New apex amp API (#1465)

* use new apex amp API
* make apex opt_level as option

onmt/model_builder.py

@@ -214,8 +214,6 @@ def fix_key(s):
 
     model.generator = generator
     model.to(device)
-    if model_opt.model_dtype == 'fp16':
-        model.half()
 
     return model
 

onmt/opts.py

@@ -179,6 +179,10 @@ def model_opts(parser):
     group.add('--loss_scale', '-loss_scale', type=float, default=0,
               help="For FP16 training, the static loss scale to use. If not "
                    "set, the loss scale is dynamically computed.")
+    group.add('--apex_opt_level', '-apex_opt_level', type=str, default="O2",
+              choices=["O0", "O1", "O2", "O3"],
+              help="For FP16 training, the opt_level to use."
+                   "See https://nvidia.github.io/apex/amp.html#opt-levels.")
 
 
 def preprocess_opts(parser):

onmt/trainer.py

@@ -294,8 +294,7 @@ def validate(self, valid_iter, moving_average=None):
             valid_model = deepcopy(self.model)
             for avg, param in zip(self.moving_average,
                                   valid_model.parameters()):
-                param.data = avg.data.half() if self.model_dtype == "fp16" \
-                    else avg.data
+                param.data = avg.data
         else:
             valid_model = self.model
 

onmt/utils/optimizers.py

@@ -7,8 +7,6 @@
 from copy import copy
 from math import sqrt
 
-from onmt.utils.misc import fn_args
-
 
 def build_torch_optimizer(model, opt):
     """Builds the PyTorch optimizer.
@@ -87,17 +85,14 @@ def build_torch_optimizer(model, opt):
 
     if opt.model_dtype == 'fp16':
         import apex
-        static_loss_scale = opt.loss_scale
-        dynamic_loss_scale = opt.loss_scale == 0
-        # TODO: clean this up when APEX unify its optimizer API.
-        if opt.optim.startswith('fused'):
-            namespace = apex.optimizers  # Faster wrapper.
-        else:
-            namespace = apex.fp16_utils
-        optimizer = namespace.FP16_Optimizer(
+        loss_scale = "dynamic" if opt.loss_scale == 0 else opt.loss_scale
+        model, optimizer = apex.amp.initialize(
+            [model, model.generator],
             optimizer,
-            static_loss_scale=static_loss_scale,
-            dynamic_loss_scale=dynamic_loss_scale)
+            opt_level=opt.apex_opt_level,
+            loss_scale=loss_scale,
+            keep_batchnorm_fp32=False if opt.optim == "fusedadam" else None)
+
     return optimizer
 
 
@@ -317,10 +312,9 @@ def backward(self, loss):
         """Wrapper for backward pass. Some optimizer requires ownership of the
         backward pass."""
         if self._with_fp16_wrapper:
-            kwargs = {}
-            if "update_master_grads" in fn_args(self._optimizer.backward):
-                kwargs["update_master_grads"] = True
-            self._optimizer.backward(loss, **kwargs)
+            import apex
+            with apex.amp.scale_loss(loss, self._optimizer) as scaled_loss:
+                scaled_loss.backward()
         else:
             loss.backward()
 
@@ -336,7 +330,9 @@ def step(self):
                 self._optimizer.update_master_grads()
             if hasattr(self._optimizer, "clip_master_grads") and \
                self._max_grad_norm > 0:
-                self._optimizer.clip_master_grads(self._max_grad_norm)
+                import apex
+                torch.nn.utils.glip_grad_norm_(
+                    apex.amp.master_params(self), self._max_grad_norm)
         for group in self._optimizer.param_groups:
             group['lr'] = learning_rate
             if not self._with_fp16_wrapper and self._max_grad_norm > 0:

onmt/utils/parse.py

@@ -63,10 +63,6 @@ def validate_model_opts(cls, model_opt):
             if model_opt.model_type != "text":
                 raise AssertionError(
                     "--share_embeddings requires --model_type text.")
-        if model_opt.model_dtype == "fp16":
-            logger.warning(
-                "FP16 is experimental, the generated checkpoints may "
-                "be incompatible with a future version")
 
     @classmethod
     def ckpt_model_opts(cls, ckpt_opt):