Fix the other lr_decay

python/paddle/fluid/layers/learning_rate_scheduler.py

@@ -62,7 +62,7 @@ def noam_decay(d_model, warmup_steps):
         The decayed learning rate.
     """
     global_step = _decay_step_counter(1)
-    with init_on_cpu():
+
     a = global_step**-0.5
     b = (warmup_steps**-1.5) * global_step
     lr_value = (d_model**-0.5) * ops.elementwise_min(a, b)
@@ -108,8 +108,6 @@ def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     """
     global_step = _decay_step_counter()
 
-    with init_on_cpu():
-        # update learning_rate
     div_res = global_step / decay_steps
     if staircase:
         div_res = ops.floor(div_res)
@@ -138,7 +136,6 @@ def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     """
     global_step = _decay_step_counter()
 
-    with init_on_cpu():
     div_res = global_step / decay_steps
     if staircase:
         div_res = ops.floor(div_res)
@@ -184,7 +181,6 @@ def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     """
     global_step = _decay_step_counter()
 
-    with init_on_cpu():
     div_res = global_step / decay_steps
     if staircase:
         div_res = ops.floor(div_res)
@@ -224,13 +220,10 @@ def polynomial_decay(learning_rate,
     """
     global_step = _decay_step_counter()
 
-    with init_on_cpu():
     if cycle:
         div_res = ops.ceil(global_step / decay_steps)
-            zero_var = tensor.fill_constant(
-                shape=[1], dtype='float32', value=0.0)
-            one_var = tensor.fill_constant(
-                shape=[1], dtype='float32', value=1.0)
+        zero_var = tensor.fill_constant(shape=[1], dtype='float32', value=0.0)
+        one_var = tensor.fill_constant(shape=[1], dtype='float32', value=1.0)
 
         with control_flow.Switch() as switch:
             with switch.case(global_step == zero_var):