ref: ddp backend refactor (3) (#3208)

* ddp backend refactor

* ddp backend refactor

pytorch_lightning/accelerators/cpu_backend.py

@@ -36,8 +36,10 @@ class CPUBackend(Accelerator):
         self.trainer.optimizers = optimizers
         self.trainer.lr_schedulers = lr_schedulers
         self.trainer.optimizer_frequencies = optimizer_frequencies
+        self.trainer.model = model
 
-    def train(self, model):
+    def train(self):
+        model = self.trainer.model
         results = self.trainer.run_pretrain_routine(model)
         return results
 

pytorch_lightning/accelerators/ddp2_backend.py

@@ -46,6 +46,8 @@ class DDP2Backend(Accelerator):
     def setup(self, model):
         self._resolve_task_idx()
 
+        self.trainer.model = model
+
     def _resolve_task_idx(self):
         if self.trainer.is_slurm_managing_tasks:
             self.task_idx = int(os.environ['SLURM_LOCALID'])
@@ -57,7 +59,8 @@ class DDP2Backend(Accelerator):
                 m = 'ddp2 only works in SLURM or via torchelastic with the WORLD_SIZE, LOCAL_RANK, GROUP_RANK flags'
                 raise MisconfigurationException(m)
 
-    def train(self, model):
+    def train(self):
+        model = self.trainer.model
         self.ddp_train(process_idx=self.task_idx, mp_queue=None, model=model)
 
     def ddp_train(self, process_idx, mp_queue, model, is_master=False, proc_offset=0):

pytorch_lightning/accelerators/ddp_backend.py

@@ -57,13 +57,16 @@ class DDPBackend(Accelerator):
         elif self.mode == 'torchelastic_ddp':
             self.__torchelastic_setup()
 
+        self.trainer.model = model
+
     def __slurm_setup(self):
         self.task_idx = int(os.environ['SLURM_LOCALID'])
 
     def __torchelastic_setup(self):
         self.task_idx = int(os.environ['LOCAL_RANK'])
 
-    def train(self, model):
+    def train(self):
+        model = self.trainer.model
         self.ddp_train(process_idx=self.task_idx, mp_queue=None, model=model)
 
     def spawn_ddp_children(self, model):

pytorch_lightning/accelerators/ddp_spawn_backend.py

@@ -41,7 +41,11 @@ class DDPSpawnBackend(Accelerator):
         smp = mp.get_context('spawn')
         self.mp_queue = smp.SimpleQueue()
 
-    def train(self, model):
+        self.trainer.model = model
+
+    def train(self):
+        model = self.trainer.model
+
         # train in children process
         mp.spawn(self.ddp_train, nprocs=self.nprocs, args=(self.mp_queue, model,))
 

pytorch_lightning/accelerators/gpu_backend.py

@@ -46,9 +46,11 @@ class GPUBackend(Accelerator):
 
         if self.trainer.amp_backend == AMPType.APEX:
             model = self._setup_nvidia_apex(model)
-        return model
 
-    def train(self, model):
+        self.trainer.model = model
+
+    def train(self):
+        model = self.trainer.model
         results = self.trainer.run_pretrain_routine(model)
         return results
 

pytorch_lightning/accelerators/tpu_backend.py

@@ -54,6 +54,8 @@ class TPUBackend(Accelerator):
         smp = mp.get_context(self.start_method)
         self.mp_queue = smp.SimpleQueue()
 
+        self.trainer.model = model
+
     def teardown(self, model):
         # restore main state with best weights
         best_path = self.mp_queue.get()
@@ -75,8 +77,8 @@ class TPUBackend(Accelerator):
         self.__load_weights_on_main_process()
         return results
 
-    def train(self, model: LightningModule):
-        self.trainer.model = model
+    def train(self):
+        model = self.trainer.model
 
         # train
         if self.trainer.tpu_id is not None:

pytorch_lightning/trainer/trainer.py

@@ -1040,26 +1040,26 @@ class Trainer(
         if self.use_ddp2:
             self.accelerator_backend = DDP2Backend(self)
             self.accelerator_backend.setup(model)
-            results = self.accelerator_backend.train(model)
+            results = self.accelerator_backend.train()
             self.accelerator_backend.teardown()
 
         elif use_slurm_ddp:
             self.accelerator_backend = DDPBackend(self, mode='slurm_ddp')
             self.accelerator_backend.setup(model)
-            results = self.accelerator_backend.train(model)
+            results = self.accelerator_backend.train()
             self.accelerator_backend.teardown()
 
         elif use_torchelastic_ddp:
             self.accelerator_backend = DDPBackend(self, mode='torchelastic_ddp')
             self.accelerator_backend.setup(model)
-            results = self.accelerator_backend.train(model)
+            results = self.accelerator_backend.train()
             self.accelerator_backend.teardown()
 
         # regular ddp using .spawn
         elif use_ddp_spawn:
             self.accelerator_backend = DDPSpawnBackend(self, nprocs=self.num_processes)
             self.accelerator_backend.setup(model)
-            results = self.accelerator_backend.train(model)
+            results = self.accelerator_backend.train()
             self.accelerator_backend.teardown()
 
         # ddp
@@ -1082,20 +1082,20 @@ class Trainer(
 
         elif self.use_single_gpu:
             self.accelerator_backend = GPUBackend(self)
-            model = self.accelerator_backend.setup(model)
-            results = self.accelerator_backend.train(model)
+            self.accelerator_backend.setup(model)
+            results = self.accelerator_backend.train()
             self.accelerator_backend.teardown()
 
         elif self.use_tpu:
             self.accelerator_backend = TPUBackend(self)
             self.accelerator_backend.setup(model)
-            self.accelerator_backend.train(model)
+            self.accelerator_backend.train()
             self.accelerator_backend.teardown(model)
 
         else:
             self.accelerator_backend = CPUBackend(self)
             self.accelerator_backend.setup(model)
-            results = self.accelerator_backend.train(model)
+            results = self.accelerator_backend.train()
             self.accelerator_backend.teardown()
 
         # hook