fix adamw apply gradient (#30130) (#30207)

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

@@ -29,10 +29,12 @@ class TestAdamWOp(unittest.TestCase):
             parameters=linear.parameters(),
             apply_decay_param_fun=lambda name: True,
             weight_decay=0.01)
-        out = linear(a)
-        out.backward()
-        adam.step()
-        adam.clear_gradients()
+
+        for _ in range(2):
+            out = linear(a)
+            out.backward()
+            adam.step()
+            adam.clear_gradients()
 
     def test_adamw_op_coverage(self):
         paddle.disable_static()

python/paddle/optimizer/adam.py

@@ -16,6 +16,7 @@ from .optimizer import Optimizer
 from ..fluid import core
 from ..fluid import framework
 from ..fluid.framework import Variable
+from ..fluid.dygraph import base as imperative_base
 
 import paddle
 
@@ -247,6 +248,7 @@ class Adam(Optimizer):
 
         return adam_op
 
+    @imperative_base.no_grad
     @framework.dygraph_only
     def step(self):
         """

python/paddle/optimizer/adamw.py

@@ -129,6 +129,7 @@ class AdamW(Adam):
         self._params_name = set()
         self._apply_decay_param_fun = apply_decay_param_fun
         self._coeff = coeff
+        self._lr_to_coeff = dict()
         super(AdamW, self).__init__(
             learning_rate=learning_rate,
             parameters=parameters,
@@ -139,97 +140,48 @@ class AdamW(Adam):
             name=name,
             lazy_mode=lazy_mode)
 
-    def _scale_parameters(self, params_and_grads):
+    def _append_decoupled_weight_decay(self, block, param_and_grad):
         """
-        Adds weight decay ops.
-            scaled_parameter = parameter * coeff
+        Add decoupled weight decay op.
+            parameter = parameter - parameter * coeff * lr
 
         Args:
-            params_and_grads: A list of (parameters, gradients) pairs,
+            block: block in which variable is to be created
+            param_and_grad: (parameters, gradients) pairs,
                 the parameters need to decay.
         Raises:
             Exception: The type of coeff and parameter is not consistent.
         """
-
-        scaled_params = []
-        for param, grad in params_and_grads:
-            # If no gradient then we don't need to do anything
-            if grad is None:
-                continue
-            if self._apply_decay_param_fun is not None \
-                    and not self._apply_decay_param_fun(param.name):
-                continue
-
-            if isinstance(self._coeff, float):
-                assert param.dtype is not paddle.fluid.core.VarDesc.VarType.FP32, \
-                    "the type of coeff(float) and parameter(%s) is not consistent."%(self._coeff.dtype)
-            else:
-                assert self._coeff.dtype == param.dtype, \
-                    "the type of coeff(%s) and parameter(%s) is not consistent."%(self._coeff.dtype, param.dtype)
-            if isinstance(self._learning_rate, float):
-                learning_rate = self._learning_rate
-            else:
-                learning_rate = self._learning_rate()
-            with param.block.program._optimized_guard(
-                [param, grad]), framework.name_scope('weight decay'):
-                scaled_params.append(
-                    (param, grad, param * self._coeff * learning_rate))
-                if param.name not in self._params_name:
-                    self._params_name.add(param.name)
-                    param = param * self._coeff
-        return scaled_params
-
-    @imperative_base.no_grad
-    def minimize(self,
-                 loss,
-                 startup_program=None,
-                 parameters=None,
-                 no_grad_set=None):
-        parameters = parameters if parameters \
-            else self._parameter_list
-
-        params_grads = self.backward(
-            loss=loss,
-            startup_program=startup_program,
-            parameters=parameters,
-            no_grad_set=no_grad_set)
-        scaled_params = self._scale_parameters(params_grads)
-        for p_grad_sgrad in scaled_params:
-            param, grad, scaled_param = p_grad_sgrad
-            with param.block.program._optimized_guard(
-                [param, grad]), framework.name_scope('weight decay'):
-                updated_param = paddle.fluid.layers.elementwise_sub(
-                    x=param, y=scaled_param)
-                paddle.fluid.layers.assign(input=updated_param, output=param)
-
-        optimize_ops = self._apply_optimize(
-            loss=loss,
-            params_grads=params_grads,
-            startup_program=startup_program)
-        return optimize_ops, params_grads
-
-    @framework.dygraph_only
-    @imperative_base.no_grad
-    def step(self):
-        self._dtype = None
-        params_grads = []
-        for param in self._parameter_list:
-            if not param.trainable:
-                continue
-            if param._grad_ivar() is not None:
-                grad_var = param._grad_ivar()
-                params_grads.append((param, grad_var))
-
-        scaled_params = self._scale_parameters(params_grads)
-        for p_grad_sgrad in scaled_params:
-            param, grad, scaled_param = p_grad_sgrad
-            with param.block.program._optimized_guard(
-                [param, grad]), framework.name_scope('weight decay'):
-                updated_param = paddle.fluid.layers.elementwise_sub(
-                    x=param, y=scaled_param)
-                paddle.fluid.layers.assign(input=updated_param, output=param)
-        self._apply_optimize(
-            loss=None, startup_program=None, params_grads=params_grads)
+        param, grad = param_and_grad
+
+        if self._apply_decay_param_fun is not None \
+                and not self._apply_decay_param_fun(param.name):
+            return
+
+        if isinstance(self._learning_rate, float):
+            learning_rate = self._learning_rate
+        else:
+            # NOTE. We add this function to the _append_optimize_op(),
+            # for we must make sure _create_param_lr() be called after
+            # optimizer._create_global_learning_rate().
+            learning_rate = self._create_param_lr(param_and_grad)
+
+        with block.program._optimized_guard(
+            [param, grad]), framework.name_scope('weight decay'):
+            self._params_name.add(param.name)
+
+            # If it has been calculated, the result will be reused
+            decay_coeff = self._lr_to_coeff.get(learning_rate, None)
+            if decay_coeff is None:
+                decay_coeff = 1.0 - learning_rate * self._coeff
+                self._lr_to_coeff[learning_rate] = decay_coeff
+
+            scaled_param = param * decay_coeff
+            paddle.fluid.layers.assign(input=scaled_param, output=param)
+
+    def _append_optimize_op(self, block, param_and_grad):
+        self._append_decoupled_weight_decay(block, param_and_grad)
+        return super(AdamW, self)._append_optimize_op(block, param_and_grad)
 
     def __str__(self):
         return " ".join(["Weight Decay, params:", ",".join(self._params_name)])