PyTorchモデルをONNXに変換

今回の試み

Rasbperry Piカメラで撮影している画像から人物を抽出して、人物が写っていることをMQTTを使ってイベントとしてIoT Platform側に通知として送るシステムを作ろうとしています。人物の抽出には、Intel NCS2の助けを借りようと思います。

この記事では、一旦学習済みモデルをONNXベースのモデルに変換することまで試してみました。実際にIntel NCS2で実施するために、諸々パラメータなど調整が今後必要になると考えています。

環境

PC側のスペックは以下の通りで、GPUありの環境で実施しています。

$ uname -a
Linux hajime-DAIV 5.3.0-51-generic #44~18.04.2-Ubuntu SMP Thu Apr 23 14:27:18 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux

$ lscpu
Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              12
On-line CPU(s) list: 0-11
Thread(s) per core:  2
Core(s) per socket:  6
Socket(s):           1
NUMA node(s):        1
Vendor ID:           GenuineIntel
CPU family:          6
Model:               158
Model name:          Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz
Stepping:            10
CPU MHz:             1015.929
CPU max MHz:         4100.0000
CPU min MHz:         800.0000
BogoMIPS:            4399.99
Virtualization:      VT-x
L1d cache:           32K
L1i cache:           32K
L2 cache:            256K
L3 cache:            9216K
NUMA node0 CPU(s):   0-11
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb invpcid_single pti ssbd ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx rdseed adx smap clflushopt intel_pt xsaveopt xsavec xgetbv1 xsaves dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp md_clear flush_l1d

$ nvidia-smi
Fri May  8 09:39:51 2020       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.82       Driver Version: 440.82       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce RTX 2060    Off  | 00000000:01:00.0 Off |                  N/A |
(以下略)

Raspberry Pi側はModel 4B(4GB)＋Intel NCS2です。

Python環境はAnacondaにPyTorch環境を作ってそちらで実行しました。

$  conda create -n pytorch python=3.7 anaconda
$  conda activate pytorch
$  conda install pytorch torchvision cudatoolkit=10.2 -c pytorch

この環境には、pytorch 1.5.0 py3.7_cuda10.2.89_cudnn7.6.5_0がインストールされています。

学習済モデルのダウンロードをONNXへの変換

Intel NCS2で推論を行うためには、学習済みモデルをIntel OpneVinoツールキットのモデル・オプティマイザを使って、同じくIntel OpneVinoツールキットの推論エンジンで実行出来るように変換する必要があります。変換できる学習済みモデルはTensorFlow、Caffee2、ONNXなどがあります。PyTorchモデルについては、このブログ記事を書いている時点(2020/5/9)では直接変換できないので、PyTorchの学習済みモデルを一旦、ONNXのモデルに変換してから、Intel OpneVinoツールキットを使ってIntel NCS2用で推論できるようにします。

ここで、学習済みモデルとしてはhttps://pytorch.org/hub/nvidia_deeplearningexamples_ssd/
を使います。COCOデータセットに対しての学習済みモデルとなります。モデルをロードするコードについても、こちらのサイトから拝借しました。

モデルのONNXの変換にあたっては、PyTorch公式のコードを拝借しました。

import torch

precision = 'fp32'
ssd_model = torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_ssd', model_math=precision)

dummy_input = torch.randn(10, 3, 300, 300, device='cpu')

# Providing input and output names sets the display names for values
# within the model's graph. Setting these does not change the semantics
# of the graph; it is only for readabilitya
#
# The inputs to the network consist of the flat list of inputs (i.e.
# the values you would pass to the forward() method) followed by the
# flat list of parameters. You can partially specify names, i.e. provide
# a list here shorter than the number of inputs to the model, and we will
# only set that subset of names, starting from the beginning.
input_names = [ "actual_input_1" ] + [ "learned_%d" % i for i in range(16) ]
output_names = [ "output1" ]

torch.onnx.export(ssd_model, dummy_input, "nvidia_ssd.onnx", verbose=True, input_names=input_names, output_names=output_names)

