Feature/torch2trt (#165)

docker/docker_testbench/.gitignore

@@ -0,0 +1,3 @@
+mount_point
+*.zip
+*.img

docker/torch2trt/Dockerfile

@@ -0,0 +1,17 @@
+FROM nvcr.io/nvidia/pytorch:21.09-py3
+RUN python3 -m pip install nvidia-pyindex packaging && \
+  python3 -m pip install --upgrade nvidia-tensorrt
+RUN apt-get update && \
+  apt-get install -y git libgl1-mesa-dev && \
+  apt-get -y clean && \
+  rm -rf /var/lib/apt/lists/*
+WORKDIR /
+RUN git clone https://github.com/NVIDIA-AI-IOT/torch2trt.git
+WORKDIR /torch2trt
+RUN python3 setup.py install --plugins
+WORKDIR /
+RUN git clone https://github.com/Megvii-BaseDetection/YOLOX.git
+WORKDIR /YOLOX
+RUN python3 -m pip install -r requirements.txt &&\
+  python3 setup.py develop
+RUN mkdir model

docker/torch2trt/convert.sh

@@ -0,0 +1,3 @@
+docker-compose build
+docker-compose up
+sudo chown $USER -R model

docker/torch2trt/docker-compose.yaml

@@ -0,0 +1,17 @@
+version: '2.3'
+services:
+  torch2trt:
+    volumes:
+      # mount model
+      - ./model:/YOLOX/model
+      - ./model/YOLOX_outputs:/YOLOX/YOLOX_outputs
+    build: .
+    container_name: torch2trt_container
+    runtime: nvidia
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=all
+    privileged: true
+    # network_mode: host
+    #command: /bin/bash -c "while :; do sleep 10; done"
+    command: python3 tools/trt.py -n yolox-nano -c model/yolox_nano.pth

docker/torch2trt/model/.gitignore

@@ -0,0 +1,2 @@
+*.pth
+*.trt

docs/docs/index.md

@@ -21,6 +21,9 @@ See [tutorials](tutorials/supported_platforms)
 #### If you want to see navigation demo in your local machine.
 See [tutorials](tutorials/navigation_demo)
 
+#### If you want to know how to use tools.
+See [tools](tools/tools.md)
+
 #### If you want to know hardware specification.
 See [hardware documenation](https://ouxt-polaris.github.io/robotx-hardware-documentation/)
 

