Merge remote-tracking branch 'upstream/dev-1.x' into dominic/rtmpose_…

…tensorrt_project

mmpose/models/heads/coord_cls_heads/rtmcc_head.py

@@ -134,8 +134,8 @@ def __init__(
     def forward(self, feats: Tuple[Tensor]) -> Tuple[Tensor, Tensor]:
         """Forward the network.
 
-        The input is multi scale feature maps and the
-        output is the heatmap.
+        The input is the featuremap extracted by backbone and the
+        output is the simcc representation.
 
         Args:
             feats (Tuple[Tensor]): Multi scale feature maps.

mmpose/models/heads/coord_cls_heads/simcc_head.py

@@ -198,8 +198,10 @@ def _make_deconv_head(
         return deconv_head
 
     def forward(self, feats: Tuple[Tensor]) -> Tuple[Tensor, Tensor]:
-        """Forward the network. The input is multi scale feature maps and the
-        output is the heatmap.
+        """Forward the network.
+
+        The input is the featuremap extracted by backbone and the
+        output is the simcc representation.
 
         Args:
             feats (Tuple[Tensor]): Multi scale feature maps.

projects/rtmpose/README.md

@@ -320,7 +320,7 @@ We also provide the ImageNet classification pre-trained weights of the CSPNeXt b
 We provide two appoaches to try RTMPose:
 
 - MMPose demo scripts
-- Pre-compiled MMDeploy SDK (Recommended)
+- Pre-compiled MMDeploy SDK (Recommend, 6-10 times faster)
 
 ### MMPose demo scripts
 
@@ -364,6 +364,8 @@ MMDeploy provides a precompiled SDK for Pipeline reasoning on RTMPose projects,
 
 - All models must by exported by `tools/deploy.py` before PoseTracker can be used for inference.
 - For the tutorial of exporting the SDK version model, see [SDK Reasoning](#%EF%B8%8F-step3-inference-with-sdk), and for detailed parameter settings of inference, see [Pipeline Reasoning](#-step4-pipeline-inference).
+- Exported SDK models (ONNX, TRT, ncnn, etc.) can be downloaded from [OpenMMLab Deploee](https://platform.openmmlab.com/deploee).
+- You can also convert `.pth` models into SDK [online](https://platform.openmmlab.com/deploee/task-convert-list).
 
 #### Linux
 
@@ -597,6 +599,9 @@ Please refer to [Train and Test](https://mmpose.readthedocs.io/en/latest/user_gu
 
 Here is a basic example of deploy RTMPose with [MMDeploy](https://github.com/open-mmlab/mmdeploy/tree/main).
 
+- Exported SDK models (ONNX, TRT, ncnn, etc.) can be downloaded from [OpenMMLab Deploee](https://platform.openmmlab.com/deploee).
+- You can also convert `.pth` models into SDK [online](https://platform.openmmlab.com/deploee/task-convert-list).
+
 ### 🧩 Step1. Install MMDeploy
 
 Before starting the deployment, please make sure you install MMPose and MMDeploy correctly.

projects/rtmpose/README_CN.md

@@ -311,7 +311,7 @@ RTMPose 是一个长期优化迭代的项目，致力于业务场景下的高性
 我们提供了两种途径来让用户尝试 RTMPose 模型：
 
 - MMPose demo 脚本
-- MMDeploy SDK 预编译包 （推荐）
+- MMDeploy SDK 预编译包 （推荐，速度提升6-10倍）
 
 ### MMPose demo 脚本
 
@@ -357,6 +357,8 @@ MMDeploy 提供了预编译的 SDK，用于对 RTMPose 项目进行 Pipeline 推
 
 - 所有的模型必须经过 `tools/deploy.py` 导出后才能使用 PoseTracker 进行推理。
 - 导出 SDK 版模型的教程见 [SDK 推理](#%EF%B8%8F-sdk-推理)，推理的详细参数设置见 [Pipeline 推理](#-pipeline-推理)。
+- 你可以从 [硬件模型库](https://platform.openmmlab.com/deploee) 直接下载 SDK 版模型（ONNX、 TRT、ncnn 等）。
+- 同时我们也支持 [在线模型转换](https://platform.openmmlab.com/deploee/task-convert-list)。
 
 #### Linux\\
 
@@ -588,6 +590,9 @@ example\cpp\build\Release
 
 本教程将展示如何通过 [MMDeploy](https://github.com/open-mmlab/mmdeploy/tree/main) 部署 RTMPose 项目。
 
+- 你可以从 [硬件模型库](https://platform.openmmlab.com/deploee) 直接下载 SDK 版模型（ONNX、 TRT、ncnn 等）。
+- 同时我们也支持 [在线模型转换](https://platform.openmmlab.com/deploee/task-convert-list)。
+
 ### 🧩 安装
 
 在开始部署之前，首先你需要确保正确安装了 MMPose, MMDetection, MMDeploy，相关安装教程如下：