+ rows between, one ordinary 2D layer, the weights stored once — so ncnn runs them on its fast 2D paths and on the GPU.
122
+ On the self-test's frames the in-between is off by 5.23 per value from the frame truly halfway (the PyTorch network
123
+ 5.25, a blend 29.95); 1.5 s at 128 x 128 on a desktop CPU (46 s as ncnn's own 3D layers).
124
 
125
+ Changes (Apache-2.0, 4b): converted to ncnn as above — the 3D layers as 2D ones, their kernels' times as channels — weights as half floats. Licence: `LICENSE-FLAVR`.
126
 
127
  ## Other models used by MediaDroid
128
 
 
138
 
139
 
140
 
141
+
flavr/body.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d85e09c6774465d2f4fd77233c32b8d18f47c80a533ff155020aafdd8e0ed5c1
3
  size 84077848
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c09b1938d55530e65aaec54f4c54022c38d52a2340b1bc40b8fa751f2b56350
3
  size 84077848
flavr/body.param CHANGED
@@ -1,113 +1,445 @@
1
  7767517
2
- 111 136
3
  Input in0 0 1 in0 0=0
4
- Convolution3D convolution3d1 1 1 in0 b1 0=64 1=7 11=7 21=3 3=2 13=2 23=1 4=3 14=3 24=1 5=0 6=28224 9=1
5
- Split split_b1 1 3 b1 b1_0 b1_1 b1_2
6
- Convolution3D convolution3d2 1 1 b1_0 b2 0=64 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=110592 9=1
7
- Convolution3D convolution3d3 1 1 b2 b3 0=64 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=110592
8
- Split split_b3 1 2 b3 b3_0 b3_1
9
- Pooling3D pooling3d4 1 1 b3_0 b4 0=1 4=1
10
- InnerProduct innerproduct5 1 1 b4 b5 0=64 1=1 2=4096 9=4
11
- BinaryOp binaryop6 2 1 b3_1 b5 b6 0=2
12
- BinaryOp binaryop7 2 1 b6 b1_1 b7 0=0
13
- ReLU relu8 1 1 b7 b8
14
- Split split_b8 1 2 b8 b8_0 b8_1
15
- Convolution3D convolution3d9 1 1 b8_0 b9 0=64 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=110592 9=1
16
- Convolution3D convolution3d10 1 1 b9 b10 0=64 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=110592
17
- Split split_b10 1 2 b10 b10_0 b10_1
18
- Pooling3D pooling3d11 1 1 b10_0 b11 0=1 4=1
19
- InnerProduct innerproduct12 1 1 b11 b12 0=64 1=1 2=4096 9=4
20
- BinaryOp binaryop13 2 1 b10_1 b12 b13 0=2
21
- BinaryOp binaryop14 2 1 b13 b8_1 b14 0=0
22
- ReLU relu15 1 1 b14 b15
23
- Split split_b15 1 3 b15 b15_0 b15_1 b15_2
24
- Convolution3D convolution3d16 1 1 b15_0 b16 0=128 1=3 11=3 21=3 3=2 13=2 23=1 4=1 14=1 24=1 5=0 6=221184 9=1
25
- Convolution3D convolution3d17 1 1 b16 b17 0=128 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=442368
26
- Split split_b17 1 2 b17 b17_0 b17_1
27
- Pooling3D pooling3d18 1 1 b17_0 b18 0=1 4=1
28
- InnerProduct innerproduct19 1 1 b18 b19 0=128 1=1 2=16384 9=4
29
- BinaryOp binaryop20 2 1 b17_1 b19 b20 0=2
30
- Convolution3D convolution3d21 1 1 b15_1 b21 0=128 1=1 11=1 21=1 3=2 13=2 23=1 4=0 14=0 24=0 5=0 6=8192
31
- BinaryOp binaryop22 2 1 b20 b21 b22 0=0
32
- ReLU relu23 1 1 b22 b23
33
- Split split_b23 1 2 b23 b23_0 b23_1
34
- Convolution3D convolution3d24 1 1 b23_0 b24 0=128 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=442368 9=1
35
- Convolution3D convolution3d25 1 1 b24 b25 0=128 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=442368
36
- Split split_b25 1 2 b25 b25_0 b25_1
37
- Pooling3D pooling3d26 1 1 b25_0 b26 0=1 4=1
38
- InnerProduct innerproduct27 1 1 b26 b27 0=128 1=1 2=16384 9=4
39
- BinaryOp binaryop28 2 1 b25_1 b27 b28 0=2
40
- BinaryOp binaryop29 2 1 b28 b23_1 b29 0=0
41
- ReLU relu30 1 1 b29 b30
42
- Split split_b30 1 3 b30 b30_0 b30_1 b30_2
43
- Convolution3D convolution3d31 1 1 b30_0 b31 0=256 1=3 11=3 21=3 3=2 13=2 23=1 4=1 14=1 24=1 5=0 6=884736 9=1
44
- Convolution3D convolution3d32 1 1 b31 b32 0=256 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=1769472
45
- Split split_b32 1 2 b32 b32_0 b32_1
46
- Pooling3D pooling3d33 1 1 b32_0 b33 0=1 4=1
47
- InnerProduct innerproduct34 1 1 b33 b34 0=256 1=1 2=65536 9=4
48
- BinaryOp binaryop35 2 1 b32_1 b34 b35 0=2
49
- Convolution3D convolution3d36 1 1 b30_1 b36 0=256 1=1 11=1 21=1 3=2 13=2 23=1 4=0 14=0 24=0 5=0 6=32768
50
- BinaryOp binaryop37 2 1 b35 b36 b37 0=0
51
- ReLU relu38 1 1 b37 b38
52
- Split split_b38 1 2 b38 b38_0 b38_1
53
- Convolution3D convolution3d39 1 1 b38_0 b39 0=256 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=1769472 9=1
54
- Convolution3D convolution3d40 1 1 b39 b40 0=256 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=1769472
55
- Split split_b40 1 2 b40 b40_0 b40_1
56
- Pooling3D pooling3d41 1 1 b40_0 b41 0=1 4=1
57
- InnerProduct innerproduct42 1 1 b41 b42 0=256 1=1 2=65536 9=4
58
- BinaryOp binaryop43 2 1 b40_1 b42 b43 0=2
59
- BinaryOp binaryop44 2 1 b43 b38_1 b44 0=0
60
- ReLU relu45 1 1 b44 b45
61
- Split split_b45 1 3 b45 b45_0 b45_1 b45_2
62
- Convolution3D convolution3d46 1 1 b45_0 b46 0=512 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=3538944 9=1
63
- Convolution3D convolution3d47 1 1 b46 b47 0=512 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=7077888
64
- Split split_b47 1 2 b47 b47_0 b47_1
65
- Pooling3D pooling3d48 1 1 b47_0 b48 0=1 4=1
66
- InnerProduct innerproduct49 1 1 b48 b49 0=512 1=1 2=262144 9=4
67
- BinaryOp binaryop50 2 1 b47_1 b49 b50 0=2
68
- Convolution3D convolution3d51 1 1 b45_1 b51 0=512 1=1 11=1 21=1 3=1 13=1 23=1 4=0 14=0 24=0 5=0 6=131072
69
- BinaryOp binaryop52 2 1 b50 b51 b52 0=0
70
- ReLU relu53 1 1 b52 b53
71
- Split split_b53 1 2 b53 b53_0 b53_1
72
- Convolution3D convolution3d54 1 1 b53_0 b54 0=512 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=7077888 9=1
73
- Convolution3D convolution3d55 1 1 b54 b55 0=512 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=0 6=7077888
74
- Split split_b55 1 2 b55 b55_0 b55_1
75
- Pooling3D pooling3d56 1 1 b55_0 b56 0=1 4=1
76
- InnerProduct innerproduct57 1 1 b56 b57 0=512 1=1 2=262144 9=4
77
- BinaryOp binaryop58 2 1 b55_1 b57 b58 0=2
78
- BinaryOp binaryop59 2 1 b58 b53_1 b59 0=0
79
- ReLU relu60 1 1 b59 b60
80
- Convolution3D convolution3d61 1 1 b60 b61 0=256 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=1 6=3538944
