Average local optimizer param after warmup and during bmuf sync

Summary: We have seen that averaging the local param instead of doing reset or broadcast after warmup improves the WER.

Reviewed By: skritika

Differential Revision: D16739278

fbshipit-source-id: 75033d2d25f9a88fd6dd325d0d9d4c856d22d947

fairseq/optim/adam.py

@@ -8,6 +8,7 @@ import types
 
 import torch
 import torch.optim
+import torch.distributed as dist
 
 from . import FairseqOptimizer, register_optimizer
 
@@ -53,6 +54,17 @@ class FairseqAdam(FairseqOptimizer):
             'weight_decay': self.args.weight_decay,
         }
 
+    def average_params(self):
+        """Reduce Params is only used during BMUF distributed training."""
+        state_dict = self.optimizer.state_dict()
+        total_gpus = float(dist.get_world_size())
+
+        for _, value in state_dict["state"].items():
+            value["exp_avg"] /= total_gpus
+            value["exp_avg_sq"] /= total_gpus
+            dist.all_reduce(value["exp_avg"], op=dist.ReduceOp.SUM)
+            dist.all_reduce(value["exp_avg_sq"], op=dist.ReduceOp.SUM)
+
 
 class Adam(torch.optim.Optimizer):
     """Implements Adam algorithm.

fairseq/optim/bmuf.py

@@ -31,6 +31,7 @@ class FairseqBMUF(FairseqOptimizer):
         self.warmup_iteration = self.args.warmup_iterations
         self.use_nbm = self.args.use_nbm
         self.initial_state = self._optimizer.state_dict()
+        self.average_sync = self.args.average_sync
 
     @staticmethod
     def add_args(parser):
@@ -62,6 +63,12 @@ class FairseqBMUF(FairseqOptimizer):
             action="store_true",
             help="Specify whether you want to use classical BM / Nesterov BM",
         )
+        parser.add_argument(
+            "--average-sync",
+            default=True,
+            action="store_true",
+            help="Specify whether you want to average the local momentum after each sync",
+        )
 
     @property
     def optimizer(self):
@@ -91,34 +98,50 @@ class FairseqBMUF(FairseqOptimizer):
         """Clips gradient norm."""
         return self._optimizer.clip_grad_norm(max_norm)
 
+    def average_params(self):
+        self._optimizer.average_params()
+
     def _block_sync(self):
-        # Update the global model using local models from all GPUs.
-        if self._is_bmuf_iter():
+        # Update the global model using local models from all GPUs
+        # (Step-1) Calculate grad between previously synced model and
+        # currrent local model
         if self.block_momentum != 0:
-                self._BM_before_sync()
+            self._calc_grad()
 
-            self._allreduce_parameter()
+        # (Step-2) Average gradient from all GPUs
+        self._avg_grad_from_all_gpus()
 
+        # (Step-3) Calculate global momentum and update the global model
         if self.block_momentum != 0:
-                self._BM_after_sync()
+            self._update_global_model()
+
+        # (Step-4) Average local optimizer params
+        if self.average_sync:
+            self.average_params()
 
     def _is_warmup_end(self):
+        # Check whether train iterations is equal to warmup iter
         if self.get_num_updates() == self.warmup_iteration:
             return True
         return False
 
     def _is_bmuf_iter(self):
+        # Check whether train iterations is equal to bmuf sync iter
         if self.get_num_updates() % self.sync_iter == 0:
             return True
         return False
 
-    def _warmup_sync(self, rootRank=0):
-        # broadcast the local model to all GPUs
+    def _warmup_sync(self, root_rank=0):
+        # Broadcast the local model to all gpus
         for param in self.params:
-            dist.broadcast(param.data, src=rootRank)
+            dist.broadcast(param.data, src=root_rank)
 
-        # Reset the local optimizer state and local bmuf related param
+        # Update local optimizer state
+        if self.average_sync:
+            self._optimizer.average_params()
+        else:
             self._optimizer.load_state_dict(self.initial_state)
+
         self._reset_local_data()
 
     def step(self, closure=None):
@@ -127,7 +150,7 @@ class FairseqBMUF(FairseqOptimizer):
         self.set_num_updates(self.get_num_updates() + 1)
         if self._is_warmup_end():
             self._warmup_sync()
-        else:
+        elif self._is_bmuf_iter():
             self._block_sync()
 
     def zero_grad(self):
@@ -144,61 +167,56 @@ class FairseqBMUF(FairseqOptimizer):
 
