fix quantization for 1.8 (#300)

paddleslim/quant/__init__.py

@@ -20,11 +20,11 @@ from ..common import get_logger
 _logger = get_logger(__name__, level=logging.INFO)
 
 try:
-    fluid.require_version('2.0.0')
+    fluid.require_version('1.8.0')
     from .quanter import quant_aware, quant_post, convert, quant_post_only_weight
 except Exception as e:
     _logger.warning(
         "If you want to use training-aware and post-training quantization, "
-        "please use Paddle >= 2.0.0 or develop version")
+        "please use Paddle >= 1.8.0 or develop version")
 
 from .quant_embedding import quant_embedding

paddleslim/quant/quanter.py

@@ -24,8 +24,6 @@ from paddle.fluid.contrib.slim.quantization import ConvertToInt8Pass
 from paddle.fluid.contrib.slim.quantization import TransformForMobilePass
 from paddle.fluid.contrib.slim.quantization import PostTrainingQuantization
 from paddle.fluid.contrib.slim.quantization import AddQuantDequantPass
-from paddle.fluid.contrib.slim.quantization import OutScaleForTrainingPass
-from paddle.fluid.contrib.slim.quantization import OutScaleForInferencePass
 from paddle.fluid import core
 from paddle.fluid.contrib.slim.quantization import WeightQuantization
 
@@ -222,10 +220,6 @@ def quant_aware(program, place, config=None, scope=None, for_test=False):
             quantizable_op_type=quant_dequant_ops)
         quant_dequant_pass.apply(main_graph)
 
-    out_scale_training_pass = OutScaleForTrainingPass(
-        scope=scope, place=place, moving_rate=config['moving_rate'])
-    out_scale_training_pass.apply(main_graph)
-
     if for_test:
         quant_program = main_graph.to_program()
     else:
@@ -371,9 +365,6 @@ def convert(program, place, config=None, scope=None, save_int8=False):
     _logger.info("convert config {}".format(config))
     test_graph = IrGraph(core.Graph(program.desc), for_test=True)
 
-    out_scale_infer_pass = OutScaleForInferencePass(scope=scope)
-    out_scale_infer_pass.apply(test_graph)
-
     # Freeze the graph after training by adjusting the quantize
     # operators' order for the inference.
     freeze_pass = QuantizationFreezePass(