add converter

Author: BigJohnn

.vscode/settings.json

@@ -0,0 +1,3 @@
+{
+    "marscode.chatLanguage": "cn"
+}

examples/model_100.onnx

@@ -22,6 +22,12 @@ cd rsl_rl && git checkout v1.0.2 && pip install -e .
 
 `python zeroth_eval.py`
 
+模型转换。
+
+`cd genesis`
+
+`python3 utils/convert_to_onnx.py --cfg logs/zeroth-walking/cfgs.pkl --model ../../examples/model_100.pt --output ../../examples/model_100.onnx`
+
 注意⚠️
 Mac M系列芯片，使用micromamba替换conda使用，例如
 `micromamba activate genesis`

examples/model_100.pt

@@ -32,6 +32,22 @@ To monitor the training process, start TensorBoard:
 tensorboard --logdir logs
 ```
 
+## Model Conversion to ONNX
+
+To convert trained models to ONNX format:
+
+```bash
+python utils/convert_to_onnx.py \
+    --cfg logs/zeroth-walking/cfgs.pkl \
+    --model ../../examples/model_100.pt \
+    --output ../../examples/model_100.onnx
+```
+
+Arguments:
+- `--cfg`: Path to config file (.pkl)
+- `--model`: Path to model file (.pt) 
+- `--output`: Output ONNX file path
+
 ## Evaluation
 
 To view training results:

sim/genesis/README.md

@@ -0,0 +1,68 @@
+import torch
+import torch.nn as nn
+import pickle
+import argparse
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Convert model to ONNX format')
+    parser.add_argument('--cfg', type=str, required=True,
+                        help='Path to config file (.pkl)')
+    parser.add_argument('--model', type=str, required=True,
+                        help='Path to model file (.pt)')
+    parser.add_argument('--output', type=str, required=True,
+                        help='Output ONNX file path')
+    return parser.parse_args()
+
+args = parse_args()
+
+# 加载配置
+with open(args.cfg, 'rb') as f:
+    cfgs = pickle.load(f)
+
+# 加载模型
+model_dict = torch.load(args.model, weights_only=True)
+
+# 创建继承自ActorCritic的模型类
+from rsl_rl.modules import ActorCritic
+
+class ExportModel(ActorCritic):
+    def forward(self, obs):
+        # 使用actor网络生成动作
+        actions = self.actor(obs)
+        # 使用critic网络评估状态值
+        values = self.critic(obs)
+        return actions, values
+
+# 根据配置创建模型实例        
+model = ExportModel(
+    num_actor_obs=cfgs[1]['num_obs'],
+    num_critic_obs=cfgs[1]['num_obs'],
+    num_actions=cfgs[0]['num_actions'],
+    actor_hidden_dims=cfgs[4]['policy']['actor_hidden_dims'],
+    critic_hidden_dims=cfgs[4]['policy']['critic_hidden_dims'],
+    activation='elu',
+    init_noise_std=cfgs[4]['policy']['init_noise_std']
+)
+
+# 加载模型参数
+model.load_state_dict(model_dict['model_state_dict'])
+model.eval()
+
+# 创建示例输入
+obs_dim = cfgs[1]['num_obs']  # 从obs_cfg获取观测维度
+dummy_input = torch.randn(1, obs_dim)
+
+# 转换为ONNX
+torch.onnx.export(
+    model,
+    dummy_input,
+    args.output,
+    input_names=['obs'],
+    output_names=['actions'],
+    dynamic_axes={
+        'obs': {0: 'batch_size'},
+        'actions': {0: 'batch_size'}
+    }
+)
+
+print("Model successfully converted to ONNX format")

sim/genesis/README_EN.md

@@ -0,0 +1,5 @@
+import pickle
+
+with open('./logs/zeroth-walking/cfgs.pkl', 'rb') as f:
+    data = pickle.load(f)
+    print(data)

sim/genesis/utils/convert_to_onnx.py

@@ -39,14 +39,14 @@ def get_train_cfg(exp_name, max_iterations):
             "load_run": -1,
             "log_interval": 1,
             "max_iterations": max_iterations,
-            "num_steps_per_env": 24,
+            "num_steps_per_env": 48,
             "policy_class_name": "ActorCritic",
             "record_interval": -1,
             "resume": False,
             "resume_path": None,
             "run_name": "",
-            "runner_class_name": "runner_class_name",
-            "save_interval": 100,
+            "runner_class_name": "OnPolicyRunner",
+            "save_interval": 10,
         },
         "runner_class_name": "OnPolicyRunner",
         "seed": 1,