docs/docs/tools/tools.md

@@ -0,0 +1,123 @@
+# Tools
+
+## Convert yolox pytorch model into tensorrt model.
+### Requirement
+- docker
+- docker-compose
+- nvidia-docker
+- nvidia-gpu (hardware)
+
+```
+cd docker/torch2trt
+sh convert.sh
+```
+
+Output should be like below.
+```
+docker-compose build
+
+Building torch2trt
+Step 1/12 : FROM nvcr.io/nvidia/pytorch:21.09-py3
+ ---> 74d53f84c686
+Step 2/12 : RUN python3 -m pip install nvidia-pyindex packaging &&   python3 -m pip install --upgrade nvidia-tensorrt
+ ---> Using cache
+ ---> 8d655c414dd2
+Step 3/12 : RUN apt-get update &&   apt-get install -y git libgl1-mesa-dev &&   apt-get -y clean &&   rm -rf /var/lib/apt/lists/*
+ ---> Using cache
+ ---> 76bc393f04d5
+Step 4/12 : WORKDIR /
+ ---> Using cache
+ ---> 47b28fa00606
+Step 5/12 : RUN git clone https://github.com/NVIDIA-AI-IOT/torch2trt.git
+ ---> Using cache
+ ---> 1a4063050362
+Step 6/12 : WORKDIR /torch2trt
+ ---> Using cache
+ ---> 6e679ba630a9
+Step 7/12 : RUN python3 setup.py install --plugins
+ ---> Using cache
+ ---> f0f7dc347d9d
+Step 8/12 : WORKDIR /
+ ---> Using cache
+ ---> 1d4ffaaae9a0
+Step 9/12 : RUN git clone https://github.com/Megvii-BaseDetection/YOLOX.git
+ ---> Using cache
+ ---> 6cc15f5a7b36
+Step 10/12 : WORKDIR /YOLOX
+ ---> Using cache
+ ---> 3429678d2547
+Step 11/12 : RUN python3 -m pip install -r requirements.txt &&  python3 setup.py develop
+ ---> Using cache
+ ---> 377ef892e986
+Step 12/12 : RUN mkdir model
+ ---> Using cache
+ ---> 55caa4a2787d
+Successfully built 55caa4a2787d
+Successfully tagged torch2trt_torch2trt:latest
+
+docker-compose up
+
+Recreating torch2trt_torch2trt_1 ... done
+Attaching to torch2trt_torch2trt_1
+torch2trt_1  | 
+torch2trt_1  | =============
+torch2trt_1  | == PyTorch ==
+torch2trt_1  | =============
+torch2trt_1  | 
+torch2trt_1  | NVIDIA Release 21.09 (build 26760254)
+torch2trt_1  | PyTorch Version 1.10.0a0+3fd9dcf
+torch2trt_1  | 
+torch2trt_1  | Container image Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+torch2trt_1  | 
+torch2trt_1  | Copyright (c) 2014-2021 Facebook Inc.
+torch2trt_1  | Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+torch2trt_1  | Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+torch2trt_1  | Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+torch2trt_1  | Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+torch2trt_1  | Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
+torch2trt_1  | Copyright (c) 2006      Idiap Research Institute (Samy Bengio)
+torch2trt_1  | Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
+torch2trt_1  | Copyright (c) 2015      Google Inc.
+torch2trt_1  | Copyright (c) 2015      Yangqing Jia
+torch2trt_1  | Copyright (c) 2013-2016 The Caffe contributors
+torch2trt_1  | All rights reserved.
+torch2trt_1  | 
+torch2trt_1  | NVIDIA Deep Learning Profiler (dlprof) Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+torch2trt_1  | 
+torch2trt_1  | Various files include modifications (c) NVIDIA CORPORATION.  All rights reserved.
+torch2trt_1  | 
+torch2trt_1  | This container image and its contents are governed by the NVIDIA Deep Learning Container License.
+torch2trt_1  | By pulling and using the container, you accept the terms and conditions of this license:
+torch2trt_1  | https://developer.nvidia.com/ngc/nvidia-deep-learning-container-license
+torch2trt_1  | 
+torch2trt_1  | NOTE: MOFED driver for multi-node communication was not detected.
+torch2trt_1  |       Multi-node communication performance may be reduced.
+torch2trt_1  | 
+torch2trt_1  | NOTE: The SHMEM allocation limit is set to the default of 64MB.  This may be
+torch2trt_1  |    insufficient for PyTorch.  NVIDIA recommends the use of the following flags:
+torch2trt_1  |    nvidia-docker run --ipc=host ...
+torch2trt_1  | 
+torch2trt_1  | 2022-02-26 09:33:11.827 | INFO     | __main__:main:57 - loaded checkpoint done.
+torch2trt_1  | [02/26/2022-09:33:13] [TRT] [I] [MemUsageChange] Init CUDA: CPU +188, GPU +0, now: CPU 1383, GPU 2123 (MiB)
+torch2trt_1  | [02/26/2022-09:33:13] [TRT] [I] [MemUsageSnapshot] Begin constructing builder kernel library: CPU 1403 MiB, GPU 2123 MiB
+torch2trt_1  | [02/26/2022-09:33:13] [TRT] [I] [MemUsageSnapshot] End constructing builder kernel library: CPU 1410 MiB, GPU 2123 MiB
+torch2trt_1  | [02/26/2022-09:33:14] [TRT] [W] Tensor DataType is determined at build time for tensors not marked as input or output.
+torch2trt_1  | [02/26/2022-09:33:14] [TRT] [W] FP16 support requested on hardware without native FP16 support, performance will be negatively affected.
+torch2trt_1  | [02/26/2022-09:33:15] [TRT] [I] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 2107, GPU 2409 (MiB)
+torch2trt_1  | [02/26/2022-09:33:15] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 2107, GPU 2417 (MiB)
+torch2trt_1  | [02/26/2022-09:33:15] [TRT] [I] Local timing cache in use. Profiling results in this builder pass will not be stored.
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] Detected 1 inputs and 1 output network tensors.
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] Total Host Persistent Memory: 208144
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] Total Device Persistent Memory: 640512
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] Total Scratch Memory: 512
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] [MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 4 MiB, GPU 294 MiB
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] [BlockAssignment] Algorithm ShiftNTopDown took 43.9912ms to assign 7 blocks to 214 nodes requiring 7181828 bytes.
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] Total Activation Memory: 7181828
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +0, GPU +10, now: CPU 2129, GPU 2477 (MiB)
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in building engine: CPU +3, GPU +4, now: CPU 3, GPU 4 (MiB)
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 2128, GPU 2461 (MiB)
+torch2trt_1  | [02/26/2022-09:33:48] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +8, now: CPU 3, GPU 12 (MiB)
+torch2trt_1  | 2022-02-26 09:33:48.565 | INFO     | __main__:main:71 - Converted TensorRT model done.
+torch2trt_1  | 2022-02-26 09:33:48.581 | INFO     | __main__:main:79 - Converted TensorRT model engine file is saved for C++ inference.
+torch2trt_torch2trt_1 exited with code 0
+```

docs/mkdocs.yml

@@ -95,6 +95,7 @@ nav:
     - tutorials/build_instraction.md
     - tutorials/navigation_demo.md
   - Software Packages: packages.md
+  - Tools: tools/tools.md
   - Hardware Documentation: https://ouxt-polaris.github.io/robotx-hardware-documentation/
   - Automation Pipeline: automation_pipeline.md
   - LICENSE: license.md