update rknn-toolkit2 version to 1.5 (#2278)

* update rknn-toolkit2 version to 1.5

* fix lint

* cmake ..

* fix int type

* fix

Author: AllentDan

csrc/mmdeploy/net/rknn/rknn_net.cpp

@@ -170,7 +170,7 @@ Result<void> RKNNNet::Init(const Value& args) {
       MMDEPLOY_ERROR("rknn query 'RKNN_QUERY_INPUT_ATTR' fail! ret: {}", ret);
       return Status(eFail);
     }
-    if (attr.type != RKNN_TENSOR_UINT8) {
+    if (!(attr.type == RKNN_TENSOR_UINT8 || attr.type == RKNN_TENSOR_INT8)) {
       MMDEPLOY_ERROR("MMDeploy SDK only supports RKNN-INT8 model");
       return Status(eInvalidArgument);
     }

docs/en/01-how-to-build/rockchip.md

@@ -23,7 +23,7 @@ It is recommended to create a virtual environment for the project.
 2. Install RKNN python package following [rknn-toolkit2 doc](https://github.com/rockchip-linux/rknn-toolkit2/tree/master/doc) or [rknn-toolkit doc](https://github.com/rockchip-linux/rknn-toolkit/tree/master/doc). When installing rknn python package, it is better to append `--no-deps` after the commands to avoid dependency conflicts. RKNN-Toolkit2 package for example:
 
    ```
-   pip install packages/rknn_toolkit2-1.2.0_f7bb160f-cp36-cp36m-linux_x86_64.whl --no-deps
+   pip install packages/rknn_toolkit2-1.4.0_22dcfef4-cp36-cp36m-linux_x86_64.whl --no-deps
    ```
 
 3. Install ONNX==1.8.0 before reinstall MMDeploy from source following the [instructions](../01-how-to-build/build_from_source.md). Note that there are conflicts between the pip dependencies of MMDeploy and RKNN. Here is the suggested packages versions for python 3.6:
@@ -50,7 +50,7 @@ Example:
 
 ```bash
 python tools/deploy.py \
-    configs/mmpretrain/classification_rknn_static.py \
+    configs/mmpretrain/classification_rknn-fp16_static-224x224.py \
     /mmpretrain_dir/configs/resnet/resnet50_8xb32_in1k.py \
     https://download.openmmlab.com/mmclassification/v0/resnet/resnet50_batch256_imagenet_20200708-cfb998bf.pth \
     /mmpretrain_dir/demo/demo.JPEG \

docs/zh_cn/01-how-to-build/rockchip.md

@@ -63,7 +63,7 @@ MMDeploy 支持把模型部署到瑞芯微设备上。已支持的芯片：RV112
    2.2 参考表中的安装指南，安装 RKNN python 安装包。建议在安装时，使用选项 `--no-deps`，以避免依赖包的冲突。以 rknn-toolkit2 为例:
 
    ```
-   pip install packages/rknn_toolkit2-1.2.0_f7bb160f-cp36-cp36m-linux_x86_64.whl --no-deps
+   pip install packages/rknn_toolkit2-1.4.0_22dcfef4-cp36-cp36m-linux_x86_64.whl --no-deps
    ```
 
    2.3 先安装onnx==1.8.0,跟着 [instructions](../01-how-to-build/build_from_source.md)，源码安装 MMDeploy。 需要注意的是， MMDeploy 和 RKNN 依赖的安装包间有冲突的内容. 这里提供建议在 python 3.6 环境中使用的安装包版本:
@@ -88,7 +88,7 @@ git clone https://github.com/open-mmlab/mmpretrain
 # 执行转换命令
 cd /the/path/of/mmdeploy
 python tools/deploy.py \
-    configs/mmpretrain/classification_rknn_static.py \
+    configs/mmpretrain/classification_rknn-fp16_static-224x224.py \
     /the/path/of/mmpretrain/configs/resnet/resnet50_8xb32_in1k.py \
     https://download.openmmlab.com/mmclassification/v0/resnet/resnet50_batch256_imagenet_20200708-cfb998bf.pth \
     /the/path/of/mmpretrain/demo/demo.JPEG \
@@ -130,8 +130,13 @@ git clone https://github.com/open-mmlab/mmdetection
 
 # 执行转换命令
 python tools/deploy.py \
-    configs/mmpretrain/detection_rknn_static.py \
-
+    configs/mmdet/detection/detection_rknn-int8_static-320x320.py \
+    /the/path/of/mmdet/configs/yolov3/yolov3_mobilenetv2_320_300e_coco.py \
+    https://download.openmmlab.com/mmdetection/v2.0/yolo/yolov3_mobilenetv2_320_300e_coco/yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth \
+    /the/path/of/mmdet/demo/demo.jpg \
+    --work-dir mmdeploy_models/mmdet/yolov3 \
+    --device cpu \
+    --dump-info
 ```
 
 - RTMDet
@@ -299,7 +304,7 @@ export OpenCV_AARCH64_INSTALL_DIR=$(pwd)/install
 cd /path/to/mmdeploy
 mkdir -p build && cd build
 export LD_LIBRARY_PATH=$RKNN_TOOL_CHAIN/lib64:$LD_LIBRARY_PATH
-cmake \
+cmake ..\
     -DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/rknpu2-linux-gnu.cmake \
     -DMMDEPLOY_BUILD_SDK=ON \
     -DMMDEPLOY_TARGET_BACKENDS="rknn" \

mmdeploy/backend/rknn/onnx2rknn.py

@@ -2,6 +2,7 @@
 from typing import Optional, Union
 
 import mmengine
+from packaging import version
 from rknn.api import RKNN
 
 from mmdeploy.utils import (get_common_config, get_normalization,
@@ -50,10 +51,16 @@ def onnx2rknn(onnx_model: str,
     onnx_params = get_onnx_config(deploy_cfg)
     quantization_cfg = get_quantization_config(deploy_cfg)
 
+    package_info = rknn_package_info()
     input_names = onnx_params.get('input_names', None)
     output_names = onnx_params.get('output_names', None)
     input_size_list = get_backend_config(deploy_cfg).get(
         'input_size_list', None)
+    # rknn-toolkit 1.5+ can not pass input output info, which is weird
+    if package_info['name'] == 'rknn-toolkit2' and version.parse(
+            package_info['version']) > version.parse('1.4'):
+        input_names, output_names, input_size_list = [None] * 3
+
     # update norm value
     if get_rknn_quantization(deploy_cfg) is True and model_cfg is not None:
         transform = get_normalization(model_cfg)
@@ -85,7 +92,7 @@ def onnx2rknn(onnx_model: str,
         if dataset_file is None:
             quantization_cfg.update(dict(do_quantization=False))
             logger.warning('no dataset passed in, quantization is skipped')
-    if rknn_package_info()['name'] == 'rknn-toolkit2':
+    if package_info['name'] == 'rknn-toolkit2':
         quantization_cfg.pop('pre_compile', None)
     ret = rknn.build(**quantization_cfg)
     if ret != 0: