Merge branch 'gm_r1.1' into 'master'

Merge r1.1 code into master branch

See merge request intelai/tools!42

Author: chuanqi129

api/README.md

@@ -46,7 +46,7 @@ of specific models in Model Zoo as examples.
 
   ```bash
   $ cd ~
-  $ git clone https://github.com/IntelAI/tools.git  quantization && cd quantization
+  $ git clone https://github.com/IntelAI/tools.git quantization && cd quantization
   ```
 
 ## Step-by-step Procedure for ResNet-50 Quantization

api/examples/quantize_cmd.py

@@ -41,12 +41,12 @@ def main(_):
         outputs = []
 
     if args.excluded_ops:
-        excluded_ops = args.exclude_ops.split(',')
+        excluded_ops = args.excluded_ops.split(',')
     else:
         excluded_ops = []
 
     if args.excluded_nodes:
-        excluded_nodes = args.exclude_nodes.split(',')
+        excluded_nodes = args.excluded_nodes.split(',')
     else:
         excluded_nodes = []
 
@@ -61,7 +61,7 @@ def main(_):
         callback_cmd = prefix + 'input_graph={} ' + postfix
     else:
         callback_cmd = args.callback
-    qt.gen_calib_data_cmds = args.callback
+    qt.gen_calib_data_cmds = callback_cmd
     qt.convert()
 
 

api/intel_quantization/__init__.py

@@ -15,3 +15,4 @@
 #  See the License for the specific language governing permissions and
 #  limitations under the License.
 #
+__version__="1.1"

api/intel_quantization/graph_converter.py

@@ -25,21 +25,22 @@
 # from tensorflow.python.tools.optimize_for_inference_lib import optimize_for_inference
 
 # from intel_quantization.quantize_graph import GraphRewriter
-from intel_quantization.transform_graph.strip_unused import StripUnusedNodes
-from intel_quantization.transform_graph.fold_batch_norm import FoldBatchNormNodes
-from intel_quantization.transform_graph.insert_logging import InsertLogging
-from intel_quantization.transform_graph.freeze_max_min import freeze_max
-from intel_quantization.transform_graph.freeze_max_min import freeze_min
-from intel_quantization.transform_graph.freeze_max_min import freeze_requantization_range
-from intel_quantization.transform_graph.fuse_quantized_conv_and_requantize import fuse_quantized_conv_and_requantize
-from intel_quantization.transform_graph.fuse_column_wise_mul import FuseColumnWiseMul
-from intel_quantization.transform_graph.rerange_quantized_concat import RerangeQuantizedConcat
-from intel_quantization.util import read_graph, write_graph
-from intel_quantization.quantize_graph.quantize_graph_for_intel_cpu import QuantizeGraphForIntel
+from .transform_graph.strip_unused import StripUnusedNodes
+from .transform_graph.fold_batch_norm import FoldBatchNormNodes
+from .transform_graph.insert_logging import InsertLogging
+from .transform_graph.freeze_max_min import freeze_max
+from .transform_graph.freeze_max_min import freeze_min
+from .transform_graph.freeze_max_min import freeze_requantization_range
+from .transform_graph.freeze_max_min import get_all_fp32_data, get_tensor_histogram, combine_histogram
+from .transform_graph.fuse_quantized_conv_and_requantize import fuse_quantized_conv_and_requantize
+from .transform_graph.fuse_column_wise_mul import FuseColumnWiseMul
+from .transform_graph.rerange_quantized_concat import RerangeQuantizedConcat
+from .util import read_graph, write_graph
+from .quantize_graph.quantize_graph_for_intel_cpu import QuantizeGraphForIntel
+from .quantize_graph.quantize_graph_common import QuantizeGraphHelper
 import os
 import shlex
 import subprocess
-import sys
 import logging
 
 logging.getLogger().setLevel(level=logging.INFO)
@@ -52,7 +53,7 @@
 
 class GraphConverter:
     def __init__(self, input_graph, output_graph, inputs=[], outputs=[], excluded_ops=[], excluded_nodes=[],
-                 per_channel=False, input_graph_is_binary=True):
+                 per_channel=False, input_graph_is_binary=True, algo='DIRECT'):
         """Convert graph.
 