とりあえず、つつがなく実行され、ONNXのモデルが出力されました。モデルの詳細についても出力してくれています。

以下、実行した結果です。SSD300のネットワークが出力されているようですが、とても長く、ラズパイで動くのか不安です。

$  python convertTorch2ONNX.py 
Using cache found in /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub
graph(%actual_input_1 : Float(10, 3, 300, 300),
      %learned_0 : Float(64, 3, 7, 7),
      %learned_1 : Float(64),
      %learned_2 : Float(64),
      %learned_3 : Float(64),
      %learned_4 : Float(64),
      %learned_6 : Float(64, 64, 1, 1),
      %learned_7 : Float(64),
      %learned_8 : Float(64),
      %learned_9 : Float(64),
      %learned_10 : Float(64),
      %learned_12 : Float(64, 64, 3, 3),
      %learned_13 : Float(64),
      %learned_14 : Float(64),
      %learned_15 : Float(64),
      %feature_extractor.feature_extractor.4.0.bn2.running_var : Float(64),
      %feature_extractor.feature_extractor.4.0.conv3.weight : Float(256, 64, 1, 1),
      %feature_extractor.feature_extractor.4.0.bn3.weight : Float(256),
      %feature_extractor.feature_extractor.4.0.bn3.bias : Float(256),
      %feature_extractor.feature_extractor.4.0.bn3.running_mean : Float(256),
      %feature_extractor.feature_extractor.4.0.bn3.running_var : Float(256),
      %feature_extractor.feature_extractor.4.0.downsample.0.weight : Float(256, 64, 1, 1),
      %feature_extractor.feature_extractor.4.0.downsample.1.weight : Float(256),
      %feature_extractor.feature_extractor.4.0.downsample.1.bias : Float(256),
      %feature_extractor.feature_extractor.4.0.downsample.1.running_mean : Float(256),
      %feature_extractor.feature_extractor.4.0.downsample.1.running_var : Float(256),
      %feature_extractor.feature_extractor.4.1.conv1.weight : Float(64, 256, 1, 1),
      %feature_extractor.feature_extractor.4.1.bn1.weight : Float(64),
      %feature_extractor.feature_extractor.4.1.bn1.bias : Float(64),
      %feature_extractor.feature_extractor.4.1.bn1.running_mean : Float(64),
      %feature_extractor.feature_extractor.4.1.bn1.running_var : Float(64),
      %feature_extractor.feature_extractor.4.1.conv2.weight : Float(64, 64, 3, 3),
      %feature_extractor.feature_extractor.4.1.bn2.weight : Float(64),
      %feature_extractor.feature_extractor.4.1.bn2.bias : Float(64),
      %feature_extractor.feature_extractor.4.1.bn2.running_mean : Float(64),
      %feature_extractor.feature_extractor.4.1.bn2.running_var : Float(64),
      %feature_extractor.feature_extractor.4.1.conv3.weight : Float(256, 64, 1, 1),
      %feature_extractor.feature_extractor.4.1.bn3.weight : Float(256),
      %feature_extractor.feature_extractor.4.1.bn3.bias : Float(256),
      %feature_extractor.feature_extractor.4.1.bn3.running_mean : Float(256),
      %feature_extractor.feature_extractor.4.1.bn3.running_var : Float(256),
      %feature_extractor.feature_extractor.4.2.conv1.weight : Float(64, 256, 1, 1),
      %feature_extractor.feature_extractor.4.2.bn1.weight : Float(64),
      %feature_extractor.feature_extractor.4.2.bn1.bias : Float(64),
      %feature_extractor.feature_extractor.4.2.bn1.running_mean : Float(64),
      %feature_extractor.feature_extractor.4.2.bn1.running_var : Float(64),
      %feature_extractor.feature_extractor.4.2.conv2.weight : Float(64, 64, 3, 3),
      %feature_extractor.feature_extractor.4.2.bn2.weight : Float(64),
      %feature_extractor.feature_extractor.4.2.bn2.bias : Float(64),
      %feature_extractor.feature_extractor.4.2.bn2.running_mean : Float(64),
      %feature_extractor.feature_extractor.4.2.bn2.running_var : Float(64),
      %feature_extractor.feature_extractor.4.2.conv3.weight : Float(256, 64, 1, 1),
      %feature_extractor.feature_extractor.4.2.bn3.weight : Float(256),
      %feature_extractor.feature_extractor.4.2.bn3.bias : Float(256),
      %feature_extractor.feature_extractor.4.2.bn3.running_mean : Float(256),
      %feature_extractor.feature_extractor.4.2.bn3.running_var : Float(256),
      %feature_extractor.feature_extractor.5.0.conv1.weight : Float(128, 256, 1, 1),
      %feature_extractor.feature_extractor.5.0.bn1.weight : Float(128),
      %feature_extractor.feature_extractor.5.0.bn1.bias : Float(128),
      %feature_extractor.feature_extractor.5.0.bn1.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.0.bn1.running_var : Float(128),
      %feature_extractor.feature_extractor.5.0.conv2.weight : Float(128, 128, 3, 3),
      %feature_extractor.feature_extractor.5.0.bn2.weight : Float(128),
      %feature_extractor.feature_extractor.5.0.bn2.bias : Float(128),
      %feature_extractor.feature_extractor.5.0.bn2.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.0.bn2.running_var : Float(128),
      %feature_extractor.feature_extractor.5.0.conv3.weight : Float(512, 128, 1, 1),
      %feature_extractor.feature_extractor.5.0.bn3.weight : Float(512),
      %feature_extractor.feature_extractor.5.0.bn3.bias : Float(512),
      %feature_extractor.feature_extractor.5.0.bn3.running_mean : Float(512),
      %feature_extractor.feature_extractor.5.0.bn3.running_var : Float(512),
      %feature_extractor.feature_extractor.5.0.downsample.0.weight : Float(512, 256, 1, 1),
      %feature_extractor.feature_extractor.5.0.downsample.1.weight : Float(512),
      %feature_extractor.feature_extractor.5.0.downsample.1.bias : Float(512),
      %feature_extractor.feature_extractor.5.0.downsample.1.running_mean : Float(512),
      %feature_extractor.feature_extractor.5.0.downsample.1.running_var : Float(512),
      %feature_extractor.feature_extractor.5.1.conv1.weight : Float(128, 512, 1, 1),
      %feature_extractor.feature_extractor.5.1.bn1.weight : Float(128),
      %feature_extractor.feature_extractor.5.1.bn1.bias : Float(128),
      %feature_extractor.feature_extractor.5.1.bn1.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.1.bn1.running_var : Float(128),
      %feature_extractor.feature_extractor.5.1.conv2.weight : Float(128, 128, 3, 3),
      %feature_extractor.feature_extractor.5.1.bn2.weight : Float(128),
      %feature_extractor.feature_extractor.5.1.bn2.bias : Float(128),
      %feature_extractor.feature_extractor.5.1.bn2.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.1.bn2.running_var : Float(128),
      %feature_extractor.feature_extractor.5.1.conv3.weight : Float(512, 128, 1, 1),
      %feature_extractor.feature_extractor.5.1.bn3.weight : Float(512),
      %feature_extractor.feature_extractor.5.1.bn3.bias : Float(512),
      %feature_extractor.feature_extractor.5.1.bn3.running_mean : Float(512),
      %feature_extractor.feature_extractor.5.1.bn3.running_var : Float(512),
      %feature_extractor.feature_extractor.5.2.conv1.weight : Float(128, 512, 1, 1),
      %feature_extractor.feature_extractor.5.2.bn1.weight : Float(128),
      %feature_extractor.feature_extractor.5.2.bn1.bias : Float(128),
      %feature_extractor.feature_extractor.5.2.bn1.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.2.bn1.running_var : Float(128),
      %feature_extractor.feature_extractor.5.2.conv2.weight : Float(128, 128, 3, 3),
      %feature_extractor.feature_extractor.5.2.bn2.weight : Float(128),
      %feature_extractor.feature_extractor.5.2.bn2.bias : Float(128),
      %feature_extractor.feature_extractor.5.2.bn2.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.2.bn2.running_var : Float(128),
      %feature_extractor.feature_extractor.5.2.conv3.weight : Float(512, 128, 1, 1),
      %feature_extractor.feature_extractor.5.2.bn3.weight : Float(512),
      %feature_extractor.feature_extractor.5.2.bn3.bias : Float(512),
      %feature_extractor.feature_extractor.5.2.bn3.running_mean : Float(512),
      %feature_extractor.feature_extractor.5.2.bn3.running_var : Float(512),
      %feature_extractor.feature_extractor.5.3.conv1.weight : Float(128, 512, 1, 1),
      %feature_extractor.feature_extractor.5.3.bn1.weight : Float(128),
      %feature_extractor.feature_extractor.5.3.bn1.bias : Float(128),
      %feature_extractor.feature_extractor.5.3.bn1.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.3.bn1.running_var : Float(128),
      %feature_extractor.feature_extractor.5.3.conv2.weight : Float(128, 128, 3, 3),
      %feature_extractor.feature_extractor.5.3.bn2.weight : Float(128),
      %feature_extractor.feature_extractor.5.3.bn2.bias : Float(128),
      %feature_extractor.feature_extractor.5.3.bn2.running_mean : Float(128),
      %feature_extractor.feature_extractor.5.3.bn2.running_var : Float(128),
      %feature_extractor.feature_extractor.5.3.conv3.weight : Float(512, 128, 1, 1),
