Add INT8 calibration support in Paddle package (#15569)

* Add INT8 calibration support in Paddle package; test=develop

Author: hshen14

paddle/fluid/API.spec

@@ -361,6 +361,9 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b
 paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None))
 paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.contrib.Calibrator.__init__ ArgSpec(args=['self'], varargs='args', keywords='kwargs', defaults=None)
+paddle.fluid.contrib.Calibrator.sample_data ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.contrib.Calibrator.save_int8_model ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.contrib.reader.ctr_reader.ctr_reader ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
 paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))

python/paddle/fluid/contrib/__init__.py

@@ -22,6 +22,8 @@
 from .op_frequence import *
 from . import quantize
 from .quantize import *
+from . import int8_inference
+from .int8_inference import *
 from . import reader
 from .reader import *
 from . import slim
@@ -34,6 +36,7 @@
 __all__ += memory_usage_calc.__all__
 __all__ += op_frequence.__all__
 __all__ += quantize.__all__
+__all__ += int8_inference.__all__
 __all__ += reader.__all__
 __all__ += slim.__all__
 __all__ += utils.__all__

python/paddle/fluid/contrib/int8_inference/__init__.py

@@ -11,3 +11,10 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
+from __future__ import print_function
+
+from . import utility
+from .utility import *
+
+__all__ = utility.__all__

python/paddle/fluid/contrib/int8_inference/utility.py

@@ -11,11 +11,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import paddle.fluid.core as core
+
+from paddle.fluid import core
 import numpy as np
 import math
 import os
-import paddle.fluid as fluid
+from paddle.fluid.executor import global_scope
+from paddle.fluid import io
+
+__all__ = ['Calibrator']
 
 
 class Calibrator(object):
@@ -76,8 +80,7 @@ def sample_data(self):
         '''
         for i in self.sampling_program.list_vars():
             if i.name in self.sampling_vars:
-                np_data = np.array(fluid.global_scope().find_var(i.name)
-                                   .get_tensor())
+                np_data = np.array(global_scope().find_var(i.name).get_tensor())
                 if i.name not in self._sampling_data:
                     self._sampling_data[i.name] = []
                 self._sampling_data[i.name].append(np_data)
@@ -86,9 +89,9 @@ def __save_offline_model(self):
         '''
         Save the quantized model to the disk.
         '''
-        fluid.io.save_inference_model(self.output, self.feed_var_names,
-                                      self.fetch_list, self.exe,
-                                      self.sampling_program)
+        io.save_inference_model(self.output, self.feed_var_names,
+                                self.fetch_list, self.exe,
+                                self.sampling_program)
 
     def __display_debug(self):
         if self.debug:

python/paddle/fluid/contrib/tests/test_calibration.py

@@ -24,8 +24,7 @@
 from paddle.dataset.common import download
 from PIL import Image, ImageEnhance
 import math
-sys.path.append('..')
-import int8_inference.utility as int8_utility
+import paddle.fluid.contrib.int8_inference.utility as int8_utility
 
 random.seed(0)
 np.random.seed(0)

python/setup.py.in

@@ -109,6 +109,7 @@ packages=['paddle',
           'paddle.fluid.contrib',
           'paddle.fluid.contrib.decoder',
           'paddle.fluid.contrib.quantize',
+          'paddle.fluid.contrib.int8_inference',
           'paddle.fluid.contrib.reader',
           'paddle.fluid.contrib.slim',
           'paddle.fluid.contrib.slim.core',