Merge branch 'master' into ci/versions

Author: awaelchli

.gitignore

@@ -17,6 +17,7 @@ docs/source/api
 docs/source/*.md
 docs/source/generated
 docs/source/*/generated
+docs/source/notebooks
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

CHANGELOG.md

@@ -56,9 +56,6 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added `ddp_fully_sharded` support ([#7487](https://github.com/PyTorchLightning/pytorch-lightning/pull/7487))
 
 
-- Added `__len__` to `IndexBatchSamplerWrapper` ([#7681](https://github.com/PyTorchLightning/pytorch-lightning/pull/7681))
-
-
 - Added `should_rank_save_checkpoint` property to Training Plugins ([#7684](https://github.com/PyTorchLightning/pytorch-lightning/pull/7684))
 
 
@@ -85,7 +82,6 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
-- Changed calling of `untoggle_optimizer(opt_idx)` out of the closure function ([#7563](https://github.com/PyTorchLightning/pytorch-lightning/pull/7563)
 
 - Changed the `Trainer`'s `checkpoint_callback` argument to allow only boolean values ([#7539](https://github.com/PyTorchLightning/pytorch-lightning/pull/7539))
 
@@ -190,6 +186,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Deprecated default value of `monitor` argument in EarlyStopping callback to enforce `monitor` as a required argument ([#7907](https://github.com/PyTorchLightning/pytorch-lightning/pull/7907))
 
 
+- Deprecated the use of `CheckpointConnector.hpc_load()` in favor of `CheckpointConnector.restore()` ([#7652](https://github.com/PyTorchLightning/pytorch-lightning/pull/7652))
+
+
 ### Removed
 
 - Removed `ProfilerConnector` ([#7654](https://github.com/PyTorchLightning/pytorch-lightning/pull/7654))
@@ -215,51 +214,68 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
-- Fixed `DataModule.prepare_data` could only be called on the global rank 0 process ([#7945](https://github.com/PyTorchLightning/pytorch-lightning/pull/7945))
+- Fixed setting `worker_init_fn` to seed dataloaders correctly when using DDP ([#7942](https://github.com/PyTorchLightning/pytorch-lightning/pull/7942))
+
 
-- Fixed `_check_training_step_output` to be called after `train_step_end` to support more flexible accomodations ([#7868](https://github.com/PyTorchLightning/pytorch-lightning/pull/7868))
+- Fixed `DataModule.prepare_data` could only be called on the global rank 0 process ([#7945](https://github.com/PyTorchLightning/pytorch-lightning/pull/7945))
 
-- Fixed `apply_to_collection` works on Custom Collections now ([#7851](https://github.com/PyTorchLightning/pytorch-lightning/pull/7851))
 
 - Fixed ambiguous warning when both overfit and train dataloader shuffling are enabled ([#7685](https://github.com/PyTorchLightning/pytorch-lightning/pull/7685))
 
-- Fixed dataloaders are not reset when tuning the model ([#7566](https://github.com/PyTorchLightning/pytorch-lightning/pull/7566))
 
+- Fixed dev debugger memory growing due to tracking events even when disabled ([#7875](https://github.com/PyTorchLightning/pytorch-lightning/pull/7875))
 
-- Fixed global step update when the epoch is skipped ([#7677](https://github.com/PyTorchLightning/pytorch-lightning/pull/7677))
 
+- Fixed `None` loss keys getting added in `training_epoch_end` when using manual optimization and not returning a loss ([#7772](https://github.com/PyTorchLightning/pytorch-lightning/pull/7772))
 
-- Fixed dev debugger memory growing due to tracking events even when disabled ([#7875](https://github.com/PyTorchLightning/pytorch-lightning/pull/7875))
 
+- Fixed a bug where `precision=64` with `accelerator='ddp_spawn'` would throw a pickle error ([#6924](https://github.com/PyTorchLightning/pytorch-lightning/pull/6924))
 
-- Fixed training loop total batch counter when accumulate grad batches was enabled ([#7692](https://github.com/PyTorchLightning/pytorch-lightning/pull/7692))
 
+- Fixed `dataloader_idx` argument value when predicting with only one `DataLoader` ([#7941](https://github.com/PyTorchLightning/pytorch-lightning/pull/7941))
 
-- Fixed broadcasting in multi-node, multi-gpu DDP using torch 1.7 ([#7592](https://github.com/PyTorchLightning/pytorch-lightning/pull/7592))
 
