Save checkpoints and event files on Google Cloud Storage (#1316)

* Fix to be able to save checkpoints and event files on Google Cloud Storage

* Created custom normpath() function to handle the case of remote Cloud storages

tensorpack/callbacks/monitor.py

@@ -17,7 +17,7 @@ import threading
 from ..compat import tfv1 as tf
 from ..libinfo import __git_version__
 from ..tfutils.summary import create_image_summary, create_scalar_summary
-from ..utils import logger
+from ..utils import fs, logger
 from ..utils.develop import HIDE_DOC
 from .base import Callback
 
@@ -239,7 +239,7 @@ class TFEventWriter(MonitorBase):
         if logdir is None:
             logdir = logger.get_logger_dir()
         assert tf.gfile.IsDirectory(logdir), logdir
-        self._logdir = os.path.normpath(logdir)
+        self._logdir = fs.normpath(logdir)
         self._max_queue = max_queue
         self._flush_secs = flush_secs
         self._split_files = split_files

tensorpack/callbacks/saver.py

@@ -6,7 +6,7 @@ import os
 from datetime import datetime
 
 from ..compat import tfv1 as tf
-from ..utils import logger
+from ..utils import fs, logger
 from .base import Callback
 
 __all__ = ['ModelSaver', 'MinSaver', 'MaxSaver']
@@ -45,7 +45,7 @@ class ModelSaver(Callback):
         # If None, allow it to be init, but fail later if used
         # For example, if chief_only=True, it can still be safely initialized
         # in non-chief workers which don't have logger dir
-        self.checkpoint_dir = os.path.normpath(checkpoint_dir) if checkpoint_dir is not None else checkpoint_dir
+        self.checkpoint_dir = fs.normpath(checkpoint_dir) if checkpoint_dir is not None else checkpoint_dir
 
     def _setup_graph(self):
         assert self.checkpoint_dir is not None, \
@@ -121,7 +121,7 @@ class MinSaver(Callback):
         self.checkpoint_dir = checkpoint_dir
         if self.checkpoint_dir is None:
             self.checkpoint_dir = logger.get_logger_dir()
-        self.checkpoint_dir = os.path.normpath(self.checkpoint_dir)
+        self.checkpoint_dir = fs.normpath(self.checkpoint_dir)
 
     def _get_stat(self):
         try:

tensorpack/utils/fs.py

@@ -10,7 +10,7 @@ from six.moves import urllib
 from . import logger
 from .utils import execute_only_once
 
-__all__ = ['mkdir_p', 'download', 'recursive_walk', 'get_dataset_path']
+__all__ = ['mkdir_p', 'download', 'recursive_walk', 'get_dataset_path', 'normpath']
 
 
 def mkdir_p(dirname):
@@ -106,5 +106,19 @@ def get_dataset_path(*args):
     return os.path.join(d, *args)
 
 
+def normpath(path):
+    """
+    Normalizes a path to a folder by taking into consideration remote storages like Cloud storaged
+    referenced by '://' at the beginning of the path.
+
+    Args:
+        args: path to be normalized.
+
+    Returns:
+        str: normalized path.
+    """
+    return path if '://' in path else os.path.normpath(path)
+
+
 if __name__ == '__main__':
     download('http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz', '.')