[AMP]split minimize and add unscale_ for GradScaler (#35825)

* split minimize() to step() + update()

* add unscale and step for grad_scaler

* add unittest

* refine code in minimize

* delete step in loss_scaler

* fix example bug

* refine comment

* refine unittest

* add unittest

python/paddle/amp/grad_scaler.py

@@ -13,18 +13,28 @@
 # limitations under the License.
 
 from paddle.fluid.dygraph.amp import AmpScaler
+from paddle.fluid.dygraph.amp import OptimizerState
+from collections import defaultdict
 
 __all__ = []
 
 
+def _refresh_optimizer_state():
+    return {"state": OptimizerState.INIT}
+
+
 class GradScaler(AmpScaler):
     """
     GradScaler is used for Auto-Mixed-Precision training in dynamic graph mode. 
     It controls the scaling of loss, helps avoiding numerical overflow.
-    The object of this class has two methods `scale()`, `minimize()`.
+    The object of this class has nineteen methods `scale()`, `unscale_()`, `minimize()`, `step()`, `update()` and `get`/`set` api of parameters.
 
     `scale()` is used to multiply the loss by a scale ratio.
-    `minimize()` is similar as `optimizer.minimize()`, performs parameters updating.
+    `unscale_()` is used to unscale the gradients of parameters, multiplies the gradients of parameters by 1/(scale ratio)
+    `minimize()` is similar as `optimizer.minimize()`, performs parameters updating, and it will update the loss_scaling, it equal to `step()` + `update()`.
+    `step()` is similar as `optimizer.step()`, which performs parameters updating.
+    `update` is used to update the loss_scaling.
+
 
     Commonly, it is used together with `paddle.amp.auto_cast` to achieve Auto-Mixed-Precision in 
     dynamic graph mode.
@@ -115,7 +125,7 @@ class GradScaler(AmpScaler):
         This function is similar as `optimizer.minimize()`, which performs parameters updating.
         
         If the scaled gradients of parameters contains NAN or INF, the parameters updating is skipped.
-        Otherwise, it first unscales the scaled gradients of parameters, then updates the parameters.
+        Otherwise, if `unscale_()` has not been called, it first unscales the scaled gradients of parameters, then updates the parameters.
 
         Finally, the loss scaling ratio is updated.
 
@@ -151,16 +161,18 @@ class GradScaler(AmpScaler):
         This function is similar as `optimizer.step()`, which performs parameters updating.
         
         If the scaled gradients of parameters contains NAN or INF, the parameters updating is skipped.
-        Otherwise, it first unscales the scaled gradients of parameters, then updates the parameters.
+        Otherwise, if `unscale_()` has not been called, it first unscales the scaled gradients of parameters, then updates the parameters.
 
         Args:
             optimizer(Optimizer):  The optimizer used to update parameters.
 
         Examples:
+
             .. code-block:: python
             
                 # required: gpu
                 import paddle
+
                 model = paddle.nn.Conv2D(3, 2, 3, bias_attr=True)
                 optimizer = paddle.optimizer.SGD(learning_rate=0.01, parameters=model.parameters())
                 scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
@@ -170,13 +182,20 @@ class GradScaler(AmpScaler):
                     loss = paddle.mean(conv)
                 scaled = scaler.scale(loss)  # scale the loss 
                 scaled.backward()            # do backward
-                scaler.step(optimizer)
+                scaler.step(optimizer)       # update parameters
+                scaler.update()              # update the loss scaling ratio
                 optimizer.clear_grad()
         """
         if not self._enable:
             return optimizer.step()
 
+        optimizer_state = self._optimizer_states[id(optimizer)]
+        if optimizer_state["state"] is OptimizerState.STEPPED:
+            raise RuntimeError(
+                "step() has already been called since the last update().")
+
         #  unscale the grad
+        if optimizer_state["state"] is OptimizerState.INIT:
             self._unscale(optimizer)
 
         if self._found_inf:
@@ -185,9 +204,75 @@ class GradScaler(AmpScaler):
             optimizer.step()
             self._cache_founf_inf = False
 