81
- Split split_b61 1 2 b61 b61_0 b61_1
82
- Pooling3D pooling3d62 1 1 b61_0 b62 0=1 4=1
83
- InnerProduct innerproduct63 1 1 b62 b63 0=256 1=1 2=65536 9=4
84
- BinaryOp binaryop64 2 1 b61_1 b63 b64 0=2
85
- ReLU relu65 1 1 b64 b65 0=2.000000e-01
86
- Concat concat66 2 1 b65 b45_2 b66 0=0
87
- Deconvolution3D deconvolution3d67 1 1 b66 b67 0=128 1=4 11=4 21=3 3=2 13=2 23=1 4=1 14=1 24=1 5=1 6=3145728
88
- Split split_b67 1 2 b67 b67_0 b67_1
89
- Pooling3D pooling3d68 1 1 b67_0 b68 0=1 4=1
90
- InnerProduct innerproduct69 1 1 b68 b69 0=128 1=1 2=16384 9=4
91
- BinaryOp binaryop70 2 1 b67_1 b69 b70 0=2
92
- ReLU relu71 1 1 b70 b71 0=2.000000e-01
93
- Concat concat72 2 1 b71 b30_2 b72 0=0
94
- Deconvolution3D deconvolution3d73 1 1 b72 b73 0=64 1=4 11=4 21=3 3=2 13=2 23=1 4=1 14=1 24=1 5=1 6=786432
95
- Split split_b73 1 2 b73 b73_0 b73_1
96
- Pooling3D pooling3d74 1 1 b73_0 b74 0=1 4=1
97
- InnerProduct innerproduct75 1 1 b74 b75 0=64 1=1 2=4096 9=4
98
- BinaryOp binaryop76 2 1 b73_1 b75 b76 0=2
99
- ReLU relu77 1 1 b76 b77 0=2.000000e-01
100
- Concat concat78 2 1 b77 b15_2 b78 0=0
101
- Convolution3D convolution3d79 1 1 b78 b79 0=64 1=3 11=3 21=3 3=1 13=1 23=1 4=1 14=1 24=1 5=1 6=221184
102
- Split split_b79 1 2 b79 b79_0 b79_1
103
- Pooling3D pooling3d80 1 1 b79_0 b80 0=1 4=1
104
- InnerProduct innerproduct81 1 1 b80 b81 0=64 1=1 2=4096 9=4
105
- BinaryOp binaryop82 2 1 b79_1 b81 b82 0=2
106
- ReLU relu83 1 1 b82 b83 0=2.000000e-01
107
- Concat concat84 2 1 b83 b1_2 b84 0=0
108
- Deconvolution3D deconvolution3d85 1 1 b84 b85 0=64 1=4 11=4 21=3 3=2 13=2 23=1 4=1 14=1 24=1 5=1 6=393216
109
- Split split_b85 1 2 b85 b85_0 b85_1
110
- Pooling3D pooling3d86 1 1 b85_0 b86 0=1 4=1
111
- InnerProduct innerproduct87 1 1 b86 b87 0=64 1=1 2=4096 9=4
112
- BinaryOp binaryop88 2 1 b85_1 b87 b88 0=2
113
- ReLU relu89 1 1 b88 out 0=2.000000e-01
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  7767517
2
+ 443 556
3
  Input in0 0 1 in0 0=0
4
+ Split split_in0 1 5 in0 in0_0 in0_1 in0_2 in0_3 in0_4
5
+ Crop crop1 1 1 in0_0 b1 13=0 15=3 3=-233 4=-233 14=-233 5=-233
6
+ BinaryOp binaryop2 1 1 b1 b2 0=2 1=1 2=0.000000e+00
7
+ Crop crop3 1 1 in0_1 b3 13=0 15=1 3=-233 4=-233 14=-233 5=-233
8
+ Concat concat4 2 1 b2 b3 b4 0=1
9
+ Crop crop5 1 1 in0_2 b5 13=0 15=3 3=-233 4=-233 14=-233 5=-233
10
+ BinaryOp binaryop6 1 1 b5 b6 0=2 1=1 2=0.000000e+00
11
+ Crop crop7 1 1 in0_3 b7 13=1 15=0 3=-233 4=-233 14=-233 5=-233
12
+ Concat concat8 2 1 b7 b6 b8 0=1
13
+ Concat concat9 3 1 b4 in0_4 b8 b9 0=0
14
+ Padding padding10 1 1 b9 b10 0=0 1=4 2=0 3=0 4=0
15
+ Reshape reshape11 1 1 b10 b11 0=0 1=-1 2=9
16
+ Convolution convolution12 1 1 b11 b12 0=64 1=7 11=7 3=2 13=2 4=3 14=3 5=0 6=28224 9=1
17
+ Reshape reshape13 1 1 b12 b13 0=0 1=-1 11=4 2=64
18
+ Crop crop14 1 1 b13 b14 7=2 3=-233 4=-233 14=-233 5=-233
19
+ Split split_b14 1 7 b14 b14_0 b14_1 b14_2 b14_3 b14_4 b14_5 b14_6
20
+ Crop crop15 1 1 b14_0 b15 13=0 15=3 3=-233 4=-233 14=-233 5=-233
21
+ BinaryOp binaryop16 1 1 b15 b16 0=2 1=1 2=0.000000e+00
22
+ Crop crop17 1 1 b14_1 b17 13=0 15=1 3=-233 4=-233 14=-233 5=-233
23
+ Concat concat18 2 1 b16 b17 b18 0=1
24
+ Crop crop19 1 1 b14_2 b19 13=0 15=3 3=-233 4=-233 14=-233 5=-233
25
+ BinaryOp binaryop20 1 1 b19 b20 0=2 1=1 2=0.000000e+00
26
+ Crop crop21 1 1 b14_3 b21 13=1 15=0 3=-233 4=-233 14=-233 5=-233
27
+ Concat concat22 2 1 b21 b20 b22 0=1
28
+ Concat concat23 3 1 b18 b14_4 b22 b23 0=0
29
+ Padding padding24 1 1 b23 b24 0=0 1=1 2=0 3=0 4=0
30
+ Reshape reshape25 1 1 b24 b25 0=0 1=-1 2=192
31
+ Convolution convolution26 1 1 b25 b26 0=64 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=110592 9=1
32
+ Reshape reshape27 1 1 b26 b27 0=0 1=-1 11=4 2=64
33
+ Crop crop28 1 1 b27 b28 7=1 3=-233 4=-233 14=-233 5=-233
34
+ Split split_b28 1 5 b28 b28_0 b28_1 b28_2 b28_3 b28_4
35
+ Crop crop29 1 1 b28_0 b29 13=0 15=3 3=-233 4=-233 14=-233 5=-233
36
+ BinaryOp binaryop30 1 1 b29 b30 0=2 1=1 2=0.000000e+00
37
+ Crop crop31 1 1 b28_1 b31 13=0 15=1 3=-233 4=-233 14=-233 5=-233
38
+ Concat concat32 2 1 b30 b31 b32 0=1
39
+ Crop crop33 1 1 b28_2 b33 13=0 15=3 3=-233 4=-233 14=-233 5=-233
40
+ BinaryOp binaryop34 1 1 b33 b34 0=2 1=1 2=0.000000e+00
41
+ Crop crop35 1 1 b28_3 b35 13=1 15=0 3=-233 4=-233 14=-233 5=-233
42
+ Concat concat36 2 1 b35 b34 b36 0=1
43
+ Concat concat37 3 1 b32 b28_4 b36 b37 0=0
44
+ Padding padding38 1 1 b37 b38 0=0 1=1 2=0 3=0 4=0
45
+ Reshape reshape39 1 1 b38 b39 0=0 1=-1 2=192
46
+ Convolution convolution40 1 1 b39 b40 0=64 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=110592
47
+ Reshape reshape41 1 1 b40 b41 0=0 1=-1 11=4 2=64
48
+ Crop crop42 1 1 b41 b42 7=1 3=-233 4=-233 14=-233 5=-233
49
+ Reshape reshape43 1 1 b42 b43 0=0 1=-1 2=64
50
+ Split split_b43 1 2 b43 b43_0 b43_1
51
+ Pooling pooling44 1 1 b43_0 b44 0=1 4=1
52
+ InnerProduct innerproduct45 1 1 b44 b45 0=64 1=1 2=4096 9=4
53
+ BinaryOp binaryop46 2 1 b43_1 b45 b46 0=2
54
+ Reshape reshape47 1 1 b46 b47 0=0 1=-1 11=4 2=64
55
+ BinaryOp binaryop48 2 1 b47 b14_5 b48 0=0
56
+ ReLU relu49 1 1 b48 b49
57
+ Split split_b49 1 6 b49 b49_0 b49_1 b49_2 b49_3 b49_4 b49_5
58
+ Crop crop50 1 1 b49_0 b50 13=0 15=3 3=-233 4=-233 14=-233 5=-233
59
+ BinaryOp binaryop51 1 1 b50 b51 0=2 1=1 2=0.000000e+00
60
+ Crop crop52 1 1 b49_1 b52 13=0 15=1 3=-233 4=-233 14=-233 5=-233
61
+ Concat concat53 2 1 b51 b52 b53 0=1
62
+ Crop crop54 1 1 b49_2 b54 13=0 15=3 3=-233 4=-233 14=-233 5=-233
63
+ BinaryOp binaryop55 1 1 b54 b55 0=2 1=1 2=0.000000e+00
64
+ Crop crop56 1 1 b49_3 b56 13=1 15=0 3=-233 4=-233 14=-233 5=-233
65
+ Concat concat57 2 1 b56 b55 b57 0=1
66
+ Concat concat58 3 1 b53 b49_4 b57 b58 0=0
67
+ Padding padding59 1 1 b58 b59 0=0 1=1 2=0 3=0 4=0
68
+ Reshape reshape60 1 1 b59 b60 0=0 1=-1 2=192
69
+ Convolution convolution61 1 1 b60 b61 0=64 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=110592 9=1
