From 0773fff20fbfaf7db67dac6e69df10397e0bbb34 Mon Sep 17 00:00:00 2001
From: coincheung <867153576@qq.com>
Date: Sun, 5 Feb 2023 03:04:28 +0000
Subject: [PATCH] ncnn

---
 .gitignore           |  2 ++
 lib/models/resnet.py |  1 +
 ncnn/README.md       | 43 +++++++++++++++++++++++++------------------
 ncnn/segment.cpp     | 10 +++++++---
 4 files changed, 35 insertions(+), 21 deletions(-)

diff --git a/.gitignore b/.gitignore
index 7a1c8d1..25e57be 100644
--- a/.gitignore
+++ b/.gitignore
@@ -117,7 +117,9 @@ pretrained/*
 run.sh
 openvino/build/*
 openvino/output*
+ncnn/models/*
 *.onnx
 tis/cpp_client/build/*
+log*txt
 
 tvm/
diff --git a/lib/models/resnet.py b/lib/models/resnet.py
index 815f564..7b039ed 100644
--- a/lib/models/resnet.py
+++ b/lib/models/resnet.py
@@ -8,6 +8,7 @@
 
 resnet18_url = 'https://download.pytorch.org/models/resnet18-5c106cde.pth'
 
+
 from torch.nn import BatchNorm2d
 
 
diff --git a/ncnn/README.md b/ncnn/README.md
index 4861967..fb27b93 100644
--- a/ncnn/README.md
+++ b/ncnn/README.md
@@ -1,5 +1,5 @@
 
-### My platform
+### My platform  
 
 * raspberry pi 3b
 * 2022-04-04-raspios-bullseye-armhf-lite.img 
@@ -7,14 +7,8 @@
 
 
 
-### Install ncnn
+### Install ncnn  
 
-#### 1. dependencies  
-```
-$ python -m pip install onnx-simplifier
-```
-
-#### 2. build ncnn  
 Just follow the ncnn official tutoral of [build-for-linux](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-linux) to install ncnn. Following steps are all carried out on my raspberry pi:  
 
 **step 1:** install dependencies  
@@ -25,11 +19,11 @@ $ sudo apt install build-essential git cmake libprotobuf-dev protobuf-compiler l
 **step 2:** (optional) install vulkan  
 
 **step 3:** build   
-I am using commit `5725c028c0980efd`, and I have not tested over other commits.  
+I am using commit `6869c81ed3e7170dc0`, and I have not tested over other commits.  
 ```
 $ git clone https://github.com/Tencent/ncnn.git
 $ cd ncnn
-$ git reset --hard 5725c028c0980efd
+$ git reset --hard 6869c81ed3e7170dc0
 $ git submodule update --init
 $ mkdir -p build
 $ cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_VULKAN=OFF -DNCNN_BUILD_TOOLS=ON -DCMAKE_TOOLCHAIN_FILE=../toolchains/pi3.toolchain.cmake ..
@@ -37,9 +31,14 @@ $ make -j2
 $ make install 
 ```
 
-### Convert model, build and run the demo
+### Convert pytorch model to ncnn model   
 
-#### 1. convert pytorch model to ncnn model via onnx  
+#### 1. dependencies  
+```
+$ python -m pip install onnx-simplifier
+```
+
+#### 2. convert pytorch model to ncnn model via onnx  
 On your training platform:
 ```
 $ cd BiSeNet/
@@ -52,13 +51,21 @@ Then copy your `model_v2_sim.onnx` from training platform to raspberry device.
 On raspberry device:  
 ```
 $ /path/to/ncnn/build/tools/onnx/onnx2ncnn model_v2_sim.onnx model_v2_sim.param model_v2_sim.bin
-$ cd BiSeNet/ncnn/
-$ mkdir -p models
-$ mv model_v2_sim.param models/
-$ mv model_v2_sim.bin models/
 ```
 
-#### 2. compile demo code  
+You can optimize the ncnn model by fusing the layers and save the weights with fp16 datatype.   
+On raspberry device:
+```
+$ /path/to/ncnn/build/tools/ncnnoptimize model_v2_sim.param model_v2_sim.bin model_v2_sim_opt.param model_v2_sim_opt.bin 65536
+$ mv model_v2_sim_opt.param model_v2_sim.param
+$ mv model_v2_sim_opt.bin model_v2_sim.bin
+```
+
+You can also quantize the model for int8 inference, following this [tutorial](https://github.com/Tencent/ncnn/wiki/quantized-int8-inference). Make sure your device support int8 inference.  
+
+
+### build and run the demo
+#### 1. compile demo code  
 On raspberry device:  
 ```
 $ mkdir -p BiSeNet/ncnn/build
@@ -67,7 +74,7 @@ $ cmake .. -DNCNN_ROOT=/path/to/ncnn/build/install
 $ make
 ```
 
-#### 3. run demo  
+#### 2. run demo  
 ```
 ./segment
 ```
diff --git a/ncnn/segment.cpp b/ncnn/segment.cpp
index ee6e3ba..e570b09 100644
--- a/ncnn/segment.cpp
+++ b/ncnn/segment.cpp
@@ -51,12 +51,16 @@ void inference() {
     mod.opt.use_vulkan_compute = 1;
     mod.set_vulkan_device(1);
 #endif 
-    // ncnn enable fp16 by default, so we do not need these options
-    // int8 depends on the model itself, so we do not set here
+    //// switch off fp16
     // bool use_fp16 = false;
     // mod.opt.use_fp16_packed = use_fp16;
     // mod.opt.use_fp16_storage = use_fp16;
     // mod.opt.use_fp16_arithmetic = use_fp16;
+    //// switch on bf16
+    // mod.opt.use_packing_layout = true;
+    // mod.opt.use_ff16_storage = true;
+    //// reduce cpu usage
+    // net.opt.openmp_blocktime = 0; 
     mod.opt.use_winograd_convolution = true;
 
     // we should set opt before load model
@@ -78,7 +82,7 @@ void inference() {
 
     // set input, run, get output
     ncnn::Extractor ex = mod.create_extractor();
-    ex.set_light_mode(true); // not sure what this mean
+    ex.set_light_mode(true); 
     ex.set_num_threads(nthreads);
 #if NCNN_VULKAN
     ex.set_vulkan_compute(true);