      %feature_extractor.feature_extractor.5.3.bn3.weight : Float(512),
      %feature_extractor.feature_extractor.5.3.bn3.bias : Float(512),
      %feature_extractor.feature_extractor.5.3.bn3.running_mean : Float(512),
      %feature_extractor.feature_extractor.5.3.bn3.running_var : Float(512),
      %feature_extractor.feature_extractor.6.0.conv1.weight : Float(256, 512, 1, 1),
      %feature_extractor.feature_extractor.6.0.bn1.weight : Float(256),
      %feature_extractor.feature_extractor.6.0.bn1.bias : Float(256),
      %feature_extractor.feature_extractor.6.0.bn1.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.0.bn1.running_var : Float(256),
      %feature_extractor.feature_extractor.6.0.conv2.weight : Float(256, 256, 3, 3),
      %feature_extractor.feature_extractor.6.0.bn2.weight : Float(256),
      %feature_extractor.feature_extractor.6.0.bn2.bias : Float(256),
      %feature_extractor.feature_extractor.6.0.bn2.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.0.bn2.running_var : Float(256),
      %feature_extractor.feature_extractor.6.0.conv3.weight : Float(1024, 256, 1, 1),
      %feature_extractor.feature_extractor.6.0.bn3.weight : Float(1024),
      %feature_extractor.feature_extractor.6.0.bn3.bias : Float(1024),
      %feature_extractor.feature_extractor.6.0.bn3.running_mean : Float(1024),
      %feature_extractor.feature_extractor.6.0.bn3.running_var : Float(1024),
      %feature_extractor.feature_extractor.6.0.downsample.0.weight : Float(1024, 512, 1, 1),
      %feature_extractor.feature_extractor.6.0.downsample.1.weight : Float(1024),
      %feature_extractor.feature_extractor.6.0.downsample.1.bias : Float(1024),
      %feature_extractor.feature_extractor.6.0.downsample.1.running_mean : Float(1024),
      %feature_extractor.feature_extractor.6.0.downsample.1.running_var : Float(1024),
      %feature_extractor.feature_extractor.6.1.conv1.weight : Float(256, 1024, 1, 1),
      %feature_extractor.feature_extractor.6.1.bn1.weight : Float(256),
      %feature_extractor.feature_extractor.6.1.bn1.bias : Float(256),
      %feature_extractor.feature_extractor.6.1.bn1.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.1.bn1.running_var : Float(256),
      %feature_extractor.feature_extractor.6.1.conv2.weight : Float(256, 256, 3, 3),
      %feature_extractor.feature_extractor.6.1.bn2.weight : Float(256),
      %feature_extractor.feature_extractor.6.1.bn2.bias : Float(256),
      %feature_extractor.feature_extractor.6.1.bn2.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.1.bn2.running_var : Float(256),
      %feature_extractor.feature_extractor.6.1.conv3.weight : Float(1024, 256, 1, 1),
      %feature_extractor.feature_extractor.6.1.bn3.weight : Float(1024),
      %feature_extractor.feature_extractor.6.1.bn3.bias : Float(1024),
      %feature_extractor.feature_extractor.6.1.bn3.running_mean : Float(1024),
      %feature_extractor.feature_extractor.6.1.bn3.running_var : Float(1024),
      %feature_extractor.feature_extractor.6.2.conv1.weight : Float(256, 1024, 1, 1),
      %feature_extractor.feature_extractor.6.2.bn1.weight : Float(256),
      %feature_extractor.feature_extractor.6.2.bn1.bias : Float(256),
      %feature_extractor.feature_extractor.6.2.bn1.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.2.bn1.running_var : Float(256),
      %feature_extractor.feature_extractor.6.2.conv2.weight : Float(256, 256, 3, 3),
      %feature_extractor.feature_extractor.6.2.bn2.weight : Float(256),
      %feature_extractor.feature_extractor.6.2.bn2.bias : Float(256),
      %feature_extractor.feature_extractor.6.2.bn2.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.2.bn2.running_var : Float(256),
      %feature_extractor.feature_extractor.6.2.conv3.weight : Float(1024, 256, 1, 1),
      %feature_extractor.feature_extractor.6.2.bn3.weight : Float(1024),
      %feature_extractor.feature_extractor.6.2.bn3.bias : Float(1024),
      %feature_extractor.feature_extractor.6.2.bn3.running_mean : Float(1024),
      %feature_extractor.feature_extractor.6.2.bn3.running_var : Float(1024),
      %feature_extractor.feature_extractor.6.3.conv1.weight : Float(256, 1024, 1, 1),
      %feature_extractor.feature_extractor.6.3.bn1.weight : Float(256),
      %feature_extractor.feature_extractor.6.3.bn1.bias : Float(256),
      %feature_extractor.feature_extractor.6.3.bn1.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.3.bn1.running_var : Float(256),
      %feature_extractor.feature_extractor.6.3.conv2.weight : Float(256, 256, 3, 3),
      %feature_extractor.feature_extractor.6.3.bn2.weight : Float(256),
      %feature_extractor.feature_extractor.6.3.bn2.bias : Float(256),
      %feature_extractor.feature_extractor.6.3.bn2.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.3.bn2.running_var : Float(256),
      %feature_extractor.feature_extractor.6.3.conv3.weight : Float(1024, 256, 1, 1),
      %feature_extractor.feature_extractor.6.3.bn3.weight : Float(1024),
      %feature_extractor.feature_extractor.6.3.bn3.bias : Float(1024),
      %feature_extractor.feature_extractor.6.3.bn3.running_mean : Float(1024),
      %feature_extractor.feature_extractor.6.3.bn3.running_var : Float(1024),
      %feature_extractor.feature_extractor.6.4.conv1.weight : Float(256, 1024, 1, 1),
      %feature_extractor.feature_extractor.6.4.bn1.weight : Float(256),
      %feature_extractor.feature_extractor.6.4.bn1.bias : Float(256),
      %feature_extractor.feature_extractor.6.4.bn1.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.4.bn1.running_var : Float(256),
      %feature_extractor.feature_extractor.6.4.conv2.weight : Float(256, 256, 3, 3),
      %feature_extractor.feature_extractor.6.4.bn2.weight : Float(256),
      %feature_extractor.feature_extractor.6.4.bn2.bias : Float(256),
      %feature_extractor.feature_extractor.6.4.bn2.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.4.bn2.running_var : Float(256),
      %feature_extractor.feature_extractor.6.4.conv3.weight : Float(1024, 256, 1, 1),
      %feature_extractor.feature_extractor.6.4.bn3.weight : Float(1024),
      %feature_extractor.feature_extractor.6.4.bn3.bias : Float(1024),
      %feature_extractor.feature_extractor.6.4.bn3.running_mean : Float(1024),
      %feature_extractor.feature_extractor.6.4.bn3.running_var : Float(1024),
      %feature_extractor.feature_extractor.6.5.conv1.weight : Float(256, 1024, 1, 1),
      %feature_extractor.feature_extractor.6.5.bn1.weight : Float(256),
      %feature_extractor.feature_extractor.6.5.bn1.bias : Float(256),
      %feature_extractor.feature_extractor.6.5.bn1.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.5.bn1.running_var : Float(256),
      %feature_extractor.feature_extractor.6.5.conv2.weight : Float(256, 256, 3, 3),
      %feature_extractor.feature_extractor.6.5.bn2.weight : Float(256),
      %feature_extractor.feature_extractor.6.5.bn2.bias : Float(256),
      %feature_extractor.feature_extractor.6.5.bn2.running_mean : Float(256),
      %feature_extractor.feature_extractor.6.5.bn2.running_var : Float(256),
      %feature_extractor.feature_extractor.6.5.conv3.weight : Float(1024, 256, 1, 1),
      %feature_extractor.feature_extractor.6.5.bn3.weight : Float(1024),
      %feature_extractor.feature_extractor.6.5.bn3.bias : Float(1024),
      %feature_extractor.feature_extractor.6.5.bn3.running_mean : Float(1024),
      %feature_extractor.feature_extractor.6.5.bn3.running_var : Float(1024),
      %additional_blocks.0.0.weight : Float(256, 1024, 1, 1),
      %additional_blocks.0.1.weight : Float(256),
      %additional_blocks.0.1.bias : Float(256),
      %additional_blocks.0.1.running_mean : Float(256),
      %additional_blocks.0.1.running_var : Float(256),
      %additional_blocks.0.3.weight : Float(512, 256, 3, 3),
      %additional_blocks.0.4.weight : Float(512),
      %additional_blocks.0.4.bias : Float(512),
      %additional_blocks.0.4.running_mean : Float(512),
      %additional_blocks.0.4.running_var : Float(512),
      %additional_blocks.1.0.weight : Float(256, 512, 1, 1),
      %additional_blocks.1.1.weight : Float(256),
      %additional_blocks.1.1.bias : Float(256),
      %additional_blocks.1.1.running_mean : Float(256),
      %additional_blocks.1.1.running_var : Float(256),
      %additional_blocks.1.3.weight : Float(512, 256, 3, 3),
      %additional_blocks.1.4.weight : Float(512),
      %additional_blocks.1.4.bias : Float(512),