70
+ Reshape reshape62 1 1 b61 b62 0=0 1=-1 11=4 2=64
71
+ Crop crop63 1 1 b62 b63 7=1 3=-233 4=-233 14=-233 5=-233
72
+ Split split_b63 1 5 b63 b63_0 b63_1 b63_2 b63_3 b63_4
73
+ Crop crop64 1 1 b63_0 b64 13=0 15=3 3=-233 4=-233 14=-233 5=-233
74
+ BinaryOp binaryop65 1 1 b64 b65 0=2 1=1 2=0.000000e+00
75
+ Crop crop66 1 1 b63_1 b66 13=0 15=1 3=-233 4=-233 14=-233 5=-233
76
+ Concat concat67 2 1 b65 b66 b67 0=1
77
+ Crop crop68 1 1 b63_2 b68 13=0 15=3 3=-233 4=-233 14=-233 5=-233
78
+ BinaryOp binaryop69 1 1 b68 b69 0=2 1=1 2=0.000000e+00
79
+ Crop crop70 1 1 b63_3 b70 13=1 15=0 3=-233 4=-233 14=-233 5=-233
80
+ Concat concat71 2 1 b70 b69 b71 0=1
81
+ Concat concat72 3 1 b67 b63_4 b71 b72 0=0
82
+ Padding padding73 1 1 b72 b73 0=0 1=1 2=0 3=0 4=0
83
+ Reshape reshape74 1 1 b73 b74 0=0 1=-1 2=192
84
+ Convolution convolution75 1 1 b74 b75 0=64 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=110592
85
+ Reshape reshape76 1 1 b75 b76 0=0 1=-1 11=4 2=64
86
+ Crop crop77 1 1 b76 b77 7=1 3=-233 4=-233 14=-233 5=-233
87
+ Reshape reshape78 1 1 b77 b78 0=0 1=-1 2=64
88
+ Split split_b78 1 2 b78 b78_0 b78_1
89
+ Pooling pooling79 1 1 b78_0 b79 0=1 4=1
90
+ InnerProduct innerproduct80 1 1 b79 b80 0=64 1=1 2=4096 9=4
91
+ BinaryOp binaryop81 2 1 b78_1 b80 b81 0=2
92
+ Reshape reshape82 1 1 b81 b82 0=0 1=-1 11=4 2=64
93
+ BinaryOp binaryop83 2 1 b82 b49_5 b83 0=0
94
+ ReLU relu84 1 1 b83 b84
95
+ Split split_b84 1 7 b84 b84_0 b84_1 b84_2 b84_3 b84_4 b84_5 b84_6
96
+ Crop crop85 1 1 b84_0 b85 13=0 15=3 3=-233 4=-233 14=-233 5=-233
97
+ BinaryOp binaryop86 1 1 b85 b86 0=2 1=1 2=0.000000e+00
98
+ Crop crop87 1 1 b84_1 b87 13=0 15=1 3=-233 4=-233 14=-233 5=-233
99
+ Concat concat88 2 1 b86 b87 b88 0=1
100
+ Crop crop89 1 1 b84_2 b89 13=0 15=3 3=-233 4=-233 14=-233 5=-233
101
+ BinaryOp binaryop90 1 1 b89 b90 0=2 1=1 2=0.000000e+00
102
+ Crop crop91 1 1 b84_3 b91 13=1 15=0 3=-233 4=-233 14=-233 5=-233
103
+ Concat concat92 2 1 b91 b90 b92 0=1
104
+ Concat concat93 3 1 b88 b84_4 b92 b93 0=0
105
+ Padding padding94 1 1 b93 b94 0=0 1=2 2=0 3=0 4=0
106
+ Reshape reshape95 1 1 b94 b95 0=0 1=-1 2=192
107
+ Convolution convolution96 1 1 b95 b96 0=128 1=3 11=3 3=2 13=2 4=1 14=1 5=0 6=221184 9=1
108
+ Reshape reshape97 1 1 b96 b97 0=0 1=-1 11=4 2=128
109
+ Crop crop98 1 1 b97 b98 7=1 3=-233 4=-233 14=-233 5=-233
110
+ Split split_b98 1 5 b98 b98_0 b98_1 b98_2 b98_3 b98_4
111
+ Crop crop99 1 1 b98_0 b99 13=0 15=3 3=-233 4=-233 14=-233 5=-233
112
+ BinaryOp binaryop100 1 1 b99 b100 0=2 1=1 2=0.000000e+00
113
+ Crop crop101 1 1 b98_1 b101 13=0 15=1 3=-233 4=-233 14=-233 5=-233
114
+ Concat concat102 2 1 b100 b101 b102 0=1
115
+ Crop crop103 1 1 b98_2 b103 13=0 15=3 3=-233 4=-233 14=-233 5=-233
116
+ BinaryOp binaryop104 1 1 b103 b104 0=2 1=1 2=0.000000e+00
117
+ Crop crop105 1 1 b98_3 b105 13=1 15=0 3=-233 4=-233 14=-233 5=-233
118
+ Concat concat106 2 1 b105 b104 b106 0=1
119
+ Concat concat107 3 1 b102 b98_4 b106 b107 0=0
120
+ Padding padding108 1 1 b107 b108 0=0 1=1 2=0 3=0 4=0
121
+ Reshape reshape109 1 1 b108 b109 0=0 1=-1 2=384
122
+ Convolution convolution110 1 1 b109 b110 0=128 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=442368
123
+ Reshape reshape111 1 1 b110 b111 0=0 1=-1 11=4 2=128
124
+ Crop crop112 1 1 b111 b112 7=1 3=-233 4=-233 14=-233 5=-233
125
+ Reshape reshape113 1 1 b112 b113 0=0 1=-1 2=128
126
+ Split split_b113 1 2 b113 b113_0 b113_1
127
+ Pooling pooling114 1 1 b113_0 b114 0=1 4=1
128
+ InnerProduct innerproduct115 1 1 b114 b115 0=128 1=1 2=16384 9=4
129
+ BinaryOp binaryop116 2 1 b113_1 b115 b116 0=2
130
+ Reshape reshape117 1 1 b116 b117 0=0 1=-1 11=4 2=128
131
+ Reshape reshape118 1 1 b84_5 b118 0=0 1=-1 2=64
132
+ Convolution convolution119 1 1 b118 b119 0=128 1=1 11=1 3=2 13=2 4=0 14=0 5=0 6=8192
133
+ Reshape reshape120 1 1 b119 b120 0=0 1=-1 11=4 2=128
134
+ BinaryOp binaryop121 2 1 b117 b120 b121 0=0
135
+ ReLU relu122 1 1 b121 b122
136
+ Split split_b122 1 6 b122 b122_0 b122_1 b122_2 b122_3 b122_4 b122_5
137
+ Crop crop123 1 1 b122_0 b123 13=0 15=3 3=-233 4=-233 14=-233 5=-233
138
+ BinaryOp binaryop124 1 1 b123 b124 0=2 1=1 2=0.000000e+00
139
+ Crop crop125 1 1 b122_1 b125 13=0 15=1 3=-233 4=-233 14=-233 5=-233
140
+ Concat concat126 2 1 b124 b125 b126 0=1
141
+ Crop crop127 1 1 b122_2 b127 13=0 15=3 3=-233 4=-233 14=-233 5=-233
142
+ BinaryOp binaryop128 1 1 b127 b128 0=2 1=1 2=0.000000e+00
143
+ Crop crop129 1 1 b122_3 b129 13=1 15=0 3=-233 4=-233 14=-233 5=-233
144
+ Concat concat130 2 1 b129 b128 b130 0=1
145
+ Concat concat131 3 1 b126 b122_4 b130 b131 0=0
146
+ Padding padding132 1 1 b131 b132 0=0 1=1 2=0 3=0 4=0
147
+ Reshape reshape133 1 1 b132 b133 0=0 1=-1 2=384
148
+ Convolution convolution134 1 1 b133 b134 0=128 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=442368 9=1
149
+ Reshape reshape135 1 1 b134 b135 0=0 1=-1 11=4 2=128
150
+ Crop crop136 1 1 b135 b136 7=1 3=-233 4=-233 14=-233 5=-233
151
+ Split split_b136 1 5 b136 b136_0 b136_1 b136_2 b136_3 b136_4
152
+ Crop crop137 1 1 b136_0 b137 13=0 15=3 3=-233 4=-233 14=-233 5=-233
153
+ BinaryOp binaryop138 1 1 b137 b138 0=2 1=1 2=0.000000e+00
154
+ Crop crop139 1 1 b136_1 b139 13=0 15=1 3=-233 4=-233 14=-233 5=-233
155
+ Concat concat140 2 1 b138 b139 b140 0=1
156
+ Crop crop141 1 1 b136_2 b141 13=0 15=3 3=-233 4=-233 14=-233 5=-233
157
+ BinaryOp binaryop142 1 1 b141 b142 0=2 1=1 2=0.000000e+00
158
+ Crop crop143 1 1 b136_3 b143 13=1 15=0 3=-233 4=-233 14=-233 5=-233
159
+ Concat concat144 2 1 b143 b142 b144 0=1
160
+ Concat concat145 3 1 b140 b136_4 b144 b145 0=0
161
+ Padding padding146 1 1 b145 b146 0=0 1=1 2=0 3=0 4=0
162
+ Reshape reshape147 1 1 b146 b147 0=0 1=-1 2=384
163
+ Convolution convolution148 1 1 b147 b148 0=128 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=442368
164
+ Reshape reshape149 1 1 b148 b149 0=0 1=-1 11=4 2=128
165
