Add logging to ClearML

CONTRIBUTORS.md

@@ -33,6 +33,7 @@ Please keep the lists sorted alphabetically.
 * Eric Vollenweider
 * Fabian Jenelten
 * Lorenzo Terenzi
+* Maciej Aleksandrowicz
 * Marko Bjelonic
 * Matthijs van der Boon
 * Özhan Özen

README.md

@@ -44,6 +44,7 @@ The package supports the following logging frameworks which can be configured th
 * Tensorboard: https://www.tensorflow.org/tensorboard/
 * Weights & Biases: https://wandb.ai/site
 * Neptune: https://docs.neptune.ai/
+* ClearML: https://clear.ml/docs/latest/docs/
 
 For a demo configuration of PPO, please check the [example_config.yaml](config/example_config.yaml) file.
 

config/example_config.yaml

@@ -11,9 +11,10 @@ runner:
   experiment_name: walking_experiment
   run_name: ""
   # Logging writer
-  logger: tensorboard  # tensorboard, neptune, wandb
+  logger: tensorboard  # tensorboard, neptune, wandb, clearml
   neptune_project: legged_gym
   wandb_project: legged_gym
+  clearml_project: legged_gym
 
   # Policy
   policy:

rsl_rl/utils/clearml_utils.py

@@ -0,0 +1,77 @@
+# Copyright (c) 2021-2026, ETH Zurich and NVIDIA CORPORATION
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+from __future__ import annotations
+
+import os
+from dataclasses import asdict
+from torch.utils.tensorboard import SummaryWriter
+
+try:
+    from clearml import Task
+except ModuleNotFoundError:
+    raise ModuleNotFoundError("clearml package is required to log to ClearML.") from None
+
+
+class ClearmlSummaryWriter(SummaryWriter):
+    """Summary writer for ClearML."""
+
+    def __init__(self, log_dir: str, flush_secs: int, cfg: dict) -> None:
+        super().__init__(log_dir, flush_secs)
+
+        # Get the run name
+        run_name = os.path.split(log_dir)[-1]
+
+        # Get ClearML task
+        try:
+            project_name = cfg["clearml_project"]
+        except KeyError:
+            raise KeyError("Please specify clearml_project in the runner config, e.g. `legged_gym`.") from None
+
+        # Initialize ClearML Task
+        self.task = Task.init(
+            project_name=project_name, task_name=run_name, auto_connect_frameworks={"tensorboard": False}
+        )
+
+    def store_config(self, env_cfg: dict | object, train_cfg: dict) -> None:
+        runner_cfg = dict(train_cfg)
+        runner_cfg.pop("policy", None)
+        runner_cfg.pop("algorithm", None)
+
+        if isinstance(env_cfg, dict):
+            env_dict = env_cfg
+        else:
+            env_dict = env_cfg.to_dict() if hasattr(env_cfg, "to_dict") else asdict(env_cfg)
+
+        self.task.connect(runner_cfg, name="runner_cfg")
+        self.task.connect(train_cfg.get("policy", {}), name="policy_cfg")
+        self.task.connect(train_cfg.get("algorithm", {}), name="alg_cfg")
+        self.task.connect(env_dict, name="env_cfg")
+
+    def add_scalar(
+        self,
+        tag: str,
+        scalar_value: float,
+        global_step: int | None = None,
+        walltime: float | None = None,
+        new_style: bool = False,
+    ) -> None:
+        super().add_scalar(
+            tag,
+            scalar_value,
+            global_step=global_step,
+            walltime=walltime,
+            new_style=new_style,
+        )
+        self.task.get_logger().report_scalar(tag, "series", scalar_value, iteration=global_step)
+
+    def stop(self) -> None:
+        self.task.close()
+
+    def save_model(self, model_path: str, it: int) -> None:
+        self.task.upload_artifact(name=f"model_{it}", artifact_object=model_path)
+
+    def save_file(self, path: str) -> None:
+        self.task.upload_artifact(name=os.path.basename(path), artifact_object=path)

rsl_rl/utils/logger.py

@@ -19,6 +19,8 @@
 class Logger:
     """Logger to save the learning metrics to different logging services."""
 
+    LOGGER_TYPES = ("wandb", "neptune", "clearml")
+
     def __init__(
         self,
         log_dir: str | None,
@@ -64,7 +66,7 @@ def __init__(
         self._store_code_state()
 
         # Log configuration
-        if self.writer and not self.disable_logs and self.logger_type in ["wandb", "neptune"]:
+        if self.writer and not self.disable_logs and self.logger_type in Logger.LOGGER_TYPES:
             self.writer.store_config(env_cfg, self.cfg)
 
     def process_env_step(
@@ -230,11 +232,11 @@ def log(
 
     def save_model(self, path: str, it: int) -> None:
         """Save the model to external logging services if specified."""
-        if self.writer and not self.disable_logs and self.logger_type in ["neptune", "wandb"]:
+        if self.writer and not self.disable_logs and self.logger_type in Logger.LOGGER_TYPES:
             self.writer.save_model(path, it)
 
     def _prepare_logging_writer(self) -> None:
-        """Prepare the logging writer, which can be either Tensorboard, W&B or Neptune."""
+        """Prepare the logging writer, which can be either Tensorboard, W&B, Neptune or ClearML."""
         if self.log_dir is not None and not self.disable_logs:
             self.logger_type = self.cfg.get("logger", "tensorboard")
             self.logger_type = self.logger_type.lower()
@@ -247,12 +249,16 @@ def _prepare_logging_writer(self) -> None:
                 from rsl_rl.utils.wandb_utils import WandbSummaryWriter
 
                 self.writer = WandbSummaryWriter(log_dir=self.log_dir, flush_secs=10, cfg=self.cfg)
+            elif self.logger_type == "clearml":
+                from rsl_rl.utils.clearml_utils import ClearmlSummaryWriter
+
+                self.writer = ClearmlSummaryWriter(log_dir=self.log_dir, flush_secs=10, cfg=self.cfg)
             elif self.logger_type == "tensorboard":
                 from torch.utils.tensorboard import SummaryWriter
 
                 self.writer = SummaryWriter(log_dir=self.log_dir, flush_secs=10)
             else:
-                raise ValueError("Logger type not found. Please choose 'wandb', 'neptune', or 'tensorboard'.")
+                raise ValueError(f"Logger type not found. Please choose one of: {', '.join(Logger.LOGGER_TYPES)}.")
         else:
             self.writer = None
 
@@ -285,6 +291,6 @@ def _store_code_state(self) -> None:
                 file_paths.append(diff_file_name)
 
             # Upload diff files to external logging services
-            if self.writer and self.logger_type in ["wandb", "neptune"] and file_paths:
+            if self.writer and self.logger_type in Logger.LOGGER_TYPES and file_paths:
                 for path in file_paths:
                     self.writer.save_file(path)