Add lr scheduler for auto_compress (PaddlePaddle#1127)

* Add lr scheduler. * Add lr scheduler. * add notes for lr scheduler. * add notes for lr scheduler.
lizexu123 · May 24, 2022 · 009b0a9 · 009b0a9
1 parent 78332fc
commit 009b0a9
Show file tree

Hide file tree

Showing 4 changed files with 29 additions and 7 deletions.
diff --git a/demo/auto_compression/hyperparameter_tutorial.md b/demo/auto_compression/hyperparameter_tutorial.md
@@ -116,7 +116,13 @@ TrainConfig:
   optim_args:
     weight_decay: 0.0005
 ```
-
+- 学习率衰减策略：主要设置策略类名和策略参数，如下所示。目前在paddle中已经实现了多种衰减策略，请参考[lr文档](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.2/api/paddle/optimizer/lr/LRScheduler_cn.html)，策略参数即类初始化参数。
+```yaml
+  learning_rate:
+    type: PiecewiseDecay # 学习率衰减策略类名
+    boundaries: [4500] # 设置策略参数
+    values: [0.005, 0.0005] # 设置策略参数
+```
 ## 其他参数配置
 
 #### 1.自动蒸馏效果不理想，怎么自主选择蒸馏节点？

diff --git a/paddleslim/auto_compression/compressor.py b/paddleslim/auto_compression/compressor.py
@@ -494,7 +494,8 @@ def _start_train(self, train_program_info, test_program_info, strategy):
                 np_probs_float, = self._exe.run(train_program_info.program, \
                     feed=data, \
                     fetch_list=train_program_info.fetch_targets)
-
+                if not isinstance(train_program_info.learning_rate, float):
+                    train_program_info.learning_rate.step()
                 if 'unstructure' in strategy:
                     self._pruner.step()
 

diff --git a/paddleslim/auto_compression/create_compressed_program.py b/paddleslim/auto_compression/create_compressed_program.py
@@ -30,6 +30,18 @@
 ]
 
 
+def _create_lr_scheduler(train_config):
+    if 'learning_rate' not in train_config:
+        raise RuntimeError(
+            'No `learning_rate` specified in the configuration file.')
+    if isinstance(train_config.get('learning_rate'), float):
+        return train_config.get('learning_rate')
+
+    params = train_config.get('learning_rate')
+    lr_type = params.pop('type')
+    return getattr(optimizer.lr, lr_type)(**params)
+
+
 def _create_optimizer(train_config):
     """create optimizer"""
     opt = getattr(optimizer, train_config.get('optimizer') or
@@ -54,10 +66,11 @@ def _create_optimizer(train_config):
         train_config['optim_args'] = {}
         grad_clip = None
 
-    op = opt(learning_rate=train_config["learning_rate"],
+    lr = _create_lr_scheduler(train_config)
+    op = opt(learning_rate=lr,
              grad_clip=grad_clip,
              **train_config['optim_args'])
-    return op
+    return op, lr
 
 
 def _parse_distill_loss(distill_node_pair,
@@ -223,7 +236,7 @@ def build_distill_program(executor,
     train_fetch_list = []
     with paddle.static.program_guard(train_program, startup_program):
         with paddle.utils.unique_name.guard('merge'):
-            optimizer = _create_optimizer(train_config)
+            optimizer, learning_rate = _create_optimizer(train_config)
 
             if train_config.get('use_fleet'):
                 optimizer = fleet.distributed_optimizer(optimizer,
@@ -277,7 +290,7 @@ def build_distill_program(executor,
 
     train_program_info = ProgramInfo(startup_program, train_program,
                                      feed_target_names, train_fetch_list,
-                                     optimizer)
+                                     optimizer, learning_rate)
     test_program_info = ProgramInfo(startup_program, test_program,
                                     feed_target_names, fetch_targets)
     return train_program_info, test_program_info

diff --git a/paddleslim/auto_compression/strategy_config.py b/paddleslim/auto_compression/strategy_config.py
@@ -128,9 +128,11 @@ def __init__(self,
                  program,
                  feed_target_names,
                  fetch_targets,
-                 optimizer=None):
+                 optimizer=None,
+                 learning_rate=None):
         self.startup_program = startup_program
         self.program = program
         self.feed_target_names = feed_target_names
         self.fetch_targets = fetch_targets
         self.optimizer = optimizer
+        self.learning_rate = learning_rate