+ Crop crop150 1 1 b149 b150 7=1 3=-233 4=-233 14=-233 5=-233
166
+ Reshape reshape151 1 1 b150 b151 0=0 1=-1 2=128
167
+ Split split_b151 1 2 b151 b151_0 b151_1
168
+ Pooling pooling152 1 1 b151_0 b152 0=1 4=1
169
+ InnerProduct innerproduct153 1 1 b152 b153 0=128 1=1 2=16384 9=4
170
+ BinaryOp binaryop154 2 1 b151_1 b153 b154 0=2
171
+ Reshape reshape155 1 1 b154 b155 0=0 1=-1 11=4 2=128
172
+ BinaryOp binaryop156 2 1 b155 b122_5 b156 0=0
173
+ ReLU relu157 1 1 b156 b157
174
+ Split split_b157 1 7 b157 b157_0 b157_1 b157_2 b157_3 b157_4 b157_5 b157_6
175
+ Crop crop158 1 1 b157_0 b158 13=0 15=3 3=-233 4=-233 14=-233 5=-233
176
+ BinaryOp binaryop159 1 1 b158 b159 0=2 1=1 2=0.000000e+00
177
+ Crop crop160 1 1 b157_1 b160 13=0 15=1 3=-233 4=-233 14=-233 5=-233
178
+ Concat concat161 2 1 b159 b160 b161 0=1
179
+ Crop crop162 1 1 b157_2 b162 13=0 15=3 3=-233 4=-233 14=-233 5=-233
180
+ BinaryOp binaryop163 1 1 b162 b163 0=2 1=1 2=0.000000e+00
181
+ Crop crop164 1 1 b157_3 b164 13=1 15=0 3=-233 4=-233 14=-233 5=-233
182
+ Concat concat165 2 1 b164 b163 b165 0=1
183
+ Concat concat166 3 1 b161 b157_4 b165 b166 0=0
184
+ Padding padding167 1 1 b166 b167 0=0 1=2 2=0 3=0 4=0
185
+ Reshape reshape168 1 1 b167 b168 0=0 1=-1 2=384
186
+ Convolution convolution169 1 1 b168 b169 0=256 1=3 11=3 3=2 13=2 4=1 14=1 5=0 6=884736 9=1
187
+ Reshape reshape170 1 1 b169 b170 0=0 1=-1 11=4 2=256
188
+ Crop crop171 1 1 b170 b171 7=1 3=-233 4=-233 14=-233 5=-233
189
+ Split split_b171 1 5 b171 b171_0 b171_1 b171_2 b171_3 b171_4
190
+ Crop crop172 1 1 b171_0 b172 13=0 15=3 3=-233 4=-233 14=-233 5=-233
191
+ BinaryOp binaryop173 1 1 b172 b173 0=2 1=1 2=0.000000e+00
192
+ Crop crop174 1 1 b171_1 b174 13=0 15=1 3=-233 4=-233 14=-233 5=-233
193
+ Concat concat175 2 1 b173 b174 b175 0=1
194
+ Crop crop176 1 1 b171_2 b176 13=0 15=3 3=-233 4=-233 14=-233 5=-233
195
+ BinaryOp binaryop177 1 1 b176 b177 0=2 1=1 2=0.000000e+00
196
+ Crop crop178 1 1 b171_3 b178 13=1 15=0 3=-233 4=-233 14=-233 5=-233
197
+ Concat concat179 2 1 b178 b177 b179 0=1
198
+ Concat concat180 3 1 b175 b171_4 b179 b180 0=0
199
+ Padding padding181 1 1 b180 b181 0=0 1=1 2=0 3=0 4=0
200
+ Reshape reshape182 1 1 b181 b182 0=0 1=-1 2=768
201
+ Convolution convolution183 1 1 b182 b183 0=256 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=1769472
202
+ Reshape reshape184 1 1 b183 b184 0=0 1=-1 11=4 2=256
203
+ Crop crop185 1 1 b184 b185 7=1 3=-233 4=-233 14=-233 5=-233
204
+ Reshape reshape186 1 1 b185 b186 0=0 1=-1 2=256
205
+ Split split_b186 1 2 b186 b186_0 b186_1
206
+ Pooling pooling187 1 1 b186_0 b187 0=1 4=1
207
+ InnerProduct innerproduct188 1 1 b187 b188 0=256 1=1 2=65536 9=4
208
+ BinaryOp binaryop189 2 1 b186_1 b188 b189 0=2
209
+ Reshape reshape190 1 1 b189 b190 0=0 1=-1 11=4 2=256
210
+ Reshape reshape191 1 1 b157_5 b191 0=0 1=-1 2=128
211
+ Convolution convolution192 1 1 b191 b192 0=256 1=1 11=1 3=2 13=2 4=0 14=0 5=0 6=32768
212
+ Reshape reshape193 1 1 b192 b193 0=0 1=-1 11=4 2=256
213
+ BinaryOp binaryop194 2 1 b190 b193 b194 0=0
214
+ ReLU relu195 1 1 b194 b195
215
+ Split split_b195 1 6 b195 b195_0 b195_1 b195_2 b195_3 b195_4 b195_5
216
+ Crop crop196 1 1 b195_0 b196 13=0 15=3 3=-233 4=-233 14=-233 5=-233
217
+ BinaryOp binaryop197 1 1 b196 b197 0=2 1=1 2=0.000000e+00
218
+ Crop crop198 1 1 b195_1 b198 13=0 15=1 3=-233 4=-233 14=-233 5=-233
219
+ Concat concat199 2 1 b197 b198 b199 0=1
220
+ Crop crop200 1 1 b195_2 b200 13=0 15=3 3=-233 4=-233 14=-233 5=-233
221
+ BinaryOp binaryop201 1 1 b200 b201 0=2 1=1 2=0.000000e+00
222
+ Crop crop202 1 1 b195_3 b202 13=1 15=0 3=-233 4=-233 14=-233 5=-233
223
+ Concat concat203 2 1 b202 b201 b203 0=1
224
+ Concat concat204 3 1 b199 b195_4 b203 b204 0=0
225
+ Padding padding205 1 1 b204 b205 0=0 1=1 2=0 3=0 4=0
226
+ Reshape reshape206 1 1 b205 b206 0=0 1=-1 2=768
227
+ Convolution convolution207 1 1 b206 b207 0=256 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=1769472 9=1
228
+ Reshape reshape208 1 1 b207 b208 0=0 1=-1 11=4 2=256
229
+ Crop crop209 1 1 b208 b209 7=1 3=-233 4=-233 14=-233 5=-233
230
+ Split split_b209 1 5 b209 b209_0 b209_1 b209_2 b209_3 b209_4
231
+ Crop crop210 1 1 b209_0 b210 13=0 15=3 3=-233 4=-233 14=-233 5=-233
232
+ BinaryOp binaryop211 1 1 b210 b211 0=2 1=1 2=0.000000e+00
233
+ Crop crop212 1 1 b209_1 b212 13=0 15=1 3=-233 4=-233 14=-233 5=-233
234
+ Concat concat213 2 1 b211 b212 b213 0=1
235
+ Crop crop214 1 1 b209_2 b214 13=0 15=3 3=-233 4=-233 14=-233 5=-233
236
+ BinaryOp binaryop215 1 1 b214 b215 0=2 1=1 2=0.000000e+00
237
+ Crop crop216 1 1 b209_3 b216 13=1 15=0 3=-233 4=-233 14=-233 5=-233
238
+ Concat concat217 2 1 b216 b215 b217 0=1
239
+ Concat concat218 3 1 b213 b209_4 b217 b218 0=0
240
+ Padding padding219 1 1 b218 b219 0=0 1=1 2=0 3=0 4=0
241
+ Reshape reshape220 1 1 b219 b220 0=0 1=-1 2=768
242
+ Convolution convolution221 1 1 b220 b221 0=256 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=1769472
243
+ Reshape reshape222 1 1 b221 b222 0=0 1=-1 11=4 2=256
244
+ Crop crop223 1 1 b222 b223 7=1 3=-233 4=-233 14=-233 5=-233
245
+ Reshape reshape224 1 1 b223 b224 0=0 1=-1 2=256
246
+ Split split_b224 1 2 b224 b224_0 b224_1
247
+ Pooling pooling225 1 1 b224_0 b225 0=1 4=1
248
+ InnerProduct innerproduct226 1 1 b225 b226 0=256 1=1 2=65536 9=4
249
+ BinaryOp binaryop227 2 1 b224_1 b226 b227 0=2
250
+ Reshape reshape228 1 1 b227 b228 0=0 1=-1 11=4 2=256
251
+ BinaryOp binaryop229 2 1 b228 b195_5 b229 0=0
252
+ ReLU relu230 1 1 b229 b230
253
+ Split split_b230 1 7 b230 b230_0 b230_1 b230_2 b230_3 b230_4 b230_5 b230_6
254
+ Crop crop231 1 1 b230_0 b231 13=0 15=3 3=-233 4=-233 14=-233 5=-233
255
+ BinaryOp binaryop232 1 1 b231 b232 0=2 1=1 2=0.000000e+00
256
+ Crop crop233 1 1 b230_1 b233 13=0 15=1 3=-233 4=-233 14=-233 5=-233
257
+ Concat concat234 2 1 b232 b233 b234 0=1
258
+ Crop crop235 1 1 b230_2 b235 13=0 15=3 3=-233 4=-233 14=-233 5=-233
