refactor 4 (#2711)

* refactor ddp_spawn

pytorch_lightning/accelerator_backends/__init__.py

@@ -1,3 +1,4 @@
 from pytorch_lightning.accelerator_backends.gpu_backend import GPUBackend
 from pytorch_lightning.accelerator_backends.tpu_backend import TPUBackend
 from pytorch_lightning.accelerator_backends.dp_backend import DataParallelBackend
+from pytorch_lightning.accelerator_backends.ddp_spawn_backend import DDPSpawnBackend

pytorch_lightning/accelerator_backends/ddp_spawn_backend.py

@@ -0,0 +1,172 @@
+# Copyright The PyTorch Lightning team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License
+
+import os
+import torch
+import torch.multiprocessing as mp
+from pytorch_lightning.utilities.distributed import rank_zero_only
+from pytorch_lightning import _logger as log
+
+try:
+    from apex import amp
+except ImportError:
+    APEX_AVAILABLE = False
+else:
+    APEX_AVAILABLE = True
+
+
+class DDPSpawnBackend(object):
+
+    def __init__(self, trainer):
+        self.trainer = trainer
+        self.q = None
+
+    def setup(self):
+        self.trainer.set_random_port()
+
+        # pass in a state q
+        smp = mp.get_context('spawn')
+        self.q = smp.SimpleQueue()
+
+    def train(self, model, nprocs):
+        mp.spawn(self.ddp_train, nprocs=nprocs, args=(self.q, model,))
+
+    def teardown(self, model):
+        # restore main state with best weights
+        best_path = self.q.get()
+        results = self.q.get()
+        last_path = self.q.get()
+
+        # transfer back the best path to the trainer
+        self.trainer.checkpoint_callback.best_model_path = best_path
+
+        # load last weights
+        if last_path is not None and not self.trainer.testing:
+            ckpt = torch.load(last_path, map_location=lambda storage, loc: storage)
+            model.load_state_dict(ckpt)
+
+        self.trainer.model = model
+        return results
+
+    def ddp_train(self, process_idx, q, model, is_master=False, proc_offset=0):
+        """
+        Entry point for ddp
+
+        Args:
+            process_idx:
+            q:
+            model:
+            is_master:
+            proc_offset:
+
+        Returns:
+
+        """
+        # offset the process id if requested
+        process_idx = process_idx + proc_offset
+
+        # show progressbar only on progress_rank 0
+        if (self.trainer.node_rank != 0 or process_idx != 0) and self.trainer.progress_bar_callback is not None:
+            self.trainer.progress_bar_callback.disable()
+
+        # determine which process we are and world size
+        if self.trainer.use_ddp:
+            self.trainer.local_rank = process_idx
+            self.trainer.global_rank = self.trainer.node_rank * self.trainer.num_processes + process_idx
+            self.trainer.world_size = self.trainer.num_nodes * self.trainer.num_processes
+
+        elif self.trainer.use_ddp2:
+            self.trainer.local_rank = self.trainer.node_rank
+            self.trainer.global_rank = self.trainer.node_rank
+            self.trainer.world_size = self.trainer.num_nodes
+
+        # set warning rank
+        rank_zero_only.rank = self.trainer.global_rank
+
+        # set up server using proc 0's ip address
+        # try to init for 20 times at max in case ports are taken
+        # where to store ip_table
+        model.trainer = self.trainer
+        model.init_ddp_connection(
+            self.trainer.global_rank,
+            self.trainer.world_size,
+            self.trainer.is_slurm_managing_tasks
+        )
+
+        # call setup after the ddp process has connected
+        if not self.trainer.testing:
+            self.trainer.setup('fit')
+            model.setup('fit')
+
+        # on world_size=0 let everyone know training is starting
+        if self.trainer.is_global_zero:
+            log.info('-' * 100)
+            log.info(f'distributed_backend={self.trainer.distributed_backend}')
+            log.info(f'All DDP processes registered. Starting ddp with {self.trainer.world_size} processes')
+            log.info('-' * 100)
+
+        # CHOOSE OPTIMIZER
+        # allow for lr schedulers as well
+        optimizers, lr_schedulers, optimizer_frequencies = self.trainer.init_optimizers(model)
+        self.trainer.optimizers = optimizers
+        self.trainer.lr_schedulers = lr_schedulers
+        self.trainer.optimizer_frequencies = optimizer_frequencies
+
+        # MODEL
+        # copy model to each gpu
+        if self.trainer.on_gpu:
+            gpu_idx = process_idx
+            if is_master:
+                # source of truth is cuda for gpu idx
+                gpus = os.environ['CUDA_VISIBLE_DEVICES'].split(',')
+                gpu_idx = int(gpus[self.trainer.local_rank])
+
+            self.trainer.root_gpu = gpu_idx
+            torch.cuda.set_device(self.trainer.root_gpu)
+            model.cuda(self.trainer.root_gpu)
+
+        # set model properties before going into wrapper
+        self.trainer.copy_trainer_model_properties(model)
+
+        # AMP
+        # run through amp wrapper before going to distributed DP
+        # TODO: remove with dropping NVIDIA AMP support
+        native_amp_available = hasattr(torch.cuda, "amp") and hasattr(torch.cuda.amp, "autocast")
+        if self.trainer.use_amp and not native_amp_available:
+            model, optimizers = model.configure_apex(amp, model, self.trainer.optimizers, self.trainer.amp_level)
+            self.trainer.optimizers = optimizers
+            self.trainer.reinit_scheduler_properties(self.trainer.optimizers, self.trainer.lr_schedulers)
+
+        # DDP2 uses all GPUs on the machine
+        if self.trainer.distributed_backend == 'ddp' or self.trainer.distributed_backend == 'ddp_spawn':
+            device_ids = [self.trainer.root_gpu]
+        elif self.trainer.use_ddp2:
+            device_ids = self.trainer.data_parallel_device_ids
+        else:  # includes ddp_cpu
+            device_ids = None
+
+        # allow user to configure ddp
+        model = model.configure_ddp(model, device_ids)
+
+        # continue training routine
+        results = self.trainer.run_pretrain_routine(model)
+
+        # get original model
+        model = self.trainer.get_model()
+
+        # persist info in ddp_spawn
+        self.trainer.transfer_ddp_spawn_state_on_fit_end(model, q, results)
+
+        # clean up memory
+        torch.cuda.empty_cache()