     @torch.no_grad()
     def _reset_local_data(self):
-        """Resetting all the BMUF specific params."""
-        self.params_localprev = [torch.zeros_like(p.data) for p in self.params]
-
-        self.smoothed_grads_localprev = [
-            p.data.new_zeros(p.data.size()) for p in self.params
-        ]
-        self.grads_localprev = [p.data.new_zeros(p.data.size()) for p in self.params]
+        # (Step-0) Initialize global momentum parameters and store global copy on each gpu
+        self.global_params = [torch.zeros_like(p.data) for p in self.params]
+        self.smoothed_grads = [p.data.new_zeros(p.data.size()) for p in self.params]
+        self.grads = [p.data.new_zeros(p.data.size()) for p in self.params]
 
         # saving the global model locally for calculating gradient during bmuf sync
-        for param, copy_param in zip(self.params, self.params_localprev):
-            copy_param.copy_(param.data)
+        for param, global_param in zip(self.params, self.global_params):
+            global_param.copy_(param.data)
 
     @torch.no_grad()
-    def _BM_before_sync(self):
-        """Calculate grad between previously synced model and currrent local model."""
-        # prev_param is basically the global copy from the previously finished
+    def _calc_grad(self):
+        # global_params is basically the global copy from the previously finished
         # synchronisation. param.data is local parameter after block_sync_freq
         # for the local gpu. so grad is difference between previously synced
         # model and currrent local model.
-        for index, (param, prev_param) in enumerate(
-            zip(self.params, self.params_localprev)
+        for index, (param, global_param) in enumerate(
+            zip(self.params, self.global_params)
         ):
-            self.grads_localprev[index] = prev_param - param.data
+            self.grads[index] = global_param - param.data
 
-    def _allreduce_parameter(self):
-        """Average gradient from all the GPUs. """
+    def _avg_grad_from_all_gpus(self):
         for index, param in enumerate(self.params):
-            sync_para = (
-                param.data if self.block_momentum == 0 else self.grads_localprev[index]
-            )
+            sync_para = param.data if self.block_momentum == 0 else self.grads[index]
             sync_para /= float(dist.get_world_size())
             dist.all_reduce(sync_para, op=dist.ReduceOp.SUM)
 
     @torch.no_grad()
-    def _BM_after_sync(self):
-        for index, (param, prev_param, smoothed_grad, grad) in enumerate(
+    def _update_global_model(self):
+        for index, (param, global_param, smoothed_grad, grad) in enumerate(
             zip(
                 self.params,
-                self.params_localprev,
-                self.smoothed_grads_localprev,
-                # all machines would share the same value of smoothed_grad, since it is
+                self.global_params,
+                self.smoothed_grads,
+                # all gpus would share the same value of smoothed_grad, since it is
                 # always computed on synchronized gradients.
-                self.grads_localprev,
+                self.grads,
             )
         ):
-            # prev_param is basically last syncrhornized parameter. though
+            # global_param is basically last syncrhornized parameter. though
             # smoothed_grad is local, all processes will have same value of
             # smoothed_grad and hence param is globally synchronized copy.
-            # smoothed_grad(t)=BM * smoothed_grad(t-1) + BM_lr*grad(t)
-            smoothed_grad = smoothed_grad * self.block_momentum + grad * self.block_lr
-            param.data.copy_(prev_param - smoothed_grad)
+            # smoothed_grad(t) = BM * smoothed_grad(t-1) + BM_lr * grad(t)
+            smoothed_grad = self.block_momentum * smoothed_grad + self.block_lr * grad
+            param.data.copy_(global_param - smoothed_grad)
+
             # A Nesterov momentum here is to do a partial weight update before
             # calculating the gradient
             if self.use_nbm:
                 param.data.copy_(param.data - self.block_momentum * smoothed_grad)
+
             # backup for the next synchronization.
-            self.smoothed_grads_localprev[index] = smoothed_grad
-            prev_param.copy_(param.data)
+            self.smoothed_grads[index] = smoothed_grad
+            global_param.copy_(param.data)

fairseq/optim/fairseq_optimizer.py

@@ -108,3 +108,6 @@ class FairseqOptimizer(object):
         if hasattr(self.optimizer, 'supports_memory_efficient_fp16'):
             return self.optimizer.supports_memory_efficient_fp16
         return False
+
+    def average_params(self):
+        pass