259
+ BinaryOp binaryop236 1 1 b235 b236 0=2 1=1 2=0.000000e+00
260
+ Crop crop237 1 1 b230_3 b237 13=1 15=0 3=-233 4=-233 14=-233 5=-233
261
+ Concat concat238 2 1 b237 b236 b238 0=1
262
+ Concat concat239 3 1 b234 b230_4 b238 b239 0=0
263
+ Padding padding240 1 1 b239 b240 0=0 1=1 2=0 3=0 4=0
264
+ Reshape reshape241 1 1 b240 b241 0=0 1=-1 2=768
265
+ Convolution convolution242 1 1 b241 b242 0=512 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=3538944 9=1
266
+ Reshape reshape243 1 1 b242 b243 0=0 1=-1 11=4 2=512
267
+ Crop crop244 1 1 b243 b244 7=1 3=-233 4=-233 14=-233 5=-233
268
+ Split split_b244 1 5 b244 b244_0 b244_1 b244_2 b244_3 b244_4
269
+ Crop crop245 1 1 b244_0 b245 13=0 15=3 3=-233 4=-233 14=-233 5=-233
270
+ BinaryOp binaryop246 1 1 b245 b246 0=2 1=1 2=0.000000e+00
271
+ Crop crop247 1 1 b244_1 b247 13=0 15=1 3=-233 4=-233 14=-233 5=-233
272
+ Concat concat248 2 1 b246 b247 b248 0=1
273
+ Crop crop249 1 1 b244_2 b249 13=0 15=3 3=-233 4=-233 14=-233 5=-233
274
+ BinaryOp binaryop250 1 1 b249 b250 0=2 1=1 2=0.000000e+00
275
+ Crop crop251 1 1 b244_3 b251 13=1 15=0 3=-233 4=-233 14=-233 5=-233
276
+ Concat concat252 2 1 b251 b250 b252 0=1
277
+ Concat concat253 3 1 b248 b244_4 b252 b253 0=0
278
+ Padding padding254 1 1 b253 b254 0=0 1=1 2=0 3=0 4=0
279
+ Reshape reshape255 1 1 b254 b255 0=0 1=-1 2=1536
280
+ Convolution convolution256 1 1 b255 b256 0=512 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=7077888
281
+ Reshape reshape257 1 1 b256 b257 0=0 1=-1 11=4 2=512
282
+ Crop crop258 1 1 b257 b258 7=1 3=-233 4=-233 14=-233 5=-233
283
+ Reshape reshape259 1 1 b258 b259 0=0 1=-1 2=512
284
+ Split split_b259 1 2 b259 b259_0 b259_1
285
+ Pooling pooling260 1 1 b259_0 b260 0=1 4=1
286
+ InnerProduct innerproduct261 1 1 b260 b261 0=512 1=1 2=262144 9=4
287
+ BinaryOp binaryop262 2 1 b259_1 b261 b262 0=2
288
+ Reshape reshape263 1 1 b262 b263 0=0 1=-1 11=4 2=512
289
+ Reshape reshape264 1 1 b230_5 b264 0=0 1=-1 2=256
290
+ Convolution convolution265 1 1 b264 b265 0=512 1=1 11=1 3=1 13=1 4=0 14=0 5=0 6=131072
291
+ Reshape reshape266 1 1 b265 b266 0=0 1=-1 11=4 2=512
292
+ BinaryOp binaryop267 2 1 b263 b266 b267 0=0
293
+ ReLU relu268 1 1 b267 b268
294
+ Split split_b268 1 6 b268 b268_0 b268_1 b268_2 b268_3 b268_4 b268_5
295
+ Crop crop269 1 1 b268_0 b269 13=0 15=3 3=-233 4=-233 14=-233 5=-233
296
+ BinaryOp binaryop270 1 1 b269 b270 0=2 1=1 2=0.000000e+00
297
+ Crop crop271 1 1 b268_1 b271 13=0 15=1 3=-233 4=-233 14=-233 5=-233
298
+ Concat concat272 2 1 b270 b271 b272 0=1
299
+ Crop crop273 1 1 b268_2 b273 13=0 15=3 3=-233 4=-233 14=-233 5=-233
300
+ BinaryOp binaryop274 1 1 b273 b274 0=2 1=1 2=0.000000e+00
301
+ Crop crop275 1 1 b268_3 b275 13=1 15=0 3=-233 4=-233 14=-233 5=-233
302
+ Concat concat276 2 1 b275 b274 b276 0=1
303
+ Concat concat277 3 1 b272 b268_4 b276 b277 0=0
304
+ Padding padding278 1 1 b277 b278 0=0 1=1 2=0 3=0 4=0
305
+ Reshape reshape279 1 1 b278 b279 0=0 1=-1 2=1536
306
+ Convolution convolution280 1 1 b279 b280 0=512 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=7077888 9=1
307
+ Reshape reshape281 1 1 b280 b281 0=0 1=-1 11=4 2=512
308
+ Crop crop282 1 1 b281 b282 7=1 3=-233 4=-233 14=-233 5=-233
309
+ Split split_b282 1 5 b282 b282_0 b282_1 b282_2 b282_3 b282_4
310
+ Crop crop283 1 1 b282_0 b283 13=0 15=3 3=-233 4=-233 14=-233 5=-233
311
+ BinaryOp binaryop284 1 1 b283 b284 0=2 1=1 2=0.000000e+00
312
+ Crop crop285 1 1 b282_1 b285 13=0 15=1 3=-233 4=-233 14=-233 5=-233
313
+ Concat concat286 2 1 b284 b285 b286 0=1
314
+ Crop crop287 1 1 b282_2 b287 13=0 15=3 3=-233 4=-233 14=-233 5=-233
315
+ BinaryOp binaryop288 1 1 b287 b288 0=2 1=1 2=0.000000e+00
316
+ Crop crop289 1 1 b282_3 b289 13=1 15=0 3=-233 4=-233 14=-233 5=-233
317
+ Concat concat290 2 1 b289 b288 b290 0=1
318
+ Concat concat291 3 1 b286 b282_4 b290 b291 0=0
319
+ Padding padding292 1 1 b291 b292 0=0 1=1 2=0 3=0 4=0
320
+ Reshape reshape293 1 1 b292 b293 0=0 1=-1 2=1536
321
+ Convolution convolution294 1 1 b293 b294 0=512 1=3 11=3 3=1 13=1 4=1 14=1 5=0 6=7077888
322
+ Reshape reshape295 1 1 b294 b295 0=0 1=-1 11=4 2=512
323
+ Crop crop296 1 1 b295 b296 7=1 3=-233 4=-233 14=-233 5=-233
324
+ Reshape reshape297 1 1 b296 b297 0=0 1=-1 2=512
325
+ Split split_b297 1 2 b297 b297_0 b297_1
326
+ Pooling pooling298 1 1 b297_0 b298 0=1 4=1
327
+ InnerProduct innerproduct299 1 1 b298 b299 0=512 1=1 2=262144 9=4
328
+ BinaryOp binaryop300 2 1 b297_1 b299 b300 0=2
329
+ Reshape reshape301 1 1 b300 b301 0=0 1=-1 11=4 2=512
330
+ BinaryOp binaryop302 2 1 b301 b268_5 b302 0=0
331
+ ReLU relu303 1 1 b302 b303
332
+ Split split_b303 1 5 b303 b303_0 b303_1 b303_2 b303_3 b303_4
333
+ Crop crop304 1 1 b303_0 b304 13=0 15=3 3=-233 4=-233 14=-233 5=-233
334
+ BinaryOp binaryop305 1 1 b304 b305 0=2 1=1 2=0.000000e+00
335
+ Crop crop306 1 1 b303_1 b306 13=0 15=1 3=-233 4=-233 14=-233 5=-233
336
+ Concat concat307 2 1 b305 b306 b307 0=1
337
+ Crop crop308 1 1 b303_2 b308 13=0 15=3 3=-233 4=-233 14=-233 5=-233
338
+ BinaryOp binaryop309 1 1 b308 b309 0=2 1=1 2=0.000000e+00
339
+ Crop crop310 1 1 b303_3 b310 13=1 15=0 3=-233 4=-233 14=-233 5=-233
340
+ Concat concat311 2 1 b310 b309 b311 0=1
341
+ Concat concat312 3 1 b307 b303_4 b311 b312 0=0
342
+ Padding padding313 1 1 b312 b313 0=0 1=1 2=0 3=0 4=0
343
+ Reshape reshape314 1 1 b313 b314 0=0 1=-1 2=1536
344
+ Convolution convolution315 1 1 b314 b315 0=256 1=3 11=3 3=1 13=1 4=1 14=1 5=1 6=3538944
345
+ Reshape reshape316 1 1 b315 b316 0=0 1=-1 11=4 2=256
346
+ Crop crop317 1 1 b316 b317 7=1 3=-233 4=-233 14=-233 5=-233
347
+ Reshape reshape318 1 1 b317 b318 0=0 1=-1 2=256
348
+ Split split_b318 1 2 b318 b318_0 b318_1
349
+ Pooling pooling319 1 1 b318_0 b319 0=1 4=1
350
+ InnerProduct innerproduct320 1 1 b319 b320 0=256 1=1 2=65536 9=4
351
+ BinaryOp binaryop321 2 1 b318_1 b320 b321 0=2
352
+ Reshape reshape322 1 1 b321 b322 0=0 1=-1 11=4 2=256