+        optimizer_state["state"] = OptimizerState.STEPPED
+
+        if not self._use_dynamic_loss_scaling:
+            self._optimizer_states = defaultdict(_refresh_optimizer_state)
+
+    def update(self):
+        """
+        Updates the loss_scaling.
+        
+        Examples:
+
+            .. code-block:: python
+            
+                # required: gpu
+                import paddle
+
+                model = paddle.nn.Conv2D(3, 2, 3, bias_attr=True)
+                optimizer = paddle.optimizer.SGD(learning_rate=0.01, parameters=model.parameters())
+                scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
+                data = paddle.rand([10, 3, 32, 32])
+                with paddle.amp.auto_cast():
+                    conv = model(data)
+                    loss = paddle.mean(conv)
+                scaled = scaler.scale(loss)     # scale the loss 
+                scaled.backward()               # do backward
+                scaler.step(optimizer)          # update parameters
+                scaler.update()                 # update the loss scaling ratio
+                optimizer.clear_grad() 
+        """
+        if not self._enable:
+            return
         if self._use_dynamic_loss_scaling:
-            # uopdate the scale
             self._update()
+            self._optimizer_states = defaultdict(_refresh_optimizer_state)
+        return
+
+    def unscale_(self, optimizer):
+        """
+        Unscale the gradients of parameters, multiplies the gradients of parameters by 1/(loss scaling ratio).  
+        If this instance of :class:`GradScaler` is not enabled, output are returned unmodified.
+
+        Args:
+            optimizer(Optimizer):  The optimizer used to update parameters.
+
+        Returns:
+            The unscaled parameters or original parameters.
+        
+        Examples:
+
+            .. code-block:: python
+
+                # required: gpu
+                import paddle
+
+                model = paddle.nn.Conv2D(3, 2, 3, bias_attr=True)
+                optimizer = paddle.optimizer.SGD(learning_rate=0.01, parameters=model.parameters())
+                scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
+                data = paddle.rand([10, 3, 32, 32])
+                with paddle.amp.auto_cast():
+                    conv = model(data)
+                    loss = paddle.mean(conv)
+                scaled = scaler.scale(loss)  # scale the loss 
+                scaled.backward()            # do backward
+                scaler.unscale_(optimizer)    # unscale the parameter
+                scaler.step(optimizer)
+                scaler.update()  
+                optimizer.clear_grad() 
+        """
+        return super(GradScaler, self)._unscale(optimizer)
 
     def is_enable(self):
         """

python/paddle/fluid/dygraph/amp/loss_scaler.py

@@ -21,8 +21,20 @@ from ...wrapped_decorator import signature_safe_contextmanager, wrap_decorator
 import warnings
 import numpy as np
 from paddle import _C_ops
+from collections import defaultdict
+from enum import Enum
 
-__all__ = ['AmpScaler']
+__all__ = ['AmpScaler', 'OptimizerState']
+
+
+class OptimizerState(Enum):
+    INIT = 0
+    UNSCALED = 1
+    STEPPED = 2
+
+
+def _refresh_optimizer_state():
+    return {"state": OptimizerState.INIT}
 
 
 class AmpScaler(object):
@@ -31,10 +43,11 @@ class AmpScaler(object):
 
     AmpScaler is used for Auto-Mixed-Precision training/inferring in imperative
     mode. It controls the scaling of loss, helps avoiding numerical overflow.
-    The object of this class has two methods `scale()`, `minimize()`.
+    The object of this class has seventeen methods `scale()`, `unscale_()`, `minimize()` and `get`/`set` api of parameters.
 