         :param input_graph: input graph pb file.
@@ -73,13 +74,18 @@ def __init__(self, input_graph, output_graph, inputs=[], outputs=[], excluded_op
         self.per_channel = per_channel
         self.excluded_ops = excluded_ops
         self.excluded_nodes = excluded_nodes
+        self.algo = algo
         self._low_precision_mode = 'eightbit'
-
+        self._calibration_data = []
+        self._fp32_print_data = []
         self.gen_calib_data_cmds = None
         self.debug = False
         self._check_tf_version()
         self._check_args()
         self._gen_tmp_filenames()
+        self._kl_op_dict = {}
+        self._kl_keys = []
+        self._print_node_mapping = {}
 
     def _check_tf_version(self):
         is_supported_version = False
@@ -113,7 +119,7 @@ def _gen_tmp_filenames(self):
         self._fp32_optimized_graph = os.path.join(self._output_path, 'fp32_optimized_graph.pb')
         self._int8_dynamic_range_graph = os.path.join(self._output_path, 'int8_dynamic_range_graph.pb')
         self._int8_logged_graph = os.path.join(self._output_path, 'int8_logged_graph.pb')
-        self._requant_min_max_log = os.path.join(self._output_path, 'requant_min_max_log.txt')
+        self._fp32_logged_graph = os.path.join(self._output_path, 'fp32_logged_graph.pb')
         self._int8_frozen_range_graph = os.path.join(self._output_path, 'int8_frozen_range_graph.pb')
         if not self.output_graph:
             self.output_graph = os.path.join(self._output_path, 'int8_final_fused_graph.pb')
@@ -137,6 +143,58 @@ def convert(self):
         else:
             self.quantize()
 
+    def _get_fp32_print_node_names(self):
+        offset_map = {
+            "QuantizedConv2DWithBiasSumAndRelu": 3,
+            "QuantizedConv2DWithBiasAndRelu": 2,
+            "QuantizedConv2DWithBias": 1,
+        }
+        target_conv_op = []
+        sorted_graph = QuantizeGraphHelper().get_sorted_graph(
+            self._fp32_origin_graph, self.outputs)
+
+        node_name_mapping = {
+            node.name: node
+            for node in self._tmp_graph_def.node if node.op != "Const"
+        }
+
+        for node in self._tmp_graph_def.node:
+            if node.op in offset_map:
+                target_conv_op.append(node.name.split('_eightbit_')[0])
+        fp32_node_name_mapping = {
+            node.name: node
+            for node in sorted_graph.node if node.op != "Const"
+        }
+        sorted_node_names = [i.name for i in sorted_graph.node if i.op != "Const"]
+
+        output_node_names = []
+        for i in target_conv_op:
+            if node_name_mapping[
+                    i + "_eightbit_quantized_conv"].op == 'QuantizedConv2DWithBiasSumAndRelu':
+                start_index = sorted_node_names.index(i)
+                for index, value in enumerate(sorted_node_names[start_index:]):
+                    if fp32_node_name_mapping[value].op.startswith(
+                            "Add") and fp32_node_name_mapping[
+                                sorted_node_names[start_index + index + 1]].op == "Relu":
+                        output_node_names.append(
+                            sorted_node_names[start_index + index + 1])
+                        self._print_node_mapping[sorted_node_names[start_index + index + 1]] = i
+            elif i in sorted_node_names:
+                start_index = sorted_node_names.index(i)
+                end_index = start_index + offset_map[node_name_mapping[
+                    i + "_eightbit_quantized_conv"].op]
+                output_node_names.append(sorted_node_names[end_index])
+                self._print_node_mapping[sorted_node_names[end_index]] = i
+
+        for i in output_node_names:
+            self._kl_keys.append(';' + i + '__print__;__KL')
+
+        InsertLogging(self._fp32_origin_graph,
+                      node_name_list=output_node_names,
+                      message="__KL:",
+                      summarize=-1, dump_fp32=True).do_transformation()
+        write_graph(self._fp32_origin_graph, self._fp32_logged_graph)
+
     def quantize(self):
         """Quantize graph only (without optimizing fp32 graph), including:
             1) quantize graph,
@@ -150,9 +208,14 @@ def quantize(self):
                              'to generate calibration data.')
         try:
             self._quantize_graph()
+            if self.algo == "KL":
+                self._get_fp32_print_node_names()
+                self._generate_calibration_data(self._fp32_logged_graph,
+                                                self._fp32_print_data, True)
+
             self._insert_logging()
-            self._generate_calibration_data()
-            self._freeze_requantization_ranges()
+            self._generate_calibration_data(self._int8_logged_graph, self._calibration_data)
+            self._freeze_requantization_ranges(self._kl_op_dict)
             self._fuse_requantize_with_fused_quantized_conv()
         except Exception as e:
             logging.error('Failed to quantize graph due to: %s', str(e))
@@ -172,6 +235,7 @@ def _optimize_frozen_fp32_graph(self):
         self._tmp_graph_def = graph_util.remove_training_nodes(self._tmp_graph_def, self.outputs)
         self._tmp_graph_def = FoldBatchNormNodes(self._tmp_graph_def).do_transform()
         write_graph(self._tmp_graph_def, self._fp32_optimized_graph)
+        self._fp32_origin_graph = self._tmp_graph_def
 