353
+ ReLU relu323 1 1 b322 b323 0=2.000000e-01
354
+ Concat concat324 2 1 b323 b230_6 b324 0=0
355
+ Split split_b324 1 5 b324 b324_0 b324_1 b324_2 b324_3 b324_4
356
+ Crop crop325 1 1 b324_0 b325 13=0 15=3 3=-233 4=-233 14=-233 5=-233
357
+ BinaryOp binaryop326 1 1 b325 b326 0=2 1=1 2=0.000000e+00
358
+ Crop crop327 1 1 b324_1 b327 13=0 15=1 3=-233 4=-233 14=-233 5=-233
359
+ Concat concat328 2 1 b326 b327 b328 0=1
360
+ Crop crop329 1 1 b324_2 b329 13=0 15=3 3=-233 4=-233 14=-233 5=-233
361
+ BinaryOp binaryop330 1 1 b329 b330 0=2 1=1 2=0.000000e+00
362
+ Crop crop331 1 1 b324_3 b331 13=1 15=0 3=-233 4=-233 14=-233 5=-233
363
+ Concat concat332 2 1 b331 b330 b332 0=1
364
+ Concat concat333 3 1 b332 b324_4 b328 b333 0=0
365
+ Padding padding334 1 1 b333 b334 0=0 1=1 2=0 3=0 4=0
366
+ Reshape reshape335 1 1 b334 b335 0=0 1=-1 2=1536
367
+ Deconvolution deconvolution336 1 1 b335 b336 0=128 1=4 11=4 3=2 13=2 4=1 14=1 5=1 6=3145728
368
+ Reshape reshape337 1 1 b336 b337 0=0 1=-1 11=4 2=128
369
+ Crop crop338 1 1 b337 b338 7=2 3=-233 4=-233 14=-233 5=-233
370
+ Reshape reshape339 1 1 b338 b339 0=0 1=-1 2=128
371
+ Split split_b339 1 2 b339 b339_0 b339_1
372
+ Pooling pooling340 1 1 b339_0 b340 0=1 4=1
373
+ InnerProduct innerproduct341 1 1 b340 b341 0=128 1=1 2=16384 9=4
374
+ BinaryOp binaryop342 2 1 b339_1 b341 b342 0=2
375
+ Reshape reshape343 1 1 b342 b343 0=0 1=-1 11=4 2=128
376
+ ReLU relu344 1 1 b343 b344 0=2.000000e-01
377
+ Concat concat345 2 1 b344 b157_6 b345 0=0
378
+ Split split_b345 1 5 b345 b345_0 b345_1 b345_2 b345_3 b345_4
379
+ Crop crop346 1 1 b345_0 b346 13=0 15=3 3=-233 4=-233 14=-233 5=-233
380
+ BinaryOp binaryop347 1 1 b346 b347 0=2 1=1 2=0.000000e+00
381
+ Crop crop348 1 1 b345_1 b348 13=0 15=1 3=-233 4=-233 14=-233 5=-233
382
+ Concat concat349 2 1 b347 b348 b349 0=1
383
+ Crop crop350 1 1 b345_2 b350 13=0 15=3 3=-233 4=-233 14=-233 5=-233
384
+ BinaryOp binaryop351 1 1 b350 b351 0=2 1=1 2=0.000000e+00
385
+ Crop crop352 1 1 b345_3 b352 13=1 15=0 3=-233 4=-233 14=-233 5=-233
386
+ Concat concat353 2 1 b352 b351 b353 0=1
387
+ Concat concat354 3 1 b353 b345_4 b349 b354 0=0
388
+ Padding padding355 1 1 b354 b355 0=0 1=1 2=0 3=0 4=0
389
+ Reshape reshape356 1 1 b355 b356 0=0 1=-1 2=768
390
+ Deconvolution deconvolution357 1 1 b356 b357 0=64 1=4 11=4 3=2 13=2 4=1 14=1 5=1 6=786432
391
+ Reshape reshape358 1 1 b357 b358 0=0 1=-1 11=4 2=64
392
+ Crop crop359 1 1 b358 b359 7=2 3=-233 4=-233 14=-233 5=-233
393
+ Reshape reshape360 1 1 b359 b360 0=0 1=-1 2=64
394
+ Split split_b360 1 2 b360 b360_0 b360_1
395
+ Pooling pooling361 1 1 b360_0 b361 0=1 4=1
396
+ InnerProduct innerproduct362 1 1 b361 b362 0=64 1=1 2=4096 9=4
397
+ BinaryOp binaryop363 2 1 b360_1 b362 b363 0=2
398
+ Reshape reshape364 1 1 b363 b364 0=0 1=-1 11=4 2=64
399
+ ReLU relu365 1 1 b364 b365 0=2.000000e-01
400
+ Concat concat366 2 1 b365 b84_6 b366 0=0
401
+ Split split_b366 1 5 b366 b366_0 b366_1 b366_2 b366_3 b366_4
402
+ Crop crop367 1 1 b366_0 b367 13=0 15=3 3=-233 4=-233 14=-233 5=-233
403
+ BinaryOp binaryop368 1 1 b367 b368 0=2 1=1 2=0.000000e+00
404
+ Crop crop369 1 1 b366_1 b369 13=0 15=1 3=-233 4=-233 14=-233 5=-233
405
+ Concat concat370 2 1 b368 b369 b370 0=1
406
+ Crop crop371 1 1 b366_2 b371 13=0 15=3 3=-233 4=-233 14=-233 5=-233
407
+ BinaryOp binaryop372 1 1 b371 b372 0=2 1=1 2=0.000000e+00
408
+ Crop crop373 1 1 b366_3 b373 13=1 15=0 3=-233 4=-233 14=-233 5=-233
409
+ Concat concat374 2 1 b373 b372 b374 0=1
410
+ Concat concat375 3 1 b370 b366_4 b374 b375 0=0
411
+ Padding padding376 1 1 b375 b376 0=0 1=1 2=0 3=0 4=0
412
+ Reshape reshape377 1 1 b376 b377 0=0 1=-1 2=384
413
+ Convolution convolution378 1 1 b377 b378 0=64 1=3 11=3 3=1 13=1 4=1 14=1 5=1 6=221184
414
+ Reshape reshape379 1 1 b378 b379 0=0 1=-1 11=4 2=64
415
+ Crop crop380 1 1 b379 b380 7=1 3=-233 4=-233 14=-233 5=-233
416
+ Reshape reshape381 1 1 b380 b381 0=0 1=-1 2=64
417
+ Split split_b381 1 2 b381 b381_0 b381_1
418
+ Pooling pooling382 1 1 b381_0 b382 0=1 4=1
419
+ InnerProduct innerproduct383 1 1 b382 b383 0=64 1=1 2=4096 9=4
420
+ BinaryOp binaryop384 2 1 b381_1 b383 b384 0=2
421
+ Reshape reshape385 1 1 b384 b385 0=0 1=-1 11=4 2=64
422
+ ReLU relu386 1 1 b385 b386 0=2.000000e-01
423
+ Concat concat387 2 1 b386 b14_6 b387 0=0
424
+ Split split_b387 1 5 b387 b387_0 b387_1 b387_2 b387_3 b387_4
425
+ Crop crop388 1 1 b387_0 b388 13=0 15=3 3=-233 4=-233 14=-233 5=-233
426
+ BinaryOp binaryop389 1 1 b388 b389 0=2 1=1 2=0.000000e+00
427
+ Crop crop390 1 1 b387_1 b390 13=0 15=1 3=-233 4=-233 14=-233 5=-233
428
+ Concat concat391 2 1 b389 b390 b391 0=1
429
+ Crop crop392 1 1 b387_2 b392 13=0 15=3 3=-233 4=-233 14=-233 5=-233
430
+ BinaryOp binaryop393 1 1 b392 b393 0=2 1=1 2=0.000000e+00
431
+ Crop crop394 1 1 b387_3 b394 13=1 15=0 3=-233 4=-233 14=-233 5=-233
432
+ Concat concat395 2 1 b394 b393 b395 0=1
433
+ Concat concat396 3 1 b395 b387_4 b391 b396 0=0
434
+ Padding padding397 1 1 b396 b397 0=0 1=1 2=0 3=0 4=0
435
+ Reshape reshape398 1 1 b397 b398 0=0 1=-1 2=384
436
+ Deconvolution deconvolution399 1 1 b398 b399 0=64 1=4 11=4 3=2 13=2 4=1 14=1 5=1 6=393216
437
+ Reshape reshape400 1 1 b399 b400 0=0 1=-1 11=4 2=64
438
+ Crop crop401 1 1 b400 b401 7=2 3=-233 4=-233 14=-233 5=-233
439
+ Reshape reshape402 1 1 b401 b402 0=0 1=-1 2=64
440
+ Split split_b402 1 2 b402 b402_0 b402_1
441
+ Pooling pooling403 1 1 b402_0 b403 0=1 4=1
442
+ InnerProduct innerproduct404 1 1 b403 b404 0=64 1=1 2=4096 9=4
443
+ BinaryOp binaryop405 2 1 b402_1 b404 b405 0=2
444
+ Reshape reshape406 1 1 b405 b406 0=0 1=-1 11=4 2=64
445
+ ReLU relu407 1 1 b406 out 0=2.000000e-01
tools/flavr/export.py CHANGED
@@ -4,8 +4,9 @@
4
  body in0 (4D: 3 channels x 4 frames x H x W, their mean taken away) -> out (64 x 4 x H x W)
5
  head in0 (256 x H x W: body's four times folded into channels, time major) -> out (3 x H x W, mean still away)
6
 