     `scale()` is used to multiply the loss by a scale ratio.
-    `minimize()` is similar as `Optimizer.minimize()`, performs parameters updating.
+    `unscale_()` is used to unscale the gradients of parameters, multiplies the gradients of parameters by 1/(scale ratio)
+    `minimize()` is similar as `optimizer.minimize()`, performs parameters updating, and it will update the loss_scaling.
 
     Commonly, it is used together with `amp_guard` to achieve Auto-Mixed-Precision in 
     imperative mode.
@@ -117,6 +130,7 @@ class AmpScaler(object):
             self._scale = to_variable(
                 np.array([self._init_loss_scaling]).astype(np.float32))
             self._cache_founf_inf = None
+            self._optimizer_states = defaultdict(_refresh_optimizer_state)
 
     def scale(self, var):
         """
@@ -129,6 +143,7 @@ class AmpScaler(object):
             The scaled variable or original variable.
         
         Examples:
+
             .. code-block:: python
 
                 import numpy as np
@@ -160,7 +175,7 @@ class AmpScaler(object):
         This function is similar as `Optimizer.minimize()`, which performs parameters updating.
         
         If the scaled gradients of parameters contains NAN or INF, the parameters updating is skipped.
-        Otherwise, it first unscales the scaled gradients of parameters, then updates the parameters.
+        Otherwise, if `unscale_()` has not been called, it first unscales the scaled gradients of parameters, then updates the parameters.
 
         Finally, the loss scaling ratio is updated.
 
@@ -170,6 +185,7 @@ class AmpScaler(object):
             kwargs: Keyword arguments, which will be forward to `Optimizer.minimize()`.
 
         Examples:
+
             .. code-block:: python
 
                 import numpy as np
@@ -192,7 +208,10 @@ class AmpScaler(object):
         if not self._enable:
             return optimizer.minimize(*args, **kwargs)
 
+        optimizer_state = self._optimizer_states[id(optimizer)]
+
         #  unscale the grad
+        if optimizer_state["state"] is OptimizerState.INIT:
             self._unscale(optimizer)
 
         optimize_ops, params_grads = (None, None)
@@ -207,12 +226,31 @@ class AmpScaler(object):
             # uopdate the scale
             self._update()
 
+        self._optimizer_states = defaultdict(_refresh_optimizer_state)
+
         return optimize_ops, params_grads
 
     def _unscale(self, optimizer):
+        """
+        Unscale the gradients of parameters, multiplies the gradients of parameters by 1/(loss scaling ratio).  
+        If this instance of :class:`GradScaler` is not enabled, output are returned unmodified.
+        Args:
+            optimizer(Optimizer):  The optimizer used to update parameters.
+        Returns:
+            The unscaled parameters or original parameters.
+        """
         if not self._enable:
             return
 
+        optimizer_state = self._optimizer_states[id(optimizer)]
+
+        if optimizer_state["state"] is OptimizerState.UNSCALED:
+            raise RuntimeError(
+                "unscale_() has already been called on this optimizer since the last update()."
+            )
+        elif optimizer_state["state"] is OptimizerState.STEPPED:
+            raise RuntimeError("unscale_() is being called after step().")
+
         if getattr(optimizer, '_param_groups', None) and isinstance(
                 optimizer._param_groups[0], dict):
             param_grads = []
@@ -256,6 +294,8 @@ class AmpScaler(object):
                                             temp_found_inf_fp32)
         self._found_inf = temp_found_inf_fp16 or temp_found_inf_fp32
 
+        optimizer_state["state"] = OptimizerState.UNSCALED
+
     def _update(self):
         """
         Updates the loss_scaling.

python/paddle/fluid/tests/unittests/test_imperative_auto_mixed_precision.py

@@ -222,6 +222,47 @@ class TestAmpScaler(unittest.TestCase):
                 np.allclose(outs_with_scaler[1][i][0].numpy(),
                             outs_no_scaler[1][i][0].numpy()), True)
 
