Fixes the checkpoint directory structure for pytorch and pytorch lightning

horovod/spark/common/store.py

@@ -30,6 +30,7 @@
 import fsspec
 from fsspec.core import split_protocol
 from fsspec.utils import update_storage_options
+from fsspec.callbacks import _DEFAULT_CALLBACK
 
 from horovod.spark.common.util import is_databricks, host_hash
 
@@ -248,7 +249,7 @@ def get_run_path(self, run_id):
         return os.path.join(self.get_runs_path(), run_id)
 
     def get_checkpoint_path(self, run_id):
-        return os.path.join(self.get_run_path(run_id), self.get_checkpoint_filename()) \
+        return self.get_run_path(run_id) \
             if self._save_runs else None
 
     def get_checkpoints(self, run_id, suffix='.ckpt'):
@@ -311,10 +312,38 @@ def sync_fn(self, run_id):
 
         def fn(local_run_path):
             print(f"Syncing dir {local_run_path} to dir {run_path}")
-            self.fs.put(local_run_path, run_path, recursive=True, overwrite=True)
+            self.copy(local_run_path, run_path, recursive=True, overwrite=True)
 
         return fn
 
+    def copy(self, lpath, rpath, recursive=False, callback=_DEFAULT_CALLBACK,**kwargs):
+        """ 
+        This method copies the contents of the local source directory to the target directory.
+        This is different from the fsspec's put() because it does not copy the source folder 
+        to the target directory in the case when target directory already exists.
+        """
+
+        from fsspec.implementations.local import LocalFileSystem, make_path_posix
+        from fsspec.utils import other_paths
+
+        rpath = (
+            self.fs._strip_protocol(rpath)
+            if isinstance(rpath, str)
+            else [self.fs._strip_protocol(p) for p in rpath]
+        )
+        if isinstance(lpath, str):
+            lpath = make_path_posix(lpath)
+        fs = LocalFileSystem()
+        lpaths = fs.expand_path(lpath, recursive=recursive)
+        rpaths = other_paths(
+            lpaths, rpath
+        )
+
+        callback.set_size(len(rpaths))
+        for lpath, rpath in callback.wrap(zip(lpaths, rpaths)):
+            callback.branch(lpath, rpath, kwargs)
+            self.fs.put_file(lpath, rpath, **kwargs)
+
     def get_filesystem(self):
         return self.fs
 

horovod/spark/keras/estimator.py

@@ -14,7 +14,7 @@
 # ==============================================================================
 import numbers
 import time
-
+import os
 import numpy as np
 import tensorflow as tf
 
@@ -282,7 +282,10 @@ def _fit_on_prepared_data(self, backend, train_rows, val_rows, metadata, avg_row
 
     def _load_model_from_checkpoint(self, run_id):
         store = self.getStore()
-        last_ckpt_path = store.get_checkpoint_path(run_id)
+        last_ckpt_path = os.path.join(store.get_checkpoint_path(run_id), store.get_checkpoint_filename())
+
+        if not store.fs.exists(last_ckpt_path):
+            return None
 
         if self.getVerbose():
             print('Resuming training from last checkpoint: {}'.format(last_ckpt_path))

horovod/spark/lightning/estimator.py

@@ -463,6 +463,10 @@ def _fit_on_prepared_data(self, backend, train_rows, val_rows, metadata, avg_row
     def _read_checkpoint(self, run_id):
         store = self.getStore()
         checkpoints = store.get_checkpoints(run_id, suffix='.ckpt')
+
+        if not checkpoints:
+            return None
+
         last_ckpt_path = checkpoints[-1]
 
         if self.getVerbose():

horovod/spark/torch/estimator.py

@@ -19,6 +19,7 @@
 import io
 import numbers
 import time
+import os
 
 from pyspark import keyword_only
 from pyspark.ml.param.shared import Param, Params, TypeConverters
@@ -287,7 +288,10 @@ def _fit_on_prepared_data(self, backend, train_rows, val_rows, metadata, avg_row
 
     def _load_checkpoint(self, run_id):
         store = self.getStore()
-        last_ckpt_path = store.get_checkpoint_path(run_id)
+        last_ckpt_path = os.path.join(store.get_checkpoint_path(run_id),store.get_checkpoint_filename())
+
+        if not store.fs.exists(last_ckpt_path):
+            return None
 
         if self.getVerbose():
             print('Resuming training from last checkpoint: {}'.format(last_ckpt_path))

test/integration/test_spark_lightning.py

@@ -215,11 +215,8 @@ def test_legacy_fit_model(self):
                 assert len(pred) == 1
                 assert pred.dtype == torch.float32
 
-    # TODO: Add this test back after checkpoint call back is supported
     def test_restore_from_checkpoint(self):
-        self.skipTest('There is a deadlock bug for checkpoint call back. ' +
-                      'Will add this test back when it is solved.')
-
+
         model = create_xor_model()
 
         with spark_session('test_restore_from_checkpoint') as spark:
@@ -253,10 +250,7 @@ def test_restore_from_checkpoint(self):
                 torch_estimator.fit(df)
                 torch_estimator._read_checkpoint.assert_called()
 
-    # TODO: Add this test back after checkpoint call back is supported
     def test_legacy_restore_from_checkpoint(self):
-        self.skipTest('There is a deadlock bug for checkpoint call back. ' +
-                      'Will add this test back when it is solved.')
 
         model = create_legacy_xor_model()
         optimizer = torch.optim.SGD(model.parameters(), lr=0.1)