Open32
ディープラーニングモデル量子化やフレームワーク間コンバージョンのアレコレ
PINTOファイルを投稿することはできないのな。
PINTOFlex OP を含むモデルを EdgeTPUモデル へ変換しようとすると Interpreter が対応していない、と怒られるんですよね。 回避する方法を一日中考えましたが、全く分かりません。 そもそも回避策はあるのかな？
PINTOnanodet_416x416_full_integer_quant.tflite
PINTO量子化を手計算してみた結果。 どう頑張ってもMin/Maxが合わないんですけど。。。
(x - zero_point) / (255 / (max - min))
(x - zero_point) * (1 / (255 / (max - min)))
(x - zero_point) * (1 / (255 / (max - min)))
min = -3.265998125076294
max = 2.2779781818389893
zero_point = 151
coefficient = 0.02182667888700962
coefficient of self-calculation = 0.021741084
(0 - 151) / (255 / (2.277978 - (-3.265998))) = −3.282903435
(151 - 151) / (255 / (2.277978 - (-3.265998))) = 0
(255 - 151) / (255 / (2.277978 - (-3.265998))) = 2.261072565
(0 - 151) * 0.02182667888700962 = −3.295828512
(151 - 151) * 0.02182667888700962 = 0
(255 - 151) * 0.02182667888700962 = 2.269974604
(0 - 151) * 0.021741084 = −3.282903684
(151 - 151) * 0.021741084 = 0
(255 - 151) * 0.021741084 = 2.261072736
git clone https://github.com/openvinotoolkit/training_extensions.git
cd training_extensions/pytorch_toolkit/instance_segmentation/model_templates/coco-instance-segmentation/instance-segmentation-0904
wget https://download.01.org/opencv/openvino_training_extensions/models/instance_segmentation/v2/instance-segmentation-0904-0912.pth
pip install torch
python ../../../../ote/tools/export.py \
--load-weights instance-segmentation-0904-0912.pth \
--save-model-to export \
--onnx \
--openvino
cd training_extensions/pytorch_toolkit
export PYTHONPATH=$PYTHONPATH:${PWD}:${PWD}/ote
wget https://download.01.org/opencv/openvino_training_extensions/models/instance_segmentation/v2/instance-segmentation-0904-0912.pth
cp instance_segmentation/model_templates/coco-instance-segmentation/instance-segmentation-0904/model.py .
cp instance_segmentation/model_templates/coco-instance-segmentation/instance-segmentation-0904/modules.yaml .
cp instance_segmentation/model_templates/coco-instance-segmentation/instance-segmentation-0904/template.yaml .
python3 ote/tools/export.py \
--load-weights instance_segmentation/model_templates/coco-instance-segmentation/instance-segmentation-0904/instance-segmentation-0904-0912.pth \
--save-model-to instance_segmentation/model_templates/coco-instance-segmentation/instance-segmentation-0904/export \
--onnx
$ dpkg -l | grep TensorRT
ii graphsurgeon-tf 7.2.2-1+cuda11.0 amd64 GraphSurgeon for TensorRT package
ii libnvinfer-bin 7.2.2-1+cuda11.0 amd64 TensorRT binaries
ii libnvinfer-dev 7.2.2-1+cuda11.0 amd64 TensorRT development libraries and headers
ii libnvinfer-doc 7.2.2-1+cuda11.0 all TensorRT documentation
ii libnvinfer-plugin-dev 7.2.2-1+cuda11.0 amd64 TensorRT plugin libraries
ii libnvinfer-plugin7 7.2.2-1+cuda11.0 amd64 TensorRT plugin libraries
ii libnvinfer-samples 7.2.2-1+cuda11.0 all TensorRT samples
ii libnvinfer7 7.2.2-1+cuda11.0 amd64 TensorRT runtime libraries
ii libnvonnxparsers-dev 7.2.2-1+cuda11.0 amd64 TensorRT ONNX libraries
ii libnvonnxparsers7 7.2.2-1+cuda11.0 amd64 TensorRT ONNX libraries
ii libnvparsers-dev 7.2.2-1+cuda11.0 amd64 TensorRT parsers libraries
ii libnvparsers7 7.2.2-1+cuda11.0 amd64 TensorRT parsers libraries
ii python3-libnvinfer 7.2.2-1+cuda11.0 amd64 Python 3 bindings for TensorRT
ii python3-libnvinfer-dev 7.2.2-1+cuda11.0 amd64 Python 3 development package for TensorRT
ii tensorrt 7.2.2.3-1+cuda11.0 amd64 Meta package of TensorRT
ii uff-converter-tf 7.2.2-1+cuda11.0 amd64 UFF converter for TensorRT package
$ /usr/local/cuda/bin/nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2020 NVIDIA Corporation
Built on Wed_Jul_22_19:09:09_PDT_2020
Cuda compilation tools, release 11.0, V11.0.221
Build cuda_11.0_bu.TC445_37.28845127_0
$ cat /usr/include/cudnn_version.h | grep CUDNN_MAJOR -A 2
#define CUDNN_MAJOR 8
#define CUDNN_MINOR 0
#define CUDNN_PATCHLEVEL 5
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)
#endif /* CUDNN_VERSION_H */
$ docker build -t pinto0309/tflite2tensorflow:latest .
