Restore quantization and distillation stategy before loading persistables. (#16959)

test=develop

Author: wanghaoshuang

python/paddle/fluid/contrib/slim/core/compressor.py

@@ -363,6 +363,9 @@ def _load_checkpoint(self, context):
                             strategies = pickle.load(
                                 strategy_file, encoding='bytes')
 
+                for strategy in strategies:
+                    strategy.restore_from_checkpoint(context)
+
                 if os.path.exists(model_path):
                     exe = SlimGraphExecutor(context.place)
                     with scope_guard(context.scope):

python/paddle/fluid/contrib/slim/core/strategy.py

@@ -46,3 +46,6 @@ def on_batch_end(self, context):
 
     def on_compression_end(self, context):
         pass
+
+    def restore_from_checkpoint(self, context):
+        pass

python/paddle/fluid/contrib/slim/distillation/distillation_strategy.py

@@ -38,7 +38,7 @@ def __init__(self, distillers=None, start_epoch=0, end_epoch=0):
         super(DistillationStrategy, self).__init__(start_epoch, end_epoch)
         self.distillers = distillers
 
-    def on_compression_begin(self, context):
+    def restore_from_checkpoint(self, context):
         # load from checkpoint
         if context.epoch_id > 0:
             if context.epoch_id > self.start_epoch and context.epoch_id < self.end_epoch:

python/paddle/fluid/contrib/slim/quantization/quantization_strategy.py

@@ -88,7 +88,7 @@ def __init__(self,
         self.save_out_nodes = save_out_nodes
         self.save_in_nodes = save_in_nodes
 
-    def on_compression_begin(self, context):
+    def restore_from_checkpoint(self, context):
         """
         Restore graph when the compressoin task is inited from checkpoint.
         """
@@ -143,10 +143,9 @@ def _modify_graph_for_quantization(self, context):
             train_ir_graph.graph).with_data_parallel(
                 loss_name=context.optimize_graph.out_nodes['loss'],
                 build_strategy=build_strategy)
-        # for evaluation. And program compiled from ir graph must be with data parallel.
-        context.eval_graph.compiled_graph = CompiledProgram(
-            test_ir_graph.graph).with_data_parallel(
-                build_strategy=build_strategy)
+
+        context.eval_graph.program = test_ir_graph.to_program()
+
         # for saving inference model after training
         context.put('quantization_test_ir_graph_backup', test_ir_graph)