+## [1.3.5] - 2021-06-08
 
-- Fixed `ProgressBar` pickling after calling `trainer.predict` ([#7608](https://github.com/PyTorchLightning/pytorch-lightning/pull/7608))
+### Added
 
+- Added warning to Training Step output ([#7779](https://github.com/PyTorchLightning/pytorch-lightning/pull/7779))
 
-- Fixed `None` loss keys getting added in `training_epoch_end` when using manual optimization and not returning a loss ([#7772](https://github.com/PyTorchLightning/pytorch-lightning/pull/7772))
+### Fixed
 
+- Fixed `LearningRateMonitor` and `BackboneFinetuning` ([#7835](https://github.com/PyTorchLightning/pytorch-lightning/pull/7835))
+- Minor improvements to `apply_to_collection` and type signature of `log_dict` ([#7851](https://github.com/PyTorchLightning/pytorch-lightning/pull/7851))
+- Fixed docker versions ([#7834](https://github.com/PyTorchLightning/pytorch-lightning/pull/7834))
+- Fixed sharded training check for fp16 precision ([#7825](https://github.com/PyTorchLightning/pytorch-lightning/pull/7825))
+- Fixed support for torch Module type hints in LightningCLI ([#7807](https://github.com/PyTorchLightning/pytorch-lightning/pull/7807))
 
-- Fixed print errors in `ProgressBar` when `trainer.fit` is not called ([#7674](https://github.com/PyTorchLightning/pytorch-lightning/pull/7674))
+### Changed
 
+- Move `training_output` validation to after `train_step_end` ([#7868](https://github.com/PyTorchLightning/pytorch-lightning/pull/7868))
 
-- Fixed formatting of info message when max training time reached ([#7780](https://github.com/PyTorchLightning/pytorch-lightning/pull/7780))
 
+## [1.3.4] - 2021-06-01
 
-- Fixed a bug where `precision=64` with `accelerator='ddp_spawn'` would throw a pickle error ([#6924](https://github.com/PyTorchLightning/pytorch-lightning/pull/6924))
+### Fixed
+
+- Fixed info message when max training time reached ([#7780](https://github.com/PyTorchLightning/pytorch-lightning/pull/7780))
+- Fixed missing `__len__` method to `IndexBatchSamplerWrapper` ([#7681](https://github.com/PyTorchLightning/pytorch-lightning/pull/7681))
 
 
-- Fixed support for `torch.nn.Module` type hints in `LightningCLI` ([#7807](https://github.com/PyTorchLightning/pytorch-lightning/pull/7807))
+## [1.3.3] - 2021-05-27
 
+### Changed
 
-- Fixed a bug where checking `trainer.precision` changed to `'mixed'` when specifying 16 in trainer ([#7825](https://github.com/PyTorchLightning/pytorch-lightning/pull/7825))
+- Changed calling of `untoggle_optimizer(opt_idx)` out of the closure function ([#7563](https://github.com/PyTorchLightning/pytorch-lightning/pull/7563))
 
+### Fixed
 
-- Fixed `LearningRateMonitor` keys not properly setup when running with `BackboneFinetuning` Callback ([#7835](https://github.com/PyTorchLightning/pytorch-lightning/pull/7835))
+- Fixed `ProgressBar` pickling after calling `trainer.predict` ([#7608](https://github.com/PyTorchLightning/pytorch-lightning/pull/7608))
+- Fixed broadcasting in multi-node, multi-gpu DDP using torch 1.7 ([#7592](https://github.com/PyTorchLightning/pytorch-lightning/pull/7592))
+- Fixed dataloaders are not reset when tuning the model ([#7566](https://github.com/PyTorchLightning/pytorch-lightning/pull/7566))
+- Fixed print errors in `ProgressBar` when `trainer.fit` is not called ([#7674](https://github.com/PyTorchLightning/pytorch-lightning/pull/7674))
+- Fixed global step update when the epoch is skipped ([#7677](https://github.com/PyTorchLightning/pytorch-lightning/pull/7677))
+- Fixed training loop total batch counter when accumulate grad batches was enabled ([#7692](https://github.com/PyTorchLightning/pytorch-lightning/pull/7692))
 
 
 ## [1.3.2] - 2021-05-18

docs/source/conf.py

@@ -122,6 +122,10 @@ def _transform_changelog(path_in: str, path_out: str) -> None:
 nbsphinx_allow_errors = True
 nbsphinx_requirejs_path = ''
 