      %additional_blocks.1.4.running_mean : Float(512),
      %additional_blocks.1.4.running_var : Float(512),
      %additional_blocks.2.0.weight : Float(128, 512, 1, 1),
      %additional_blocks.2.1.weight : Float(128),
      %additional_blocks.2.1.bias : Float(128),
      %additional_blocks.2.1.running_mean : Float(128),
      %additional_blocks.2.1.running_var : Float(128),
      %additional_blocks.2.3.weight : Float(256, 128, 3, 3),
      %additional_blocks.2.4.weight : Float(256),
      %additional_blocks.2.4.bias : Float(256),
      %additional_blocks.2.4.running_mean : Float(256),
      %additional_blocks.2.4.running_var : Float(256),
      %additional_blocks.3.0.weight : Float(128, 256, 1, 1),
      %additional_blocks.3.1.weight : Float(128),
      %additional_blocks.3.1.bias : Float(128),
      %additional_blocks.3.1.running_mean : Float(128),
      %additional_blocks.3.1.running_var : Float(128),
      %additional_blocks.3.3.weight : Float(256, 128, 3, 3),
      %additional_blocks.3.4.weight : Float(256),
      %additional_blocks.3.4.bias : Float(256),
      %additional_blocks.3.4.running_mean : Float(256),
      %additional_blocks.3.4.running_var : Float(256),
      %additional_blocks.4.0.weight : Float(128, 256, 1, 1),
      %additional_blocks.4.1.weight : Float(128),
      %additional_blocks.4.1.bias : Float(128),
      %additional_blocks.4.1.running_mean : Float(128),
      %additional_blocks.4.1.running_var : Float(128),
      %additional_blocks.4.3.weight : Float(256, 128, 3, 3),
      %additional_blocks.4.4.weight : Float(256),
      %additional_blocks.4.4.bias : Float(256),
      %additional_blocks.4.4.running_mean : Float(256),
      %additional_blocks.4.4.running_var : Float(256),
      %loc.0.weight : Float(16, 1024, 3, 3),
      %loc.0.bias : Float(16),
      %loc.1.weight : Float(24, 512, 3, 3),
      %loc.1.bias : Float(24),
      %loc.2.weight : Float(24, 512, 3, 3),
      %loc.2.bias : Float(24),
      %loc.3.weight : Float(24, 256, 3, 3),
      %loc.3.bias : Float(24),
      %loc.4.weight : Float(16, 256, 3, 3),
      %loc.4.bias : Float(16),
      %loc.5.weight : Float(16, 256, 3, 3),
      %loc.5.bias : Float(16),
      %conf.0.weight : Float(324, 1024, 3, 3),
      %conf.0.bias : Float(324),
      %conf.1.weight : Float(486, 512, 3, 3),
      %conf.1.bias : Float(486),
      %conf.2.weight : Float(486, 512, 3, 3),
      %conf.2.bias : Float(486),
      %conf.3.weight : Float(486, 256, 3, 3),
      %conf.3.bias : Float(486),
      %conf.4.weight : Float(324, 256, 3, 3),
      %conf.4.bias : Float(324),
      %conf.5.weight : Float(324, 256, 3, 3),
      %conf.5.bias : Float(324),
      %647 : Long(1),
      %648 : Long(1),
      %649 : Long(1),
      %650 : Long(1),
      %651 : Long(1),
      %652 : Long(1),
      %653 : Long(1),
      %654 : Long(1),
      %655 : Long(1),
      %656 : Long(1),
      %657 : Long(1),
      %658 : Long(1),
      %659 : Long(1),
      %660 : Long(1),
      %661 : Long(1),
      %662 : Long(1),
      %663 : Long(1),
      %664 : Long(1),
      %665 : Long(1),
      %666 : Long(1),
      %667 : Long(1),
      %668 : Long(1),
      %669 : Long(1),
      %670 : Long(1)):
  %343 : Float(10, 64, 150, 150) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[7, 7], pads=[3, 3, 3, 3], strides=[2, 2]](%actual_input_1, %learned_0) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %344 : Float(10, 64, 150, 150) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%343, %learned_1, %learned_2, %learned_3, %learned_4) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %345 : Float(10, 64, 150, 150) = onnx::Relu(%344) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %346 : Float(10, 64, 75, 75) = onnx::MaxPool[kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%345) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:539:0
  %347 : Float(10, 64, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%346, %learned_6) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %348 : Float(10, 64, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%347, %learned_7, %learned_8, %learned_9, %learned_10) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %349 : Float(10, 64, 75, 75) = onnx::Relu(%348) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %350 : Float(10, 64, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%349, %learned_12) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %351 : Float(10, 64, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%350, %learned_13, %learned_14, %learned_15, %feature_extractor.feature_extractor.4.0.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %352 : Float(10, 64, 75, 75) = onnx::Relu(%351) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %353 : Float(10, 256, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%352, %feature_extractor.feature_extractor.4.0.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %354 : Float(10, 256, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%353, %feature_extractor.feature_extractor.4.0.bn3.weight, %feature_extractor.feature_extractor.4.0.bn3.bias, %feature_extractor.feature_extractor.4.0.bn3.running_mean, %feature_extractor.feature_extractor.4.0.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %355 : Float(10, 256, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%346, %feature_extractor.feature_extractor.4.0.downsample.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %356 : Float(10, 256, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%355, %feature_extractor.feature_extractor.4.0.downsample.1.weight, %feature_extractor.feature_extractor.4.0.downsample.1.bias, %feature_extractor.feature_extractor.4.0.downsample.1.running_mean, %feature_extractor.feature_extractor.4.0.downsample.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %357 : Float(10, 256, 75, 75) = onnx::Add(%354, %356)
  %358 : Float(10, 256, 75, 75) = onnx::Relu(%357) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %359 : Float(10, 64, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%358, %feature_extractor.feature_extractor.4.1.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %360 : Float(10, 64, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%359, %feature_extractor.feature_extractor.4.1.bn1.weight, %feature_extractor.feature_extractor.4.1.bn1.bias, %feature_extractor.feature_extractor.4.1.bn1.running_mean, %feature_extractor.feature_extractor.4.1.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %361 : Float(10, 64, 75, 75) = onnx::Relu(%360) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %362 : Float(10, 64, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%361, %feature_extractor.feature_extractor.4.1.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %363 : Float(10, 64, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%362, %feature_extractor.feature_extractor.4.1.bn2.weight, %feature_extractor.feature_extractor.4.1.bn2.bias, %feature_extractor.feature_extractor.4.1.bn2.running_mean, %feature_extractor.feature_extractor.4.1.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %364 : Float(10, 64, 75, 75) = onnx::Relu(%363) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %365 : Float(10, 256, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%364, %feature_extractor.feature_extractor.4.1.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %366 : Float(10, 256, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%365, %feature_extractor.feature_extractor.4.1.bn3.weight, %feature_extractor.feature_extractor.4.1.bn3.bias, %feature_extractor.feature_extractor.4.1.bn3.running_mean, %feature_extractor.feature_extractor.4.1.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %367 : Float(10, 256, 75, 75) = onnx::Add(%366, %358)