+    def test_step(self):
+        inp_np = np.random.random(size=[1, 3, 128, 128]).astype(np.float32)
+
+        def run_simple_conv(inp_np, use_scaler=True):
+            paddle.seed(10)
+            paddle.framework.random._manual_program_seed(10)
+            with fluid.dygraph.guard():
+                model = SimpleConv(
+                    num_channels=3,
+                    num_filters=64,
+                    filter_size=7,
+                    stride=2,
+                    act='relu')
+                optimizer = paddle.optimizer.SGD(learning_rate=0.01,
+                                                 parameters=model.parameters())
+                scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
+                data = fluid.dygraph.to_variable(inp_np)
+
+                out = model(data)
+                loss = fluid.layers.mean(out)
+                if use_scaler:
+                    print('use scaler')
+                    scaled_loss = scaler.scale(loss)
+                    scaled_loss.backward()
+                    scaler.step(optimizer)
+                    scaler.update()
+                else:
+                    print('use no scaler')
+                    loss.backward()
+                    optimizer.step()
+            return optimizer._parameter_list
+
+        outs_with_scaler = run_simple_conv(inp_np, use_scaler=True)
+        outs_no_scaler = run_simple_conv(inp_np, use_scaler=False)
+
+        for i in range(len(outs_with_scaler)):
+            # check each parameter
+            self.assertEqual(
+                np.allclose(outs_with_scaler[i].numpy(),
+                            outs_no_scaler[i].numpy()), True)
+
     def test_nan_inf(self):
         inp_np = np.random.random(size=[1, 3, 128, 128]).astype(np.float32)
         inp_np[0][1][2][3] = np.nan
@@ -252,6 +293,52 @@ class TestAmpScaler(unittest.TestCase):
                 self.assertTrue(
                     np.array_equal(param.numpy(), params_init[param.name]))
 
+    def test_step_update_exception(self):
+        def func1():
+            model = paddle.nn.Conv2D(3, 2, 3, bias_attr=True)
+            optimizer = paddle.optimizer.SGD(learning_rate=0.01,
+                                             parameters=model.parameters())
+            scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
+            data = paddle.rand([10, 3, 32, 32])
+            conv = model(data)
+            loss = paddle.mean(conv)
+            scaled = scaler.scale(loss)
+            scaled.backward()
+            scaler.unscale_(optimizer)
+            scaler.unscale_(optimizer)
+
+        self.assertRaises(RuntimeError, func1)
+
+        def func2():
+            model = paddle.nn.Conv2D(3, 2, 3, bias_attr=True)
+            optimizer = paddle.optimizer.SGD(learning_rate=0.01,
+                                             parameters=model.parameters())
+            scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
+            data = paddle.rand([10, 3, 32, 32])
+            conv = model(data)
+            loss = paddle.mean(conv)
+            scaled = scaler.scale(loss)
+            scaled.backward()
+            scaler.step(optimizer)
+            scaler.unscale_(optimizer)
+
+        self.assertRaises(RuntimeError, func2)
+
+        def func3():
+            model = paddle.nn.Conv2D(3, 2, 3, bias_attr=True)
+            optimizer = paddle.optimizer.SGD(learning_rate=0.01,
+                                             parameters=model.parameters())
+            scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
+            data = paddle.rand([10, 3, 32, 32])
+            conv = model(data)
+            loss = paddle.mean(conv)
+            scaled = scaler.scale(loss)
+            scaled.backward()
+            scaler.step(optimizer)
+            scaler.step(optimizer)
+
+        self.assertRaises(RuntimeError, func3)
+
     def test_get_and_set(self):
         with fluid.dygraph.guard():
             scaler = paddle.amp.GradScaler(
@@ -838,8 +925,9 @@ class TestResnet2(unittest.TestCase):
 
             scaled_loss = scaler.scale(avg_loss)
             scaled_loss.backward()
-
+            scaler.unscale_(optimizer)
             scaler.step(optimizer)
+            scaler.update()
 
             dy_grad_value = {}
             for param in resnet.parameters():