[Enhance] refine docs and code (open-mmlab#2330)

mmpose/models/losses/classification_loss.py

@@ -139,11 +139,10 @@ def __init__(self, beta=1.0, label_softmax=False, use_target_weight=True):
 
     def criterion(self, dec_outs, labels):
         """Criterion function."""
-
-        scores = self.log_softmax(dec_outs * self.beta)
+        log_pt = self.log_softmax(dec_outs * self.beta)
         if self.label_softmax:
             labels = F.softmax(labels * self.beta, dim=1)
-        loss = torch.mean(self.kl_loss(scores, labels), dim=1)
+        loss = torch.mean(self.kl_loss(log_pt, labels), dim=1)
         return loss
 
     def forward(self, pred_simcc, gt_simcc, target_weight):
@@ -156,26 +155,19 @@ def forward(self, pred_simcc, gt_simcc, target_weight):
             target_weight (torch.Tensor[N, K] or torch.Tensor[N]):
                 Weights across different labels.
         """
-        output_x, output_y = pred_simcc
-        target_x, target_y = gt_simcc
-        num_joints = output_x.size(1)
+        num_joints = pred_simcc[0].size(1)
         loss = 0
 
-        for idx in range(num_joints):
-            coord_x_pred = output_x[:, idx].squeeze()
-            coord_y_pred = output_y[:, idx].squeeze()
-            coord_x_gt = target_x[:, idx].squeeze()
-            coord_y_gt = target_y[:, idx].squeeze()
-
-            if self.use_target_weight:
-                weight = target_weight[:, idx].squeeze()
-            else:
-                weight = 1.
-
-            loss += (
-                self.criterion(coord_x_pred, coord_x_gt).mul(weight).sum())
-            loss += (
-                self.criterion(coord_y_pred, coord_y_gt).mul(weight).sum())
+        if self.use_target_weight:
+            weight = target_weight.reshape(-1)
+        else:
+            weight = 1.
+
+        for pred, target in zip(pred_simcc, gt_simcc):
+            pred = pred.reshape(-1, pred.size(-1))
+            target = target.reshape(-1, target.size(-1))
+
+            loss += self.criterion(pred, target).mul(weight).sum()
 
         return loss / num_joints
 

projects/rtmpose/examples/README.md

@@ -1,9 +1,11 @@
-# Examples
-
 ## List of examples
 
-### 1. RTMPose-Deploy
+### 1. RTMPose-Deploy (without MMDeploy)
 
-RTMPose-Deploy is a C ++ code example that does not use MMDEPLOY for RTMPose localized deployment.
+RTMPose-Deploy is a C++ code example that does not use MMDEPLOY for RTMPose localized deployment.
 
 - [Original Repository](https://github.com/HW140701/RTMPose-Deploy)
+
+#### 2. RTMPose inference with ONNXRuntime (Python)
+
+This example shows how to run RTMPose inference with ONNXRuntime in Python.

projects/rtmpose/examples/RTMPose-Deploy/README.md

@@ -6,4 +6,4 @@ RTMPose-Deploy is a C ++ code example that does **NOT** use MMDEPLOY for RTMPose
 
 At present, RTMPose-Deploy has completed the deployment of RTMDetnano and RTMPose on the Windows system. This example only contains the source code. If you want a complete project example, please refer to:[https://github.com/HW140701/RTMPose-Deploy](https://github.com/HW140701/RTMPose-Deploy) . This project provides a complete VS2019 project and release package.
 
-Subsequent will consider adding the use of C ++ Tensorrt SDK on the Windows system to deploy RTMDetnano and RTMPose.
+Subsequent will consider adding the use of C ++ Tensorrt SDK on the Windows system to deploy RTMDet-nano and RTMPose.