7
- The 3D layers are ncnn's Convolution3D, Deconvolution3D and Pooling3D; SE gating is a global average, an InnerProduct
8
- with a sigmoid and a per-channel multiplication. The weights load as tensors only.
 
9
  """
10
  import os
11
  import struct
@@ -17,6 +18,8 @@ import torch.nn as nn
17
 
18
  from flavr_net import BasicBlock, SEGating, load
19
 
 
 
20
 
21
  class Graph:
22
  """An ncnn network written layer by layer: a Split added where a blob feeds more than one layer."""
@@ -47,41 +50,97 @@ class Graph:
47
  def floats(self, b):
48
  self.weights += b.detach().float().numpy().astype("<f4").tobytes()
49
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
50
  def conv3d(self, x, conv, act=0):
51
  out_c, in_c, kd, kh, kw = conv.weight.shape
52
  sd, sh, sw = conv.stride
53
  pd, ph, pw = conv.padding
 
54
  bias = conv.bias is not None
55
- p = (f"0={out_c} 1={kw} 11={kh} 21={kd} 3={sw} 13={sh} 23={sd} 4={pw} 14={ph} 24={pd} "
56
- f"5={int(bias)} 6={conv.weight.numel()}")
 
 
 
 
 
 
 
57
  if act == 1:
58
  p += " 9=1"
59
- y = self.layer("Convolution3D", [x], p)
60
- self.half(conv.weight)
61
  if bias:
62
  self.floats(conv.bias)
63
- return y
64
 
