Add non-existing resume_from_checkpoint acceptance for auto-resubmit

pytorch_lightning/trainer/connectors/checkpoint_connector.py

@@ -52,12 +52,12 @@ def __init__(self, trainer):
         # used to validate checkpointing logic
         self.has_trained = False
 
-    def restore_weights(self, model: LightningModule):
+    def restore_weights(self, model: LightningModule) -> None:
         """
-        We attempt to restore weights in this order:
-        1. HPC weights.
-        2. if no HPC weights restore checkpoint_path weights
-        3. otherwise don't restore weights
+        Attempt to restore state from checkpoint in this priority:
+        1. HPC weights
+        2. `resume_from_checkpoint` file
+        3. don't restore
         """
         # clear cache before restore
         if self.trainer.on_gpu:
@@ -70,9 +70,8 @@ def restore_weights(self, model: LightningModule):
         if self.trainer.on_gpu:
             torch.cuda.empty_cache()
 
-        if not did_restore_hpc_weights:
-            if self.trainer.resume_from_checkpoint is not None:
-                self.restore(self.trainer.resume_from_checkpoint, on_gpu=self.trainer.on_gpu)
+        if (not did_restore_hpc_weights) and (self.trainer.resume_from_checkpoint is not None):
+            self.restore(self.trainer.resume_from_checkpoint, on_gpu=self.trainer.on_gpu)
 
         # wait for all to catch up
         self.trainer.accelerator_backend.barrier('TrainerIOMixin.restore_weights')
@@ -81,16 +80,25 @@ def restore_weights(self, model: LightningModule):
         if self.trainer.on_gpu:
             torch.cuda.empty_cache()
 
-    def restore(self, checkpoint_path: str, on_gpu: bool):
+    def restore(self, checkpoint_path: str, on_gpu: bool) -> bool:
         """
-        Restore training state from checkpoint.
+        Try to restore training state from checkpoint.
         Also restores all training state like:
         - epoch
         - callbacks
         - schedulers
         - optimizer
+
+        Returns:
+           `True` if restored successfully else `False`
         """
 
+        # Try to read the checkpoint file at `checkpoint_path`. If not exist, do not restore checkpoint.
+        fs = get_filesystem(checkpoint_path)
+        if not fs.exists(checkpoint_path):
+            log.info(f'failed to load model from checkpoint:{checkpoint_path}')
+            return False
+
         # if on_gpu:
         #     checkpoint = torch.load(checkpoint_path)
         # else:
@@ -122,6 +130,9 @@ def restore(self, checkpoint_path: str, on_gpu: bool):
         # load training state (affects trainer only)
         self.restore_training_state(checkpoint)
 
+        log.info(f'Model restored from: {checkpoint_path}')
+        return True
+
     def restore_training_state(self, checkpoint):
         """
         Restore trainer state.
@@ -186,7 +197,7 @@ def restore_training_state(self, checkpoint):
         for scheduler, lrs_state in zip(self.trainer.lr_schedulers, lr_schedulers):
             scheduler['scheduler'].load_state_dict(lrs_state)
 
-    def restore_hpc_weights_if_needed(self, model: LightningModule):
+    def restore_hpc_weights_if_needed(self, model: LightningModule) -> bool:
         """If there is a set of hpc weights, use as signal to restore model."""
         did_restore = False
 

pytorch_lightning/trainer/trainer.py

@@ -248,6 +248,7 @@ def __init__(
 
             resume_from_checkpoint: To resume training from a specific checkpoint pass in the path here.
                 This can be a URL.
+                If there isn't the checkpoint file at specified path, start training from scratch.
 
             sync_batchnorm: Synchronize batch norm layers between process groups/whole world.
 

requirements.txt

@@ -6,5 +6,5 @@ future>=0.17.1  # required for builtins in setup.py
 # pyyaml>=3.13
 PyYAML>=5.1  # OmegaConf requirement >=5.1
 tqdm>=4.41.0
-fsspec>=0.8.0
+fsspec[http]>=0.8.1
 tensorboard>=2.2.0

tests/models/test_restore.py

@@ -71,6 +71,18 @@ def test_model_properties_resume_from_checkpoint(tmpdir):
     trainer.fit(model)
 
 
+def test_try_resume_from_non_existing_checkpoint(tmpdir):
+    """ Test that trying to resume from non-existing `resume_from_checkpoint` fail without error."""
+    model = EvalModelTemplate()
+    checkpoint_callback = ModelCheckpoint(dirpath=tmpdir, monitor="early_stop_on", save_last=True)
+    trainer = Trainer(default_root_dir=tmpdir, max_epochs=2, logger=False, checkpoint_callback=checkpoint_callback)
+    # Generate checkpoint `last.ckpt` with template model
+    trainer.fit(model)
+    # `True` if resume/restore successfully else `False`
+    assert trainer.checkpoint_connector.restore(str(tmpdir / "last.ckpt"), trainer.on_gpu)
+    assert not trainer.checkpoint_connector.restore(str(tmpdir / "last_non_existing.ckpt"), trainer.on_gpu)
+
+
 class CaptureCallbacksBeforeTraining(Callback):
     callbacks = []