$ docker run --gpus all -it --rm -v `pwd`:/workspace/resources -v ${HOME}/TFDS:/workspace/resources/TFDS pinto0309/tflite2tensorflow:latest
TensorRT DEB ダウンロードパス
PINTOdocker build -t pinto0309/tflite2tensorflow:latest .
docker run --gpus all -it --rm \
-v `pwd`:/workspace/resources \
-e LOCAL_UID=$(id -u $USER) \
-e LOCAL_GID=$(id -g $USER) \
pinto0309/tflite2tensorflow:latest bash
docker run --gpus all -it --rm \
-v `pwd`:/workspace/resources \
-v ${HOME}/TFDS:/workspace/resources/TFDS \
-e LOCAL_UID=$(id -u $USER) \
-e LOCAL_GID=$(id -g $USER) \
pinto0309/tflite2tensorflow:latest bash
PyTorch で DataParallel が使用されているモデルのONNXエクスポート用のトリック
PINTO双方向性 LSTM - UnidirectionalSequenceLSTM の実装
PINTO双方向性 LSTM - UnidirectionalSequenceLSTM のKeras実装サンプル
PINTOadaptive_max_pool2d -> max_pool2d
PINTOimport torch
import torch.nn.functional as F
import torch.nn as nn
x = torch.randn(1, 3, 512, 512)
F.max_pool2d(x, kernel_size=x.size()[2:])
tensor([[[[4.5801]],
[[4.4516]],
[[4.4280]]]])
F.adaptive_max_pool2d(x, (1, 1))
tensor([[[[4.5801]],
[[4.4516]],
[[4.4280]]]])
nn.AdaptiveMaxPool2d(1)(x)
tensor([[[[4.9739]],
[[4.5847]],
[[4.5794]]]])
F.max_pool2d(x, kernel_size=(170, 170))
tensor([[[[3.8038, 4.9739, 4.0777],
[4.5678, 4.1971, 4.0989],
[3.9696, 3.9876, 3.8003]],
[[3.9727, 4.0108, 4.5847],
[4.0080, 4.2751, 3.9599],
[4.0119, 3.9289, 4.2282]],
[[4.1449, 4.1064, 3.7096],
[4.0888, 3.6905, 4.5794],
[4.0994, 4.1094, 4.4437]]]])
F.adaptive_max_pool2d(x, (3, 3))
tensor([[[[3.8038, 4.9739, 4.0777],
[4.5678, 4.1971, 4.0989],
[3.9696, 3.9876, 3.8003]],
[[3.9727, 4.0108, 4.5847],
[4.0080, 4.2751, 3.9599],
[4.0119, 3.9289, 4.2282]],
[[4.1449, 4.1064, 3.7096],
[4.0888, 3.6905, 4.5794],
[4.0994, 4.1094, 4.4437]]]])
nn.AdaptiveMaxPool2d(3)(x)
tensor([[[[3.8038, 4.9739, 4.0777],
[4.5678, 4.1971, 4.0989],
[3.9696, 3.9876, 3.8003]],
[[3.9727, 4.0108, 4.5847],
[4.0080, 4.2751, 3.9599],
[4.0119, 3.9289, 4.2282]],
[[4.1449, 4.1064, 3.7096],
[4.0888, 3.6905, 4.5794],
[4.0994, 4.1094, 4.4437]]]])
nn.MaxPool2d(kernel_size=(170, 170))(x)
tensor([[[[3.8038, 4.9739, 4.0777],
[4.5678, 4.1971, 4.0989],
[3.9696, 3.9876, 3.8003]],
[[3.9727, 4.0108, 4.5847],
[4.0080, 4.2751, 3.9599],