  %368 : Float(10, 256, 75, 75) = onnx::Relu(%367) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %369 : Float(10, 64, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%368, %feature_extractor.feature_extractor.4.2.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %370 : Float(10, 64, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%369, %feature_extractor.feature_extractor.4.2.bn1.weight, %feature_extractor.feature_extractor.4.2.bn1.bias, %feature_extractor.feature_extractor.4.2.bn1.running_mean, %feature_extractor.feature_extractor.4.2.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %371 : Float(10, 64, 75, 75) = onnx::Relu(%370) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %372 : Float(10, 64, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%371, %feature_extractor.feature_extractor.4.2.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %373 : Float(10, 64, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%372, %feature_extractor.feature_extractor.4.2.bn2.weight, %feature_extractor.feature_extractor.4.2.bn2.bias, %feature_extractor.feature_extractor.4.2.bn2.running_mean, %feature_extractor.feature_extractor.4.2.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %374 : Float(10, 64, 75, 75) = onnx::Relu(%373) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %375 : Float(10, 256, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%374, %feature_extractor.feature_extractor.4.2.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %376 : Float(10, 256, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%375, %feature_extractor.feature_extractor.4.2.bn3.weight, %feature_extractor.feature_extractor.4.2.bn3.bias, %feature_extractor.feature_extractor.4.2.bn3.running_mean, %feature_extractor.feature_extractor.4.2.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %377 : Float(10, 256, 75, 75) = onnx::Add(%376, %368)
  %378 : Float(10, 256, 75, 75) = onnx::Relu(%377) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %379 : Float(10, 128, 75, 75) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%378, %feature_extractor.feature_extractor.5.0.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %380 : Float(10, 128, 75, 75) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%379, %feature_extractor.feature_extractor.5.0.bn1.weight, %feature_extractor.feature_extractor.5.0.bn1.bias, %feature_extractor.feature_extractor.5.0.bn1.running_mean, %feature_extractor.feature_extractor.5.0.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %381 : Float(10, 128, 75, 75) = onnx::Relu(%380) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %382 : Float(10, 128, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%381, %feature_extractor.feature_extractor.5.0.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %383 : Float(10, 128, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%382, %feature_extractor.feature_extractor.5.0.bn2.weight, %feature_extractor.feature_extractor.5.0.bn2.bias, %feature_extractor.feature_extractor.5.0.bn2.running_mean, %feature_extractor.feature_extractor.5.0.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %384 : Float(10, 128, 38, 38) = onnx::Relu(%383) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %385 : Float(10, 512, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%384, %feature_extractor.feature_extractor.5.0.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %386 : Float(10, 512, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%385, %feature_extractor.feature_extractor.5.0.bn3.weight, %feature_extractor.feature_extractor.5.0.bn3.bias, %feature_extractor.feature_extractor.5.0.bn3.running_mean, %feature_extractor.feature_extractor.5.0.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %387 : Float(10, 512, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[2, 2]](%378, %feature_extractor.feature_extractor.5.0.downsample.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %388 : Float(10, 512, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%387, %feature_extractor.feature_extractor.5.0.downsample.1.weight, %feature_extractor.feature_extractor.5.0.downsample.1.bias, %feature_extractor.feature_extractor.5.0.downsample.1.running_mean, %feature_extractor.feature_extractor.5.0.downsample.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %389 : Float(10, 512, 38, 38) = onnx::Add(%386, %388)
  %390 : Float(10, 512, 38, 38) = onnx::Relu(%389) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %391 : Float(10, 128, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%390, %feature_extractor.feature_extractor.5.1.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %392 : Float(10, 128, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%391, %feature_extractor.feature_extractor.5.1.bn1.weight, %feature_extractor.feature_extractor.5.1.bn1.bias, %feature_extractor.feature_extractor.5.1.bn1.running_mean, %feature_extractor.feature_extractor.5.1.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %393 : Float(10, 128, 38, 38) = onnx::Relu(%392) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %394 : Float(10, 128, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%393, %feature_extractor.feature_extractor.5.1.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %395 : Float(10, 128, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%394, %feature_extractor.feature_extractor.5.1.bn2.weight, %feature_extractor.feature_extractor.5.1.bn2.bias, %feature_extractor.feature_extractor.5.1.bn2.running_mean, %feature_extractor.feature_extractor.5.1.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %396 : Float(10, 128, 38, 38) = onnx::Relu(%395) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %397 : Float(10, 512, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%396, %feature_extractor.feature_extractor.5.1.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %398 : Float(10, 512, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%397, %feature_extractor.feature_extractor.5.1.bn3.weight, %feature_extractor.feature_extractor.5.1.bn3.bias, %feature_extractor.feature_extractor.5.1.bn3.running_mean, %feature_extractor.feature_extractor.5.1.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %399 : Float(10, 512, 38, 38) = onnx::Add(%398, %390)
  %400 : Float(10, 512, 38, 38) = onnx::Relu(%399) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %401 : Float(10, 128, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%400, %feature_extractor.feature_extractor.5.2.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %402 : Float(10, 128, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%401, %feature_extractor.feature_extractor.5.2.bn1.weight, %feature_extractor.feature_extractor.5.2.bn1.bias, %feature_extractor.feature_extractor.5.2.bn1.running_mean, %feature_extractor.feature_extractor.5.2.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %403 : Float(10, 128, 38, 38) = onnx::Relu(%402) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %404 : Float(10, 128, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%403, %feature_extractor.feature_extractor.5.2.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %405 : Float(10, 128, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%404, %feature_extractor.feature_extractor.5.2.bn2.weight, %feature_extractor.feature_extractor.5.2.bn2.bias, %feature_extractor.feature_extractor.5.2.bn2.running_mean, %feature_extractor.feature_extractor.5.2.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %406 : Float(10, 128, 38, 38) = onnx::Relu(%405) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %407 : Float(10, 512, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%406, %feature_extractor.feature_extractor.5.2.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %408 : Float(10, 512, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%407, %feature_extractor.feature_extractor.5.2.bn3.weight, %feature_extractor.feature_extractor.5.2.bn3.bias, %feature_extractor.feature_extractor.5.2.bn3.running_mean, %feature_extractor.feature_extractor.5.2.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %409 : Float(10, 512, 38, 38) = onnx::Add(%408, %400)
  %410 : Float(10, 512, 38, 38) = onnx::Relu(%409) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %411 : Float(10, 128, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%410, %feature_extractor.feature_extractor.5.3.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %412 : Float(10, 128, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%411, %feature_extractor.feature_extractor.5.3.bn1.weight, %feature_extractor.feature_extractor.5.3.bn1.bias, %feature_extractor.feature_extractor.5.3.bn1.running_mean, %feature_extractor.feature_extractor.5.3.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %413 : Float(10, 128, 38, 38) = onnx::Relu(%412) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %414 : Float(10, 128, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%413, %feature_extractor.feature_extractor.5.3.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %415 : Float(10, 128, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%414, %feature_extractor.feature_extractor.5.3.bn2.weight, %feature_extractor.feature_extractor.5.3.bn2.bias, %feature_extractor.feature_extractor.5.3.bn2.running_mean, %feature_extractor.feature_extractor.5.3.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %416 : Float(10, 128, 38, 38) = onnx::Relu(%415) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %417 : Float(10, 512, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%416, %feature_extractor.feature_extractor.5.3.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %418 : Float(10, 512, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%417, %feature_extractor.feature_extractor.5.3.bn3.weight, %feature_extractor.feature_extractor.5.3.bn3.bias, %feature_extractor.feature_extractor.5.3.bn3.running_mean, %feature_extractor.feature_extractor.5.3.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %419 : Float(10, 512, 38, 38) = onnx::Add(%418, %410)