pytorch_lightning/trainer/trainer.py

@@ -51,7 +51,7 @@ from pytorch_lightning.utilities import parsing, rank_zero_info, rank_zero_only,
 from pytorch_lightning.utilities.debugging import InternalDebugger
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.trainer.configuration_validator import ConfigValidator
-from pytorch_lightning.accelerator_backends import GPUBackend, TPUBackend, DataParallelBackend
+from pytorch_lightning.accelerator_backends import GPUBackend, TPUBackend, DataParallelBackend, DDPSpawnBackend
 
 # warnings to ignore in trainer
 warnings.filterwarnings(
@@ -1055,10 +1055,16 @@ class Trainer(
                 self.ddp_train(process_idx=task, q=None, model=model)
 
             elif self.distributed_backend == 'ddp_cpu':
-                results = self.__run_ddp_spawn(model, nprocs=self.num_processes)
+                self.accelerator_backend = DDPSpawnBackend(self)
+                self.accelerator_backend.setup()
+                self.accelerator_backend.train(model, nprocs=self.num_processes)
+                results = self.accelerator_backend.teardown(model)
 
             elif self.distributed_backend == 'ddp_spawn':
-                results = self.__run_ddp_spawn(model, nprocs=self.num_processes)
+                self.accelerator_backend = DDPSpawnBackend(self)
+                self.accelerator_backend.setup()
+                self.accelerator_backend.train(model, nprocs=self.num_processes)
+                results = self.accelerator_backend.teardown(model)
 
             elif self.distributed_backend == 'ddp':
                 self.set_random_port()
@@ -1116,31 +1122,6 @@ class Trainer(
         # used for testing or when we need to know that training succeeded
         return results or 1
 
-    def __run_ddp_spawn(self, model, nprocs):
-        self.set_random_port()
-
-        # pass in a state q
-        smp = mp.get_context('spawn')
-        q = smp.SimpleQueue()
-
-        mp.spawn(self.ddp_train, nprocs=nprocs, args=(q, model,))
-
-        # restore main state with best weights
-        best_path = q.get()
-        results = q.get()
-        last_path = q.get()
-
-        # transfer back the best path to the trainer
-        self.checkpoint_callback.best_model_path = best_path
-
-        # load last weights
-        if last_path is not None and not self.testing:
-            ckpt = torch.load(last_path, map_location=lambda storage, loc: storage)
-            model.load_state_dict(ckpt)
-
-        self.model = model
-        return results
-
     def can_prepare_data(self):
         if self.prepare_data_per_node:
             return self.local_rank == 0