add amp example (#7062)

examples/train_resnet_amp.py

@@ -0,0 +1,35 @@
+from train_resnet_base import TrainResNetBase
+
+import itertools
+
+import torch_xla.distributed.xla_multiprocessing as xmp
+import torch_xla.core.xla_model as xm
+from torch_xla.amp import autocast
+
+
+# For more details check https://github.com/pytorch/xla/blob/master/docs/amp.md
+class TrainResNetXLAAMP(TrainResNetBase):
+
+  def train_loop_fn(self, loader, epoch):
+    tracker = xm.RateTracker()
+    self.model.train()
+    loader = itertools.islice(loader, self.num_steps)
+    for step, (data, target) in enumerate(loader):
+      self.optimizer.zero_grad()
+      # Enables autocasting for the forward pass
+      with autocast(xm.xla_device()):
+        output = self.model(data)
+        loss = self.loss_fn(output, target)
+      # TPU amp uses bf16 hence gradient scaling is not necessary. If runnign with XLA:GPU
+      # check https://github.com/pytorch/xla/blob/master/docs/amp.md#amp-for-xlagpu.
+      loss.backward()
+      self.run_optimizer()
+      tracker.add(self.batch_size)
+      if step % 10 == 0:
+        xm.add_step_closure(
+            self._train_update, args=(step, loss, tracker, epoch))
+
+
+if __name__ == '__main__':
+  xla_amp = TrainResNetXLAAMP()
+  xla_amp.start_training()

examples/train_resnet_base.py

@@ -13,10 +13,6 @@
 import torch.nn as nn
 
 
-def _train_update(step, loss, tracker, epoch):
-  print(f'epoch: {epoch}, step: {step}, loss: {loss}, rate: {tracker.rate()}')
-
-
 class TrainResNetBase():
 
   def __init__(self):
@@ -37,29 +33,33 @@ def __init__(self):
     self.optimizer = optim.SGD(self.model.parameters(), weight_decay=1e-4)
     self.loss_fn = nn.CrossEntropyLoss()
 
+  def _train_update(self, step, loss, tracker, epoch):
+    print(f'epoch: {epoch}, step: {step}, loss: {loss}, rate: {tracker.rate()}')
+
   def run_optimizer(self):
     self.optimizer.step()
 
-  def start_training(self):
+  def train_loop_fn(self, loader, epoch):
+    tracker = xm.RateTracker()
+    self.model.train()
+    loader = itertools.islice(loader, self.num_steps)
+    for step, (data, target) in enumerate(loader):
+      self.optimizer.zero_grad()
+      output = self.model(data)
+      loss = self.loss_fn(output, target)
+      loss.backward()
+      self.run_optimizer()
+      tracker.add(self.batch_size)
+      if step % 10 == 0:
+        xm.add_step_closure(
+            self._train_update, args=(step, loss, tracker, epoch))
 
-    def train_loop_fn(loader, epoch):
-      tracker = xm.RateTracker()
-      self.model.train()
-      loader = itertools.islice(loader, self.num_steps)
-      for step, (data, target) in enumerate(loader):
-        self.optimizer.zero_grad()
-        output = self.model(data)
-        loss = self.loss_fn(output, target)
-        loss.backward()
-        self.run_optimizer()
-        tracker.add(self.batch_size)
-        if step % 10 == 0:
-          xm.add_step_closure(_train_update, args=(step, loss, tracker, epoch))
+  def start_training(self):
 
     for epoch in range(1, self.num_epochs + 1):
       xm.master_print('Epoch {} train begin {}'.format(
           epoch, time.strftime('%l:%M%p %Z on %b %d, %Y')))
-      train_loop_fn(self.train_device_loader, epoch)
+      self.train_loop_fn(self.train_device_loader, epoch)
       xm.master_print('Epoch {} train end {}'.format(
           epoch, time.strftime('%l:%M%p %Z on %b %d, %Y')))
     xm.wait_device_ops()