     def _quantize_graph(self):
         """quantize graph."""
@@ -199,32 +263,50 @@ def _insert_logging(self):
                       ops=["RequantizationRange{}".format("PerChannel" if self.per_channel else "")],
                       message="__requant_min_max:").do_transformation()
         InsertLogging(self._tmp_graph_def, ops=["Min"], message="__min:").do_transformation()
-        InsertLogging(self._tmp_graph_def, ops=["Max"], message="__max:").do_transformation()
+        InsertLogging(self._tmp_graph_def, ops=["Max"],
+                      message="__max:").do_transformation()
+        # InsertLogging(
+        #     self._tmp_graph_def,
+        #     ops=["QuantizedConv2DWithBiasAndRelu",
+        #     "QuantizedConv2DWithBias"
+        #     ],
+        #     message="__KL:",
+        #     summarize=-1).do_transformation()
+
         write_graph(self._tmp_graph_def, self._int8_logged_graph)
         self._tmp_graph_def.CopyFrom(int8_dynamic_range_graph_def)
 
-    def _generate_calibration_data(self):
+    def _generate_calibration_data(self, graph, output, enable_kl_algo=False):
         cmd = self.gen_calib_data_cmds
-        cmd = cmd.format(self._int8_logged_graph)
-        f = open(self._requant_min_max_log, 'w', buffering=1)
-        p = subprocess.Popen(shlex.split(cmd), stderr=subprocess.STDOUT, stdout=subprocess.PIPE)
-        try:
-            for line in p.stdout:
-                line_str = line.decode(sys.stdout.encoding)
-                sys.stdout.write(line_str)
-                f.write(line_str)
-            p.communicate()
-        except Exception:
-            p.kill()
-            p.wait()
-            raise
-        if p.poll():
-            raise SystemExit('ERROR generating calibration data, command: \n{}'.format(cmd))
-
-    def _freeze_requantization_ranges(self):
-        self._tmp_graph_def = freeze_max(self._tmp_graph_def, self._requant_min_max_log)
-        self._tmp_graph_def = freeze_min(self._tmp_graph_def, self._requant_min_max_log)
-        self._tmp_graph_def = freeze_requantization_range(self._tmp_graph_def, self._requant_min_max_log)
+        cmd = cmd.format(graph)
+        p = subprocess.Popen(shlex.split(cmd),
+                             stderr=subprocess.STDOUT,
+                             stdout=subprocess.PIPE)
+        while p.poll() is None:
+            line = p.stdout.readline().strip().decode()
+            if line and line.startswith(';'):
+                if not enable_kl_algo:
+                    output.append(line)
+
+                if enable_kl_algo and line.rsplit(':')[0] in self._kl_keys:
+                    fp32_data = get_all_fp32_data(line.rsplit(':')[-1])
+                    key = self._print_node_mapping[line[1:].split('__print')[0]] + '_eightbit_requant_range'
+                    if key not in self._kl_op_dict:
+                        self._kl_op_dict[key] = get_tensor_histogram(fp32_data)
+                    else:
+                        self._kl_op_dict[key] = combine_histogram(self._kl_op_dict[key], fp32_data)
+
+    def _freeze_requantization_ranges(self, additional_data=None):
+        use_moving_average = self.algo == "MA"
+        self._tmp_graph_def = freeze_max(self._tmp_graph_def,
+                                         self._calibration_data,
+                                         use_moving_average)
+        self._tmp_graph_def = freeze_min(self._tmp_graph_def,
+                                         self._calibration_data,
+                                         use_moving_average)
+        self._tmp_graph_def = freeze_requantization_range(
+            self._tmp_graph_def, self._calibration_data, use_moving_average,
+            additional_data)
         if self.debug:
             write_graph(self._tmp_graph_def, self._int8_frozen_range_graph)
 