  %420 : Float(10, 512, 38, 38) = onnx::Relu(%419) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %421 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%420, %feature_extractor.feature_extractor.6.0.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %422 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%421, %feature_extractor.feature_extractor.6.0.bn1.weight, %feature_extractor.feature_extractor.6.0.bn1.bias, %feature_extractor.feature_extractor.6.0.bn1.running_mean, %feature_extractor.feature_extractor.6.0.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %423 : Float(10, 256, 38, 38) = onnx::Relu(%422) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %424 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%423, %feature_extractor.feature_extractor.6.0.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %425 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%424, %feature_extractor.feature_extractor.6.0.bn2.weight, %feature_extractor.feature_extractor.6.0.bn2.bias, %feature_extractor.feature_extractor.6.0.bn2.running_mean, %feature_extractor.feature_extractor.6.0.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %426 : Float(10, 256, 38, 38) = onnx::Relu(%425) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %427 : Float(10, 1024, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%426, %feature_extractor.feature_extractor.6.0.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %428 : Float(10, 1024, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%427, %feature_extractor.feature_extractor.6.0.bn3.weight, %feature_extractor.feature_extractor.6.0.bn3.bias, %feature_extractor.feature_extractor.6.0.bn3.running_mean, %feature_extractor.feature_extractor.6.0.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %429 : Float(10, 1024, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%420, %feature_extractor.feature_extractor.6.0.downsample.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %430 : Float(10, 1024, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%429, %feature_extractor.feature_extractor.6.0.downsample.1.weight, %feature_extractor.feature_extractor.6.0.downsample.1.bias, %feature_extractor.feature_extractor.6.0.downsample.1.running_mean, %feature_extractor.feature_extractor.6.0.downsample.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %431 : Float(10, 1024, 38, 38) = onnx::Add(%428, %430)
  %432 : Float(10, 1024, 38, 38) = onnx::Relu(%431) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %433 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%432, %feature_extractor.feature_extractor.6.1.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %434 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%433, %feature_extractor.feature_extractor.6.1.bn1.weight, %feature_extractor.feature_extractor.6.1.bn1.bias, %feature_extractor.feature_extractor.6.1.bn1.running_mean, %feature_extractor.feature_extractor.6.1.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %435 : Float(10, 256, 38, 38) = onnx::Relu(%434) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %436 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%435, %feature_extractor.feature_extractor.6.1.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %437 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%436, %feature_extractor.feature_extractor.6.1.bn2.weight, %feature_extractor.feature_extractor.6.1.bn2.bias, %feature_extractor.feature_extractor.6.1.bn2.running_mean, %feature_extractor.feature_extractor.6.1.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %438 : Float(10, 256, 38, 38) = onnx::Relu(%437) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %439 : Float(10, 1024, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%438, %feature_extractor.feature_extractor.6.1.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %440 : Float(10, 1024, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%439, %feature_extractor.feature_extractor.6.1.bn3.weight, %feature_extractor.feature_extractor.6.1.bn3.bias, %feature_extractor.feature_extractor.6.1.bn3.running_mean, %feature_extractor.feature_extractor.6.1.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %441 : Float(10, 1024, 38, 38) = onnx::Add(%440, %432)
  %442 : Float(10, 1024, 38, 38) = onnx::Relu(%441) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %443 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%442, %feature_extractor.feature_extractor.6.2.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %444 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%443, %feature_extractor.feature_extractor.6.2.bn1.weight, %feature_extractor.feature_extractor.6.2.bn1.bias, %feature_extractor.feature_extractor.6.2.bn1.running_mean, %feature_extractor.feature_extractor.6.2.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %445 : Float(10, 256, 38, 38) = onnx::Relu(%444) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %446 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%445, %feature_extractor.feature_extractor.6.2.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %447 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%446, %feature_extractor.feature_extractor.6.2.bn2.weight, %feature_extractor.feature_extractor.6.2.bn2.bias, %feature_extractor.feature_extractor.6.2.bn2.running_mean, %feature_extractor.feature_extractor.6.2.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %448 : Float(10, 256, 38, 38) = onnx::Relu(%447) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %449 : Float(10, 1024, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%448, %feature_extractor.feature_extractor.6.2.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %450 : Float(10, 1024, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%449, %feature_extractor.feature_extractor.6.2.bn3.weight, %feature_extractor.feature_extractor.6.2.bn3.bias, %feature_extractor.feature_extractor.6.2.bn3.running_mean, %feature_extractor.feature_extractor.6.2.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %451 : Float(10, 1024, 38, 38) = onnx::Add(%450, %442)
  %452 : Float(10, 1024, 38, 38) = onnx::Relu(%451) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %453 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%452, %feature_extractor.feature_extractor.6.3.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %454 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%453, %feature_extractor.feature_extractor.6.3.bn1.weight, %feature_extractor.feature_extractor.6.3.bn1.bias, %feature_extractor.feature_extractor.6.3.bn1.running_mean, %feature_extractor.feature_extractor.6.3.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %455 : Float(10, 256, 38, 38) = onnx::Relu(%454) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %456 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%455, %feature_extractor.feature_extractor.6.3.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %457 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%456, %feature_extractor.feature_extractor.6.3.bn2.weight, %feature_extractor.feature_extractor.6.3.bn2.bias, %feature_extractor.feature_extractor.6.3.bn2.running_mean, %feature_extractor.feature_extractor.6.3.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %458 : Float(10, 256, 38, 38) = onnx::Relu(%457) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %459 : Float(10, 1024, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%458, %feature_extractor.feature_extractor.6.3.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %460 : Float(10, 1024, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%459, %feature_extractor.feature_extractor.6.3.bn3.weight, %feature_extractor.feature_extractor.6.3.bn3.bias, %feature_extractor.feature_extractor.6.3.bn3.running_mean, %feature_extractor.feature_extractor.6.3.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %461 : Float(10, 1024, 38, 38) = onnx::Add(%460, %452)