+# myst-parser, forcing to parse all html pages with mathjax
+# https://github.com/executablebooks/MyST-Parser/issues/394
+myst_update_mathjax = False
+
 # The suffix(es) of source filenames.
 # You can specify multiple suffix as a list of string:
 #

pytorch_lightning/trainer/connectors/checkpoint_connector.py

@@ -21,9 +21,14 @@
 
 import pytorch_lightning
 from pytorch_lightning.core.lightning import LightningModule
-from pytorch_lightning.utilities import _OMEGACONF_AVAILABLE, DeviceType, rank_zero_info, rank_zero_warn
+from pytorch_lightning.utilities import (
+    _OMEGACONF_AVAILABLE,
+    DeviceType,
+    rank_zero_deprecation,
+    rank_zero_info,
+    rank_zero_warn,
+)
 from pytorch_lightning.utilities.cloud_io import atomic_save, get_filesystem
-from pytorch_lightning.utilities.cloud_io import load as pl_load
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.upgrade_checkpoint import KEYS_MAPPING as DEPRECATED_CHECKPOINT_KEYS
 
@@ -45,7 +50,7 @@ def hpc_resume_path(self) -> Optional[str]:
         dir_path_hpc = str(self.trainer.weights_save_path)
         max_version = self.max_ckpt_version_in_folder(dir_path_hpc, "hpc_ckpt_")
         if max_version is not None:
-            return f"{dir_path_hpc}/hpc_ckpt_{max_version}.ckpt"
+            return os.path.join(dir_path_hpc, f"hpc_ckpt_{max_version}.ckpt")
 
     def resume_start(self) -> None:
         """
@@ -129,6 +134,10 @@ def restore_model_state(self, model: LightningModule, checkpoint) -> None:
         # hook: give user access to checkpoint if needed.
         model.on_load_checkpoint(checkpoint)
 
+        # call hpc specific hook
+        if self.hpc_resume_path is not None:
+            model.on_hpc_load(self._loaded_checkpoint)
+
         # restore model state_dict
         self.trainer.training_type_plugin.load_model_state_dict(checkpoint)
 
@@ -248,6 +257,7 @@ def restore_lr_schedulers(self) -> None:
     # ----------------------------------
     # PRIVATE OPS
     # ----------------------------------
+
     def hpc_save(self, folderpath: str, logger):
         # make sure the checkpoint folder exists
         folderpath = str(folderpath)  # because the tests pass a path object
@@ -365,29 +375,18 @@ def dump_checkpoint(self, weights_only: bool = False) -> dict:
 
         return checkpoint
 
-    def hpc_load(self, checkpoint_path: str, on_gpu: bool):
-        """
-        Load model/training states from a 'PyTorch-Lightning checkpoint' file for hpc.
-        All restored states are listed in return value description of `dump_checkpoint`.
+    def hpc_load(self, checkpoint_path: str) -> None:
         """
+        Attempts to restore the full training and model state from a HPC checkpoint file.
 
-        # read a checkpoint dictionary object from the 'PyTorch-Lightning checkpoint' file at `checkpoint_path`
-        checkpoint = pl_load(checkpoint_path, map_location=lambda storage, loc: storage)
-
-        # acquire the model
-        model = self.trainer.lightning_module
-
-        # restore model and datamodule state
-        self.restore_model_state(model, checkpoint)
-
-        if self.trainer.root_gpu is not None:
-            model.cuda(self.trainer.root_gpu)
-
-        # restore training state
-        self.restore_training_state(checkpoint)
-
-        # call hpc specific hook
-        model.on_hpc_load(checkpoint)
+        .. deprecated::v1.4
+            Will be removed in v1.6. Use :meth:`restore` instead.
+        """
+        rank_zero_deprecation(
+            "`CheckpointConnector.hpc_load()` was deprecated in v1.4 and will be removed in v1.6."
+            " Use `CheckpointConnector.restore()` instead."
+        )
+        self.restore(checkpoint_path)
 
     def max_ckpt_version_in_folder(self, dir_path: Union[str, Path], name_key: str = 'ckpt_') -> Optional[int]:
         """List up files in `dir_path` with `name_key`, then yield maximum suffix number.

