Add interface so user can get scaled loss when they use customized loss (#20613)

paddle/fluid/API.spec

@@ -557,7 +557,7 @@ paddle.fluid.contrib.HDFSClient.upload (ArgSpec(args=['self', 'hdfs_path', 'loca
 paddle.fluid.contrib.multi_download (ArgSpec(args=['client', 'hdfs_path', 'local_path', 'trainer_id', 'trainers', 'multi_processes'], varargs=None, keywords=None, defaults=(5,)), ('document', '100927be598ed8f9eaa1f3ef1b23568a'))
 paddle.fluid.contrib.multi_upload (ArgSpec(args=['client', 'hdfs_path', 'local_path', 'multi_processes', 'overwrite', 'sync'], varargs=None, keywords=None, defaults=(5, False, True)), ('document', '183f34c83d30dbe16e09e8716c41958a'))
 paddle.fluid.contrib.extend_with_decoupled_weight_decay (ArgSpec(args=['base_optimizer'], varargs=None, keywords=None, defaults=None), ('document', 'a1095dfd4ec725747f662d69cd7659d4'))
-paddle.fluid.contrib.mixed_precision.decorate (ArgSpec(args=['optimizer', 'amp_lists', 'init_loss_scaling', 'incr_every_n_steps', 'decr_every_n_nan_or_inf', 'incr_ratio', 'decr_ratio', 'use_dynamic_loss_scaling'], varargs=None, keywords=None, defaults=(None, 1.0, 1000, 2, 2.0, 0.8, True)), ('document', '5f118631fc8632afb981b3a26daae731'))
+paddle.fluid.contrib.mixed_precision.decorate (ArgSpec(args=['optimizer', 'amp_lists', 'init_loss_scaling', 'incr_every_n_steps', 'decr_every_n_nan_or_inf', 'incr_ratio', 'decr_ratio', 'use_dynamic_loss_scaling'], varargs=None, keywords=None, defaults=(None, 1.0, 1000, 2, 2.0, 0.8, True)), ('document', '6b0a44eb05c8707c1eff2e786f673edb'))
 paddle.fluid.contrib.mixed_precision.AutoMixedPrecisionLists ('paddle.fluid.contrib.mixed_precision.fp16_lists.AutoMixedPrecisionLists', ('document', 'c116ec6bb5d30998792daea8db21ee40'))
 paddle.fluid.contrib.mixed_precision.AutoMixedPrecisionLists.__init__ (ArgSpec(args=['self', 'custom_white_list', 'custom_black_list'], varargs=None, keywords=None, defaults=(None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.fused_elemwise_activation (ArgSpec(args=['x', 'y', 'functor_list', 'axis', 'scale', 'save_intermediate_out'], varargs=None, keywords=None, defaults=(-1, 0.0, True)), ('document', '1c4b247a2858cea8d9d8750693688270'))

python/paddle/fluid/contrib/mixed_precision/decorator.py

@@ -58,6 +58,7 @@ class OptimizerWithMixedPrecison(object):
         self._param_grads = None
         self._train_program = default_main_program()
         self._startup_prog = default_startup_program()
+        self._scaled_loss = None
         self._loss_scaling = layers.create_global_var(
             name=unique_name.generate("loss_scaling"),
             shape=[1],
@@ -101,6 +102,13 @@ class OptimizerWithMixedPrecison(object):
         """
         return self._loss_scaling
 
+    def get_scaled_loss(self):
+        """Return the scaled loss.
+        It's useful when you feed customed loss into executor.
+        """
+
+        return self._scaled_loss
+
     def backward(self,
                  loss,
                  startup_program=None,
@@ -124,9 +132,9 @@ class OptimizerWithMixedPrecison(object):
             gradient respectively, and the scaled loss.
         """
         rewrite_program(self._train_program, self._amp_lists)
-        scaled_loss = loss * self._loss_scaling
+        self._scaled_loss = loss * self._loss_scaling
         self._params_grads = self._optimizer.backward(
-            scaled_loss, startup_program, parameter_list, no_grad_set,
+            self._scaled_loss, startup_program, parameter_list, no_grad_set,
             callbacks)
         update_role_var_grad(self._train_program, self._params_grads)
         scaled_params_grads = []
@@ -245,7 +253,7 @@ def decorate(optimizer,
 	              optimizer=optimizer, init_loss_scaling=8.0)
 	
             ops, param_grads = mp_optimizer.minimize(loss)
-            scaled_loss = mp_optimizer.get_loss_scaling()
+            scaled_loss = mp_optimizer.get_scaled_loss()
     """
     if amp_lists is None:
         amp_lists = AutoMixedPrecisionLists()

python/paddle/fluid/contrib/tests/test_image_classification_fp16.py

@@ -140,7 +140,8 @@ def train(net_type, use_cuda, save_dirname, is_local):
             use_dynamic_loss_scaling=True)
 
         mp_optimizer.minimize(avg_cost)
-        scaled_loss = mp_optimizer.get_loss_scaling()
+        loss_scaling = mp_optimizer.get_loss_scaling()
+        scaled_loss = mp_optimizer.get_scaled_loss()
 
     BATCH_SIZE = 128
     PASS_NUM = 1