[4.0119, 3.9289, 4.2282]],
[[4.1449, 4.1064, 3.7096],
[4.0888, 3.6905, 4.5794],
[4.0994, 4.1094, 4.4437]]]])
>>> F.max_pool2d(x, kernel_size=(170, 170)).shape
torch.Size([1, 3, 3, 3])
>>> F.adaptive_max_pool2d(x, (3, 3)).shape
torch.Size([1, 3, 3, 3])
>>> nn.AdaptiveMaxPool2d(3)(x).shape
torch.Size([1, 3, 3, 3])
nn.MaxPool2d(kernel_size=(170, 170))(x).shape
torch.Size([1, 3, 3, 3])
>>> x = torch.randn(1, 3, 1024, 1024)
>>> nn.MaxPool2d(kernel_size=(341, 341))(x).shape
torch.Size([1, 3, 3, 3])
>>> nn.MaxPool2d(kernel_size=(341, 341))(x)
tensor([[[[4.3699, 4.5813, 4.2206],
[4.0879, 4.7416, 4.1141],
[4.4719, 3.8822, 4.4646]],
[[4.1552, 4.3862, 4.0084],
[4.4847, 4.2279, 4.1345],
[4.1907, 4.5296, 4.6171]],
[[4.6567, 4.3410, 4.2800],
[4.7007, 4.0725, 4.6199],
[4.1732, 4.5214, 4.2547]]]])
>>> nn.AdaptiveMaxPool2d(3)(x)
tensor([[[[4.3699, 4.5813, 4.2206],
[4.2494, 4.7416, 4.1141],
[4.4719, 3.8822, 4.4646]],
[[4.1552, 4.3862, 4.0084],
[4.4847, 4.2279, 4.1345],
[4.1907, 4.5296, 4.6171]],
[[4.6567, 4.3410, 4.2800],
[4.7007, 4.0725, 4.6199],
[4.1732, 4.5214, 4.2547]]]])
if self.adaptive_diated:
self.adaptive_softmax = nn.Softmax(dim=3)
self.adaptive_layers = nn.Sequential(
nn.AdaptiveMaxPool2d(3),
nn.Conv2d(512, 512 * 3, 3, padding=0),
)
self.adaptive_bn = nn.BatchNorm2d(512)
self.adaptive_relu = nn.ReLU(inplace=True)
self.adaptive_layers1 = nn.Sequential(
nn.AdaptiveMaxPool2d(3),
nn.Conv2d(1024, 1024 * 3, 3, padding=0),
)
self.adaptive_bn1 = nn.BatchNorm2d(1024)
self.adaptive_relu1 = nn.ReLU(inplace=True)
ONNX to JSON
PINTOONNX to JSON, JSON to ONNX
import onnx
import json
from google.protobuf.json_format import MessageToJson
from google.protobuf.json_format import Parse
import os
# Convert onnx model to JSON
model_path = "ssdlite_mobiledet_edgetpu_320x320_coco.onnx"
filename, ext = os.path.splitext(model_path)
onnx_model = onnx.load(model_path)
s = MessageToJson(onnx_model)
onnx_json = json.loads(s)
with open(f'{filename}_onnx_to_json.json', 'w') as f:
json.dump(onnx_json, f, indent=2)
# Convert JSON to String
onnx_str = json.dumps(onnx_json)
# Convert String to onnx model
convert_model = Parse(onnx_str, onnx.ModelProto())
# Check
print(convert_model == onnx_model)