pytorch_lightning/trainer/predict_loop.py

@@ -98,7 +98,7 @@ def _get_num_dataloaders(self, dataloaders: List[DataLoader]) -> int:
 
     def _build_kwargs(self, batch, batch_idx, dataloader_idx):
         step_kwargs = OrderedDict([('batch', batch), ('batch_idx', batch_idx)])
-        if self.num_dataloaders:
+        if self.num_dataloaders > 1:
             step_kwargs['dataloader_idx'] = dataloader_idx
         return step_kwargs
 

pytorch_lightning/utilities/seed.py

@@ -84,8 +84,9 @@ def reset_seed() -> None:
     If :func:`pytorch_lightning.utilities.seed.seed_everything` is unused, this function will do nothing.
     """
     seed = os.environ.get("PL_GLOBAL_SEED", None)
+    workers = os.environ.get("PL_SEED_WORKERS", False)
     if seed is not None:
-        seed_everything(int(seed))
+        seed_everything(int(seed), workers=bool(workers))
 
 
 def pl_worker_init_function(worker_id: int, rank: Optional = None) -> None:  # pragma: no cover
@@ -100,6 +101,9 @@ def pl_worker_init_function(worker_id: int, rank: Optional = None) -> None:  # p
     process_seed = torch.initial_seed()
     # back out the base seed so we can use all the bits
     base_seed = process_seed - worker_id
+    log.debug(
+        f'Initializing random number generators of process {global_rank} worker {worker_id} with base seed {base_seed}'
+    )
     ss = np.random.SeedSequence([base_seed, worker_id, global_rank])
     # use 128 bits (4 x 32-bit words)
     np.random.seed(ss.generate_state(4))

tests/checkpointing/test_legacy_checkpoints.py

@@ -65,6 +65,9 @@
         "1.3.0",
         "1.3.1",
         "1.3.2",
+        "1.3.3",
+        "1.3.4",
+        "1.3.5",
     ]
 )
 def test_resume_legacy_checkpoints(tmpdir, pl_version: str):

tests/deprecated_api/test_remove_1-4.py

@@ -66,3 +66,16 @@ def training_step(self, batch, batch_idx):
 
     with pytest.deprecated_call(match=r"Relying on.*is deprecated in v1.2 and will be removed in v1.4"):
         trainer.fit(TestModel())
+
+
+def test_v1_4_0_deprecated_hpc_load(tmpdir):
+    model = BoringModel()
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        max_steps=1,
+    )
+    trainer.fit(model)
+    trainer.checkpoint_connector.hpc_save(tmpdir, trainer.logger)
+    checkpoint_path = trainer.checkpoint_connector.get_max_ckpt_path_from_folder(str(tmpdir))
+    with pytest.deprecated_call(match=r"`CheckpointConnector.hpc_load\(\)` was deprecated in v1.4"):
+        trainer.checkpoint_connector.hpc_load(checkpoint_path)

tests/helpers/boring_model.py

@@ -161,20 +161,24 @@ def __init__(self, data_dir: str = './'):
         self.checkpoint_state: Optional[str] = None
 
     def prepare_data(self):
-        self.random_full = RandomDataset(32, 192)
+        self.random_full = RandomDataset(32, 64 * 4)
 
     def setup(self, stage: Optional[str] = None):
         if stage == "fit" or stage is None:
             self.random_train = Subset(self.random_full, indices=range(64))
             self.dims = self.random_train[0].shape
 
         if stage in ("fit", "validate") or stage is None:
-            self.random_val = Subset(self.random_full, indices=range(64, 128))
+            self.random_val = Subset(self.random_full, indices=range(64, 64 * 2))
 
         if stage == "test" or stage is None:
-            self.random_test = Subset(self.random_full, indices=range(128, 192))
+            self.random_test = Subset(self.random_full, indices=range(64 * 2, 64 * 3))
             self.dims = getattr(self, "dims", self.random_test[0].shape)
 
+        if stage == "predict" or stage is None:
+            self.random_predict = Subset(self.random_full, indices=range(64 * 3, 64 * 4))
+            self.dims = getattr(self, "dims", self.random_predict[0].shape)
+
     def train_dataloader(self):
         return DataLoader(self.random_train)
 
@@ -183,3 +187,6 @@ def val_dataloader(self):
 
     def test_dataloader(self):
         return DataLoader(self.random_test)
+
+    def predict_dataloader(self):
+        return DataLoader(self.random_predict)

tests/helpers/pipelines.py

@@ -91,7 +91,7 @@ def run_model_test(
         trainer.checkpoint_connector.hpc_save(save_dir, logger)
         # test HPC loading
         checkpoint_path = trainer.checkpoint_connector.get_max_ckpt_path_from_folder(save_dir)
-        trainer.checkpoint_connector.hpc_load(checkpoint_path, on_gpu=on_gpu)
+        trainer.checkpoint_connector.restore(checkpoint_path)
 
 
 @torch.no_grad()