  %462 : Float(10, 1024, 38, 38) = onnx::Relu(%461) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %463 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%462, %feature_extractor.feature_extractor.6.4.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %464 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%463, %feature_extractor.feature_extractor.6.4.bn1.weight, %feature_extractor.feature_extractor.6.4.bn1.bias, %feature_extractor.feature_extractor.6.4.bn1.running_mean, %feature_extractor.feature_extractor.6.4.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %465 : Float(10, 256, 38, 38) = onnx::Relu(%464) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %466 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%465, %feature_extractor.feature_extractor.6.4.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %467 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%466, %feature_extractor.feature_extractor.6.4.bn2.weight, %feature_extractor.feature_extractor.6.4.bn2.bias, %feature_extractor.feature_extractor.6.4.bn2.running_mean, %feature_extractor.feature_extractor.6.4.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %468 : Float(10, 256, 38, 38) = onnx::Relu(%467) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %469 : Float(10, 1024, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%468, %feature_extractor.feature_extractor.6.4.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %470 : Float(10, 1024, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%469, %feature_extractor.feature_extractor.6.4.bn3.weight, %feature_extractor.feature_extractor.6.4.bn3.bias, %feature_extractor.feature_extractor.6.4.bn3.running_mean, %feature_extractor.feature_extractor.6.4.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %471 : Float(10, 1024, 38, 38) = onnx::Add(%470, %462)
  %472 : Float(10, 1024, 38, 38) = onnx::Relu(%471) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %473 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%472, %feature_extractor.feature_extractor.6.5.conv1.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %474 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%473, %feature_extractor.feature_extractor.6.5.bn1.weight, %feature_extractor.feature_extractor.6.5.bn1.bias, %feature_extractor.feature_extractor.6.5.bn1.running_mean, %feature_extractor.feature_extractor.6.5.bn1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %475 : Float(10, 256, 38, 38) = onnx::Relu(%474) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %476 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%475, %feature_extractor.feature_extractor.6.5.conv2.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %477 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%476, %feature_extractor.feature_extractor.6.5.bn2.weight, %feature_extractor.feature_extractor.6.5.bn2.bias, %feature_extractor.feature_extractor.6.5.bn2.running_mean, %feature_extractor.feature_extractor.6.5.bn2.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %478 : Float(10, 256, 38, 38) = onnx::Relu(%477) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %479 : Float(10, 1024, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%478, %feature_extractor.feature_extractor.6.5.conv3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %480 : Float(10, 1024, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%479, %feature_extractor.feature_extractor.6.5.bn3.weight, %feature_extractor.feature_extractor.6.5.bn3.bias, %feature_extractor.feature_extractor.6.5.bn3.running_mean, %feature_extractor.feature_extractor.6.5.bn3.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %481 : Float(10, 1024, 38, 38) = onnx::Add(%480, %472)
  %482 : Float(10, 1024, 38, 38) = onnx::Relu(%481) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %483 : Float(10, 256, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%482, %additional_blocks.0.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %484 : Float(10, 256, 38, 38) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%483, %additional_blocks.0.1.weight, %additional_blocks.0.1.bias, %additional_blocks.0.1.running_mean, %additional_blocks.0.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %485 : Float(10, 256, 38, 38) = onnx::Relu(%484) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %486 : Float(10, 512, 19, 19) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%485, %additional_blocks.0.3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %487 : Float(10, 512, 19, 19) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%486, %additional_blocks.0.4.weight, %additional_blocks.0.4.bias, %additional_blocks.0.4.running_mean, %additional_blocks.0.4.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %488 : Float(10, 512, 19, 19) = onnx::Relu(%487) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %489 : Float(10, 256, 19, 19) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%488, %additional_blocks.1.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %490 : Float(10, 256, 19, 19) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%489, %additional_blocks.1.1.weight, %additional_blocks.1.1.bias, %additional_blocks.1.1.running_mean, %additional_blocks.1.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %491 : Float(10, 256, 19, 19) = onnx::Relu(%490) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %492 : Float(10, 512, 10, 10) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%491, %additional_blocks.1.3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %493 : Float(10, 512, 10, 10) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%492, %additional_blocks.1.4.weight, %additional_blocks.1.4.bias, %additional_blocks.1.4.running_mean, %additional_blocks.1.4.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %494 : Float(10, 512, 10, 10) = onnx::Relu(%493) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %495 : Float(10, 128, 10, 10) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%494, %additional_blocks.2.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %496 : Float(10, 128, 10, 10) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%495, %additional_blocks.2.1.weight, %additional_blocks.2.1.bias, %additional_blocks.2.1.running_mean, %additional_blocks.2.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %497 : Float(10, 128, 10, 10) = onnx::Relu(%496) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %498 : Float(10, 256, 5, 5) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2]](%497, %additional_blocks.2.3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %499 : Float(10, 256, 5, 5) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%498, %additional_blocks.2.4.weight, %additional_blocks.2.4.bias, %additional_blocks.2.4.running_mean, %additional_blocks.2.4.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %500 : Float(10, 256, 5, 5) = onnx::Relu(%499) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %501 : Float(10, 128, 5, 5) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%500, %additional_blocks.3.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %502 : Float(10, 128, 5, 5) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%501, %additional_blocks.3.1.weight, %additional_blocks.3.1.bias, %additional_blocks.3.1.running_mean, %additional_blocks.3.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %503 : Float(10, 128, 5, 5) = onnx::Relu(%502) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %504 : Float(10, 256, 3, 3) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[0, 0, 0, 0], strides=[1, 1]](%503, %additional_blocks.3.3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %505 : Float(10, 256, 3, 3) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%504, %additional_blocks.3.4.weight, %additional_blocks.3.4.bias, %additional_blocks.3.4.running_mean, %additional_blocks.3.4.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %506 : Float(10, 256, 3, 3) = onnx::Relu(%505) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %507 : Float(10, 128, 3, 3) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%506, %additional_blocks.4.0.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %508 : Float(10, 128, 3, 3) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%507, %additional_blocks.4.1.weight, %additional_blocks.4.1.bias, %additional_blocks.4.1.running_mean, %additional_blocks.4.1.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %509 : Float(10, 128, 3, 3) = onnx::Relu(%508) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %510 : Float(10, 256, 1, 1) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[0, 0, 0, 0], strides=[1, 1]](%509, %additional_blocks.4.3.weight) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %511 : Float(10, 256, 1, 1) = onnx::BatchNormalization[epsilon=1.0000000000000001e-05, momentum=0.90000000000000002](%510, %additional_blocks.4.4.weight, %additional_blocks.4.4.bias, %additional_blocks.4.4.running_mean, %additional_blocks.4.4.running_var) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1923:0
