Merge pull request #12254 from JiayiFeng/fix_lr_decay

Fix learning rate scheduler performance issue

python/paddle/fluid/layers/learning_rate_scheduler.py

@@ -62,10 +62,10 @@ 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)
+
+    a = global_step**-0.5
+    b = (warmup_steps**-1.5) * global_step
+    lr_value = (d_model**-0.5) * ops.elementwise_min(a, b)
 
     return lr_value
 
@@ -108,12 +108,10 @@ 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)
-        decayed_lr = learning_rate * (decay_rate**div_res)
+    div_res = global_step / decay_steps
+    if staircase:
+        div_res = ops.floor(div_res)
+    decayed_lr = learning_rate * (decay_rate**div_res)
 
     return decayed_lr
 
@@ -138,11 +136,10 @@ 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)
-        decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res)
+    div_res = global_step / decay_steps
+    if staircase:
+        div_res = ops.floor(div_res)
+    decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res)
 
     return decayed_lr
 
@@ -184,12 +181,11 @@ 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)
+    div_res = global_step / decay_steps
+    if staircase:
+        div_res = ops.floor(div_res)
 
-        decayed_lr = learning_rate / (1 + decay_rate * div_res)
+    decayed_lr = learning_rate / (1 + decay_rate * div_res)
 
     return decayed_lr
 
@@ -224,25 +220,22 @@ 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)
-
-            with control_flow.Switch() as switch:
-                with switch.case(global_step == zero_var):
-                    tensor.assign(input=one_var, output=div_res)
-            decay_steps = decay_steps * div_res
-        else:
-            decay_steps_var = tensor.fill_constant(
-                shape=[1], dtype='float32', value=float(decay_steps))
-            global_step = ops.elementwise_min(x=global_step, y=decay_steps_var)
+    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)
 
-        decayed_lr = (learning_rate - end_learning_rate) * \
-                     ((1 - global_step / decay_steps) ** power) + end_learning_rate
+        with control_flow.Switch() as switch:
+            with switch.case(global_step == zero_var):
+                tensor.assign(input=one_var, output=div_res)
+        decay_steps = decay_steps * div_res
+    else:
+        decay_steps_var = tensor.fill_constant(
+            shape=[1], dtype='float32', value=float(decay_steps))
+        global_step = ops.elementwise_min(x=global_step, y=decay_steps_var)
+
+    decayed_lr = (learning_rate - end_learning_rate) * \
+        ((1 - global_step / decay_steps) ** power) + end_learning_rate
     return decayed_lr
 
 
@@ -277,28 +270,28 @@ def piecewise_decay(boundaries, values):
 
     global_step = _decay_step_counter()
 
-    with init_on_cpu():
-        lr = tensor.create_global_var(
-            shape=[1],
-            value=0.0,
-            dtype='float32',
-            persistable=True,
-            name="learning_rate")
+    lr = tensor.create_global_var(
+        shape=[1],
+        value=0.0,
+        dtype='float32',
+        persistable=True,
+        name="learning_rate")
 
-        with control_flow.Switch() as switch:
-            for i in range(len(boundaries)):
-                boundary_val = tensor.fill_constant(
-                    shape=[1], dtype='float32', value=float(boundaries[i]))
-                value_var = tensor.fill_constant(
-                    shape=[1], dtype='float32', value=float(values[i]))
-                with switch.case(global_step < boundary_val):
-                    tensor.assign(value_var, lr)
-            last_value_var = tensor.fill_constant(
+    with control_flow.Switch() as switch:
+        for i in range(len(boundaries)):
+            boundary_val = tensor.fill_constant(
                 shape=[1],
                 dtype='float32',
-                value=float(values[len(values) - 1]))
-            with switch.default():
-                tensor.assign(last_value_var, lr)
+                value=float(boundaries[i]),
+                force_cpu=True)
+            value_var = tensor.fill_constant(
+                shape=[1], dtype='float32', value=float(values[i]))
+            with switch.case(global_step < boundary_val):
+                tensor.assign(value_var, lr)
+        last_value_var = tensor.fill_constant(
+            shape=[1], dtype='float32', value=float(values[len(values) - 1]))
+        with switch.default():
+            tensor.assign(last_value_var, lr)
 
     return lr
 
@@ -333,9 +326,9 @@ def append_LARS(params_grads, learning_rate, weight_decay):
         grad_norm = ops.sqrt(nn.reduce_sum(input=ops.square(grad)))
         if type(param_lr) == float and param_lr == 1.0:
             decayed_lr = learning_rate * param_norm \
-                         / _balanced_weight(param_norm, grad_norm)
+                / _balanced_weight(param_norm, grad_norm)
         else:
             decayed_lr = learning_rate * param_lr * param_norm \
-                         / _balanced_weight(param_norm, grad_norm)
+                / _balanced_weight(param_norm, grad_norm)
         # set back param local learning rate
         param.optimize_attr['learning_rate'] = decayed_lr

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

@@ -91,20 +91,21 @@ class TestLearningRateDecay(unittest.TestCase):
 
     def check_decay_with_place(self, place, python_decay_fn, fluid_decay_fn,
                                kwargs):
+        main_prog = fluid.Program()
+        startup_prog = fluid.Program()
 
-        decayed_lr = fluid_decay_fn(**kwargs)
+        with fluid.program_guard(main_prog, startup_prog):
+            decayed_lr = fluid_decay_fn(**kwargs)
 
         place = fluid.CPUPlace()
         exe = fluid.Executor(place)
 
-        exe.run(fluid.default_startup_program())
+        exe.run(startup_prog)
 
-        fluid.memory_optimize(fluid.default_main_program())
+        fluid.memory_optimize(main_prog)
 
         for step in range(10):
-            lr_val, = exe.run(fluid.default_main_program(),
-                              feed={},
-                              fetch_list=[decayed_lr])
+            lr_val, = exe.run(main_prog, feed={}, fetch_list=[decayed_lr])
             python_decayed_lr = python_decay_fn(
                 global_step=float(step), **kwargs)
             self.assertAlmostEqual(