65
  def deconv3d(self, x, conv):
66
  in_c, out_c, kd, kh, kw = conv.weight.shape
67
  sd, sh, sw = conv.stride
68
  pd, ph, pw = conv.padding
69
- p = (f"0={out_c} 1={kw} 11={kh} 21={kd} 3={sw} 13={sh} 23={sd} 4={pw} 14={ph} 24={pd} "
70
- f"5=1 6={conv.weight.numel()}")
71
- y = self.layer("Deconvolution3D", [x], p)
72
- # ncnn takes a transposed convolution's weights as out x in x kernel; PyTorch keeps them in x out x kernel.
73
- self.half(conv.weight.permute(1, 0, 2, 3, 4).contiguous())
 
 
 
 
74
  self.floats(conv.bias)
75
- return y
76
 
77
  def se(self, x, gating):
78
  conv = gating.attn_layer[0]
79
  c = conv.weight.shape[0]
80
- pooled = self.layer("Pooling3D", [x], "0=1 4=1")
 
 
81
  attention = self.layer("InnerProduct", [pooled], f"0={c} 1=1 2={c * c} 9=4")
82
  self.half(conv.weight.reshape(c, c))
83
  self.floats(conv.bias)
84
- return self.layer("BinaryOp", [x, attention], "0=2")
85
 
86
  def relu(self, x, slope=0.0):
87
  return self.layer("ReLU", [x], f"0={slope:.6e}" if slope else "")
 
4
  body in0 (4D: 3 channels x 4 frames x H x W, their mean taken away) -> out (64 x 4 x H x W)
5
  head in0 (256 x H x W: body's four times folded into channels, time major) -> out (3 x H x W, mean still away)
6
 
7
+ The 3D layers are written as 2D ones over the four frames (see Graph.conv3d), so ncnn runs them on its fast 2D paths
8
+ and its GPU, the weights stored once; SE gating is a global average, an InnerProduct with a sigmoid and a per-channel
9
+ multiplication. The weights load as tensors only.
10
  """
11
  import os
12
  import struct
 
18
 
19
  from flavr_net import BasicBlock, SEGating, load
20
 
21
+ FRAMES = 4
22
+
23
 
24
  class Graph:
25
  """An ncnn network written layer by layer: a Split added where a blob feeds more than one layer."""
 
50
  def floats(self, b):
51
  self.weights += b.detach().float().numpy().astype("<f4").tobytes()
52
 
53
+ # Crop's sizes not limited (its default, 0, would leave nothing): only the offsets cut.
54
+ WHOLE = "3=-233 4=-233 14=-233 5=-233"
55
+
56
+ # The 3D layers as 2D ones. A blob is 4D — width, height, the four frames (ncnn's depth), channels. For a layer
57
+ # with a temporal kernel of 3, the frames before and after each are stacked into its channels (zeros beyond the
58
+ # first and the last), so the time part of the kernel becomes channels of a 2D one; each frame gets zero rows below
59
+ # it, enough that the 2D kernel cannot reach from one frame into the next; the four frames are put one below the
60
+ # other as one tall picture for an ordinary 2D layer — ncnn's fast paths, its GPU among them, the weights stored
61
+ # once — and the result put back into frames, the rows added cut away.
62
+
63
+ def frames(self, x, first, count):
64
+ """Frames [first, first + count) of the 4D blob [x]."""
65
+ return self.layer("Crop", [x], f"13={first} 15={FRAMES - first - count} {Graph.WHOLE}")
66
+
67
+ def stacked(self, x, order):
68
+ """[x] with the frame before and after each in its channels: (prev, cur, next) or (next, cur, prev)."""
69
+ zero = self.layer("BinaryOp", [self.frames(x, 0, 1)], "0=2 1=1 2=0.000000e+00")
70
+ prev = self.layer("Concat", [zero, self.frames(x, 0, FRAMES - 1)], "0=1")
71
+ zero2 = self.layer("BinaryOp", [self.frames(x, 0, 1)], "0=2 1=1 2=0.000000e+00")
72
+ nxt = self.layer("Concat", [self.frames(x, 1, FRAMES - 1), zero2], "0=1")
73
+ parts = [prev, x, nxt] if order == "forward" else [nxt, x, prev]
74
+ return self.layer("Concat", parts, "0=0")
75
+
76
+ def tall(self, x, gap, channels):
77
+ """The 4D blob [x] ([channels] deep) as one tall 2D picture, [gap] zero rows below each frame."""
78
+ if gap:
79
+ x = self.layer("Padding", [x], f"0=0 1={gap} 2=0 3=0 4=0")
80
+ return self.layer("Reshape", [x], f"0=0 1=-1 2={channels}")
81
+
82
+ def untall(self, x, gap, channels):
83
+ """A tall 2D picture ([channels] deep) back into the four frames, the [gap] rows below each cut away."""
84
+ y = self.layer("Reshape", [x], f"0=0 1=-1 11={FRAMES} 2={channels}")
85
+ return self.layer("Crop", [y], f"7={gap} {Graph.WHOLE}") if gap else y
86
+
87
+ @staticmethod
88
+ def gap(pad, stride):
89
+ """Rows enough below each frame for a kernel reaching [pad] rows, a multiple of [stride]."""
90
+ g = pad
91
+ while g % stride:
92
+ g += 1
93
+ return g
94
+
95
  def conv3d(self, x, conv, act=0):
96
  out_c, in_c, kd, kh, kw = conv.weight.shape
97
  sd, sh, sw = conv.stride
98
  pd, ph, pw = conv.padding
99
+ assert sd == 1 and sh == sw and ph == pw and (kd, pd) in ((1, 0), (3, 1))
100
  bias = conv.bias is not None
101
+ w = conv.weight
102
+ if kd == 3:
103
+ x = self.stacked(x, "forward")
104
+ # Channel block k of the 2D kernel is time k of the 3D one: (out, 3, in, kh, kw).
105
+ w = w.permute(0, 2, 1, 3, 4).reshape(out_c, 3 * in_c, kh, kw)
106
+ else:
107
+ w = w.reshape(out_c, in_c, kh, kw)
108
+ gap = self.gap(ph, sh)
109
+ p = f"0={out_c} 1={kw} 11={kh} 3={sw} 13={sh} 4={pw} 14={ph} 5={int(bias)} 6={w.numel()}"
110
  if act == 1:
111
  p += " 9=1"
112
+ y = self.layer("Convolution", [self.tall(x, gap, w.shape[1])], p)
113
+ self.half(w)
114
  if bias:
115
  self.floats(conv.bias)
116
+ return self.untall(y, gap // sh, out_c)
117
 
118
  def deconv3d(self, x, conv):
119
  in_c, out_c, kd, kh, kw = conv.weight.shape
120
  sd, sh, sw = conv.stride
121
  pd, ph, pw = conv.padding
122
+ assert (kd, sd, pd) == (3, 1, 1) and sh == sw and ph == pw
123
+ # A transposed kernel takes frame t + 1 - k at time k: (next, cur, prev), channel block k of the 2D kernel.
124
+ x = self.stacked(x, "backward")
125
+ # ncnn takes a transposed kernel as out x in x kh x kw; PyTorch keeps it in x out x kernel.
126
+ w = conv.weight.permute(1, 2, 0, 3, 4).reshape(out_c, 3 * in_c, kh, kw)
127
+ gap = 1
128
+ p = f"0={out_c} 1={kw} 11={kh} 3={sw} 13={sh} 4={pw} 14={ph} 5=1 6={w.numel()}"
129
+ y = self.layer("Deconvolution", [self.tall(x, gap, 3 * in_c)], p)
130
+ self.half(w)
131
  self.floats(conv.bias)
132
+ return self.untall(y, gap * sh, out_c)
133
 
134
  def se(self, x, gating):
135
  conv = gating.attn_layer[0]
136
  c = conv.weight.shape[0]
137
+ # All frames are the same size, so the average over the tall picture is the average over the frames.
138
+ t = self.tall(x, 0, c)
139
+ pooled = self.layer("Pooling", [t], "0=1 4=1")
140
  attention = self.layer("InnerProduct", [pooled], f"0={c} 1=1 2={c * c} 9=4")
141
  self.half(conv.weight.reshape(c, c))
142
  self.floats(conv.bias)
143
+ return self.untall(self.layer("BinaryOp", [t, attention], "0=2"), 0, c)
144
 
145
  def relu(self, x, slope=0.0):
146
  return self.layer("ReLU", [x], f"0={slope:.6e}" if slope else "")