  %512 : Float(10, 256, 1, 1) = onnx::Relu(%511) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/functional.py:1061:0
  %513 : Float(10, 16, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%482, %loc.0.weight, %loc.0.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %514 : Tensor = onnx::Shape(%482)
  %515 : Tensor = onnx::Constant[value={0}]()
  %516 : Long() = onnx::Gather[axis=0](%514, %515) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %519 : Tensor = onnx::Unsqueeze[axes=[0]](%516)
  %522 : Tensor = onnx::Concat[axis=0](%519, %647, %648)
  %523 : Float(10, 4, 5776) = onnx::Reshape(%513, %522) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %524 : Float(10, 324, 38, 38) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%482, %conf.0.weight, %conf.0.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %525 : Tensor = onnx::Shape(%482)
  %526 : Tensor = onnx::Constant[value={0}]()
  %527 : Long() = onnx::Gather[axis=0](%525, %526) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %530 : Tensor = onnx::Unsqueeze[axes=[0]](%527)
  %533 : Tensor = onnx::Concat[axis=0](%530, %649, %650)
  %534 : Float(10, 81, 5776) = onnx::Reshape(%524, %533) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %535 : Float(10, 24, 19, 19) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%488, %loc.1.weight, %loc.1.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %536 : Tensor = onnx::Shape(%488)
  %537 : Tensor = onnx::Constant[value={0}]()
  %538 : Long() = onnx::Gather[axis=0](%536, %537) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %541 : Tensor = onnx::Unsqueeze[axes=[0]](%538)
  %544 : Tensor = onnx::Concat[axis=0](%541, %651, %652)
  %545 : Float(10, 4, 2166) = onnx::Reshape(%535, %544) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %546 : Float(10, 486, 19, 19) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%488, %conf.1.weight, %conf.1.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %547 : Tensor = onnx::Shape(%488)
  %548 : Tensor = onnx::Constant[value={0}]()
  %549 : Long() = onnx::Gather[axis=0](%547, %548) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %552 : Tensor = onnx::Unsqueeze[axes=[0]](%549)
  %555 : Tensor = onnx::Concat[axis=0](%552, %653, %654)
  %556 : Float(10, 81, 2166) = onnx::Reshape(%546, %555) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %557 : Float(10, 24, 10, 10) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%494, %loc.2.weight, %loc.2.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %558 : Tensor = onnx::Shape(%494)
  %559 : Tensor = onnx::Constant[value={0}]()
  %560 : Long() = onnx::Gather[axis=0](%558, %559) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %563 : Tensor = onnx::Unsqueeze[axes=[0]](%560)
  %566 : Tensor = onnx::Concat[axis=0](%563, %655, %656)
  %567 : Float(10, 4, 600) = onnx::Reshape(%557, %566) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %568 : Float(10, 486, 10, 10) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%494, %conf.2.weight, %conf.2.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %569 : Tensor = onnx::Shape(%494)
  %570 : Tensor = onnx::Constant[value={0}]()
  %571 : Long() = onnx::Gather[axis=0](%569, %570) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %574 : Tensor = onnx::Unsqueeze[axes=[0]](%571)
  %577 : Tensor = onnx::Concat[axis=0](%574, %657, %658)
  %578 : Float(10, 81, 600) = onnx::Reshape(%568, %577) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %579 : Float(10, 24, 5, 5) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%500, %loc.3.weight, %loc.3.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %580 : Tensor = onnx::Shape(%500)
  %581 : Tensor = onnx::Constant[value={0}]()
  %582 : Long() = onnx::Gather[axis=0](%580, %581) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %585 : Tensor = onnx::Unsqueeze[axes=[0]](%582)
  %588 : Tensor = onnx::Concat[axis=0](%585, %659, %660)
  %589 : Float(10, 4, 150) = onnx::Reshape(%579, %588) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %590 : Float(10, 486, 5, 5) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%500, %conf.3.weight, %conf.3.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %591 : Tensor = onnx::Shape(%500)
  %592 : Tensor = onnx::Constant[value={0}]()
  %593 : Long() = onnx::Gather[axis=0](%591, %592) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %596 : Tensor = onnx::Unsqueeze[axes=[0]](%593)
  %599 : Tensor = onnx::Concat[axis=0](%596, %661, %662)
  %600 : Float(10, 81, 150) = onnx::Reshape(%590, %599) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %601 : Float(10, 16, 3, 3) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%506, %loc.4.weight, %loc.4.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %602 : Tensor = onnx::Shape(%506)
  %603 : Tensor = onnx::Constant[value={0}]()
  %604 : Long() = onnx::Gather[axis=0](%602, %603) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %607 : Tensor = onnx::Unsqueeze[axes=[0]](%604)
  %610 : Tensor = onnx::Concat[axis=0](%607, %663, %664)
  %611 : Float(10, 4, 36) = onnx::Reshape(%601, %610) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %612 : Float(10, 324, 3, 3) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%506, %conf.4.weight, %conf.4.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %613 : Tensor = onnx::Shape(%506)
  %614 : Tensor = onnx::Constant[value={0}]()
  %615 : Long() = onnx::Gather[axis=0](%613, %614) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %618 : Tensor = onnx::Unsqueeze[axes=[0]](%615)
  %621 : Tensor = onnx::Concat[axis=0](%618, %665, %666)
  %622 : Float(10, 81, 36) = onnx::Reshape(%612, %621) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %623 : Float(10, 16, 1, 1) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%512, %loc.5.weight, %loc.5.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %624 : Tensor = onnx::Shape(%512)
  %625 : Tensor = onnx::Constant[value={0}]()
  %626 : Long() = onnx::Gather[axis=0](%624, %625) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %629 : Tensor = onnx::Unsqueeze[axes=[0]](%626)
  %632 : Tensor = onnx::Concat[axis=0](%629, %667, %668)
  %633 : Float(10, 4, 4) = onnx::Reshape(%623, %632) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %634 : Float(10, 324, 1, 1) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1]](%512, %conf.5.weight, %conf.5.bias) # /home/hajime/anaconda3/envs/pytorch/lib/python3.7/site-packages/torch/nn/modules/conv.py:346:0
  %635 : Tensor = onnx::Shape(%512)
  %636 : Tensor = onnx::Constant[value={0}]()
  %637 : Long() = onnx::Gather[axis=0](%635, %636) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %640 : Tensor = onnx::Unsqueeze[axes=[0]](%637)
  %643 : Tensor = onnx::Concat[axis=0](%640, %669, %670)
  %644 : Float(10, 81, 4) = onnx::Reshape(%634, %643) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:97:0
  %output1 : Float(10, 4, 8732) = onnx::Concat[axis=2](%523, %545, %567, %589, %611, %633) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:100:0
  %646 : Float(10, 81, 8732) = onnx::Concat[axis=2](%534, %556, %578, %600, %622, %644) # /home/hajime/.cache/torch/hub/NVIDIA_DeepLearningExamples_torchhub/PyTorch/Detection/SSD/src/model.py:100:0
  return (%output1, %646)

コメント

  1. ONNXモデルをRaspi＋Intel NCS2向けのItermediate Representationに変換 | IoT & dApp BLOCKCHAIN より:
    2020-05-10 11:59 am

    […] PyTorchモデルをONNXに変換の続きです。 […]

    返信
タイトルとURLをコピーしました