add warmup

doc/doc_ch/config.md

@@ -63,8 +63,9 @@
 |         beta1           |    设置一阶矩估计的指数衰减率  |       0.9         |               \             |
 |         beta2           |    设置二阶矩估计的指数衰减率  |     0.999         |               \             |
 |         decay           |         是否使用decay       |    \              |               \             |
-|      function(decay)    |         设置decay方式       |   -    |       目前支持cosine_decay与piecewise_decay  |
+|      function(decay)    |         设置decay方式       |   -    |       目前支持cosine_decay, cosine_decay_warmup与piecewise_decay  |
-|      step_each_epoch    |      每个epoch包含多少次迭代, cosine_decay时有效   |         20       | 计算方式：total_image_num / (batch_size_per_card * card_size) |
+|      step_each_epoch    |      每个epoch包含多少次迭代, cosine_decay/cosine_decay_warmup时有效   |         20       | 计算方式：total_image_num / (batch_size_per_card * card_size) |
-|        total_epoch      |    总共迭代多少个epoch, cosine_decay时有效        |       1000      | 与Global.epoch_num 一致        |
+|        total_epoch      |    总共迭代多少个epoch, cosine_decay/cosine_decay_warmup时有效        |       1000      | 与Global.epoch_num 一致        |
+|        warmup_minibatch      |  线性warmup的迭代次数, cosine_decay_warmup时有效        |       1000      | \        |
 |        boundaries      |    学习率下降时的迭代次数间隔, piecewise_decay时有效       |       -      | 参数为列表形式        |
 |        decay_rate      |    学习率衰减系数, piecewise_decay时有效       |       -      |  \        |

doc/doc_en/config_en.md

@@ -60,8 +60,9 @@ Take `rec_icdar15_train.yml` as an example:
 |         beta1           |    Set the exponential decay rate for the 1st moment estimates  |       0.9         |               \             |
 |         beta2           |    Set the exponential decay rate for the 2nd moment estimates  |     0.999         |               \             |
 |         decay           |         Whether to use decay       |    \              |               \             |
-|      function(decay)    |         Set the decay function       |   cosine_decay    |         Support cosine_decay and piecewise_decay            |
+|      function(decay)    |         Set the decay function       |   cosine_decay    |         Support cosine_decay, cosine_decay_warmup and piecewise_decay            |
-|      step_each_epoch    |      The number of steps in an epoch. Used in cosine_decay  |         20       | Calculation ：total_image_num / (batch_size_per_card * card_size) |
+|      step_each_epoch    |      The number of steps in an epoch. Used in cosine_decay/cosine_decay_warmup  |         20       | Calculation: total_image_num / (batch_size_per_card * card_size) |
-|        total_epoch      |    The number of epochs. Used in cosine_decay      |       1000      | Consistent with Global.epoch_num      |
+|        total_epoch      |    The number of epochs. Used in cosine_decay/cosine_decay_warmup      |       1000      | Consistent with Global.epoch_num      |
+|        warmup_minibatch      |  Number of steps for linear warmup. Used in cosine_decay_warmup        |       1000      | \        |
 |        boundaries      |    The step intervals to reduce learning rate. Used in piecewise_decay       |       -      |  The format is list        |
 |        decay_rate      |    Learning rate decay rate. Used in piecewise_decay       |       -      |  \        |

ppocr/optimizer.py

@@ -14,14 +14,50 @@
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
+import math
 import paddle.fluid as fluid
 from paddle.fluid.regularizer import L2Decay
+from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter
+import paddle.fluid.layers.ops as ops
 
 from ppocr.utils.utility import initial_logger
 
 logger = initial_logger()
 
 
+def cosine_decay_with_warmup(learning_rate,
+                             step_each_epoch,
+                             epochs=500,
+                             warmup_minibatch=1000):
+    """Applies cosine decay to the learning rate.
+    lr = 0.05 * (math.cos(epoch * (math.pi / 120)) + 1)
+    decrease lr for every mini-batch and start with warmup.
+    """
+    global_step = _decay_step_counter()
+    lr = fluid.layers.tensor.create_global_var(
+        shape=[1],
+        value=0.0,
+        dtype='float32',
+        persistable=True,
+        name="learning_rate")
+
+    warmup_minibatch = fluid.layers.fill_constant(
+        shape=[1],
+        dtype='float32',
+        value=float(warmup_minibatch),
+        force_cpu=True)
+
+    with fluid.layers.control_flow.Switch() as switch:
+        with switch.case(global_step < warmup_minibatch):
+            decayed_lr = learning_rate * (1.0 * global_step / warmup_minibatch)
+            fluid.layers.tensor.assign(input=decayed_lr, output=lr)
+        with switch.default():
+            decayed_lr = learning_rate * \
+                (ops.cos((global_step - warmup_minibatch) * (math.pi / (epochs * step_each_epoch))) + 1)/2
+            fluid.layers.tensor.assign(input=decayed_lr, output=lr)
+    return lr
+
+
 def AdamDecay(params, parameter_list=None):
     """
     define optimizer function
@@ -36,7 +72,9 @@ def AdamDecay(params, parameter_list=None):
     l2_decay = params.get("l2_decay", 0.0)
 
     if 'decay' in params:
-        supported_decay_mode = ["cosine_decay", "piecewise_decay"]
+        supported_decay_mode = [
+            "cosine_decay", "cosine_decay_warmup", "piecewise_decay"
+        ]
         params = params['decay']
         decay_mode = params['function']
         assert decay_mode in supported_decay_mode, "Supported decay mode is {}, but got {}".format(
@@ -49,6 +87,15 @@ def AdamDecay(params, parameter_list=None):
                 learning_rate=base_lr,
                 step_each_epoch=step_each_epoch,
                 epochs=total_epoch)
+        elif decay_mode == "cosine_decay_warmup":
+            step_each_epoch = params['step_each_epoch']
+            total_epoch = params['total_epoch']
+            warmup_minibatch = params.get("warmup_minibatch", 1000)
+            base_lr = cosine_decay_with_warmup(
+                learning_rate=base_lr,
+                step_each_epoch=step_each_epoch,
+                epochs=total_epoch,
+                warmup_minibatch=warmup_minibatch)
         elif decay_mode == "piecewise_decay":
             boundaries = params["boundaries"]
             decay_rate = params["decay_rate"]
@@ -104,5 +151,5 @@ def RMSProp(params, parameter_list=None):
     optimizer = fluid.optimizer.RMSProp(
         learning_rate=base_lr,
         regularization=fluid.regularizer.L2Decay(regularization_coeff=l2_decay))
-        
+
-    return optimizer
+    return optimizer