[Feature] Publish models after training if published_keys is set in CheckpointHook

docs/en/tutorials/hook.md

@@ -70,6 +70,7 @@ runner.train()
 - Save the most recent checkpoints
 - Save the best checkpoints
 - Specify the path to save the checkpoints
+- Automatically publish the best and the last checkpoints
 
 For more features, please read the [CheckpointHook API documentation](mmengine.hooks.CheckpointHook).
 
@@ -120,6 +121,14 @@ The four features mentioned above are described below.
   default_hooks = dict(checkpoint=dict(type='CheckpointHook', interval=5, out_dir='/path/of/directory'))
   ```
 
+- Automatically publish the best and the last checkpoints
+
+  If you want to publish the best and the last checkpoints after training, you can set the `published_keys` parameter. You can select any keys in checkpoint to be published.
+
+  ```python
+  default_hooks = dict(checkpoint=dict(type='CheckpointHook', interval=1, save_best='accuracy', rule='less', published_keys=['meta', 'state_dict']))
+  ```
+
 [LoggerHook](mmengine.hooks.LoggerHook) collects logs from different components of `Runner` and write them to terminal, JSON file, tensorboard and wandb .etc.
 
 If we want to output (or save) the logs every 20 iterations, we can set the `interval` parameter and configure it as follows.

mmengine/hooks/checkpoint_hook.py

@@ -1,5 +1,7 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 import logging
+import subprocess
+import torch
 import os.path as osp
 from math import inf
 from pathlib import Path
@@ -84,7 +86,9 @@ class CheckpointHook(Hook):
         backend_args (dict, optional): Arguments to instantiate the
             prefix of uri corresponding backend. Defaults to None.
             New in v0.2.0.
-
+        published_keys (str, List[str], optional): If ``save_last`` is ``True``
+            or ``save_best`` is not ``None``, it will automatically
+            publish model with keys in list after train. Defaults to None.
     Examples:
         >>> # Save best based on single metric
         >>> CheckpointHook(interval=2, by_epoch=True, save_best='acc',
@@ -95,6 +99,9 @@ class CheckpointHook(Hook):
         >>> # Save best based on multi metrics with different comparison rule
         >>> CheckpointHook(interval=2, by_epoch=True,
         >>>                save_best=['FID', 'IS'], rule=['less', 'greater'])
+        >>> # Save best based on single metric and publish model after train
+        >>> CheckpointHook(interval=2, by_epoch=True, save_best='acc',
+        >>>                rule='less', published_keys=['meta', 'state_dict'])
     """
     out_dir: str
 
@@ -128,6 +135,7 @@ def __init__(self,
                  file_client_args: Optional[dict] = None,
                  filename_tmpl: Optional[str] = None,
                  backend_args: Optional[dict] = None,
+                 published_keys: Union[str, List[str], None] = None,
                  **kwargs) -> None:
         self.interval = interval
         self.by_epoch = by_epoch
@@ -218,6 +226,21 @@ def __init__(self,
             else:
                 self.best_ckpt_path_dict: Dict = dict()
 
+        # published keys
+        if not (isinstance(published_keys, str)
+                or is_list_of(published_keys, str)
+                or (published_keys is None)), (
+                    '"published_keys" should be a str or list of str or None, '
+                    f'but got {type(published_keys)}'):
+                    raise TypeError(...)
+
+        if isinstace(published_keys, str):
+            published_keys = [published_keys]    
+        elif isinstance(published_keys, list):
+            assert len(published_keys) == len(set(published_keys)), (
+                'Find duplicate element in "published_keys".')
+        self.published_keys = published_keys
+
     def before_train(self, runner) -> None:
         """Finish all operations, related to checkpoint.
 
@@ -304,6 +327,50 @@ def after_val_epoch(self, runner, metrics):
 
         self._save_best_checkpoint(runner, metrics)
 
+    def after_train(self, runner) -> None:
+        """Publish the checkpoint after train epoch.
+
+        Args:
+            runner (Runner): The runner of the training process.
+        """
+        if self.published_keys is None:
+            return
+
+        if self.save_last:
+            last_ckpt = runner.message_hub.get_info('last_ckpt')
+            assert last_ckpt ('Did not find last_checkpoint to be resumed.')
+            self._publish_model(runner, last_ckpt)
+
+        if self.save_best is not None:
+            if not self.best_ckpt_path:
+                raise RuntimeError(xxx)
+            self._publish_model(runner, best_ckpt)
+
+    def _publish_model(self, runner, out_file: str) -> None:
+
+        if published_keys is None:
+            return
+        checkpoint = runner.load_checkpoint(out_file)
+        ckpt_keys = list(checkpoint.keys())
+        published_keys = self.published_keys
+        for k in ckpt_keys:
+            if k not in published_keys:
+                print_log(
+                    f'Key `{k}` will be removed because it is not in '
+                    f'save_keys. If you want to keep it, '
+                    f'please set `{k}` in published_keys',
+                    logger='current')
+                checkpoint.pop(k)
+        out_file_name = osp.splitext(out_file)[0]
+        tmp_out_file_name = out_file + '.pth'
+        torch.save(checkpoint, tmp_out_file_name)
+        sha = subprocess.check_output(['sha256sum',
+                                       tmp_out_file_name]).decode()
+        final_file = out_file_name + f'-{sha[:8]}.pth'
+        subprocess.Popen(['mv', tmp_out_file_name, final_file])
+        print_log(
+            f'The published model is saved at {final_file}.', logger='current')
+
     def _save_checkpoint(self, runner) -> None:
         """Save the current checkpoint and delete outdated checkpoint.