@@ -256,5 +338,3 @@ def _post_clean(self):
         """
         if gfile.Exists(self._int8_logged_graph):
             os.remove(self._int8_logged_graph)
-        if gfile.Exists(self._requant_min_max_log):
-            os.remove(self._requant_min_max_log)

api/intel_quantization/quantize_graph/quantize_graph_concatv2.py

@@ -1,8 +1,8 @@
 #  -*- coding: utf-8 -*-
 from tensorflow.python.framework import dtypes
 from tensorflow.core.framework import node_def_pb2
-from intel_quantization.quantize_graph.quantize_graph_base import QuantizeNodeBase
-from intel_quantization.quantize_graph.quantize_graph_common import QuantizeGraphHelper as helper
+from .quantize_graph_base import QuantizeNodeBase
+from .quantize_graph_common import QuantizeGraphHelper as helper
 
 import re
 
@@ -53,9 +53,15 @@ def _apply_concatv2_transform(self, original_node):
             self._add_dequantize_result_node(quantized_concat_name,
                                              original_node.name)
 
+    def _quantizable_concat(self, node):
+        for input_node_name in node.input[:node.attr['N'].i]:
+            if self.node_name_mapping[helper.node_name_from_input(input_node_name)].node.op != "Dequantize":
+                return False
+        return True
+
     def _apply_concatv2_quantization(self):
         for _, v in self.node_name_mapping.items():
-            if v.node.op in ("ConcatV2") and not re.search(
+            if v.node.op in ("ConcatV2") and self._quantizable_concat(v.node) and not re.search(
                     r'map(_\d+)?/while', v.node.name) and dtypes.as_dtype(
                         v.node.attr["T"].type) == dtypes.float32:
                 self._apply_concatv2_transform(v.node)

api/intel_quantization/quantize_graph/quantize_graph_conv.py

@@ -3,8 +3,8 @@
 from tensorflow.core.framework import node_def_pb2
 from tensorflow.python.framework import dtypes
 
-from intel_quantization.quantize_graph.quantize_graph_common import QuantizeGraphHelper as helper
-from intel_quantization.quantize_graph.quantize_graph_base import QuantizeNodeBase
+from .quantize_graph_common import QuantizeGraphHelper as helper
+from .quantize_graph_base import QuantizeNodeBase
 
 import logging
 
@@ -232,7 +232,7 @@ def apply_conv_biasadd_fusion(self, match_node_name):
                 helper.set_attr_dtype(quantized_conv_node, "out_type",
                                       dtypes.qint32)
                 self.add_output_graph_node(quantized_conv_node)
-                requantize_type = dtypes.qint8 if self.per_channel else dtypes.quint8
+                requantize_type = dtypes.qint8
 
                 quantize_down_name = self._add_quantize_down_nodes(
                     node, quantized_node_name, requantize_type, False)

api/intel_quantization/quantize_graph/quantize_graph_for_intel_cpu.py

@@ -3,13 +3,13 @@
 from tensorflow.python.platform import gfile
 from tensorflow.python.framework import graph_util
 
-from intel_quantization.quantize_graph.quantize_graph_base import QuantizeGraphBase
-from intel_quantization.quantize_graph.quantize_graph_common import QuantizeGraphHelper
-from intel_quantization.quantize_graph.quantize_graph_conv import FuseNodeStartWithConv2d
-from intel_quantization.quantize_graph.quantize_graph_concatv2 import FuseNodeStartWithConcatV2
-from intel_quantization.quantize_graph.quantize_graph_matmul import FuseNodeStartWithMatmul
-from intel_quantization.quantize_graph.quantize_graph_pooling import FuseNodeStartWithPooling
-from intel_quantization.quantize_graph.quantize_graph_pad import FuseNodeStartWithPad
+from .quantize_graph_base import QuantizeGraphBase
+from .quantize_graph_common import QuantizeGraphHelper
+from .quantize_graph_conv import FuseNodeStartWithConv2d
+from .quantize_graph_concatv2 import FuseNodeStartWithConcatV2
+from .quantize_graph_matmul import FuseNodeStartWithMatmul
+from .quantize_graph_pooling import FuseNodeStartWithPooling
+from .quantize_graph_pad import FuseNodeStartWithPad
 
 
 class QuantizeGraphForIntel(QuantizeGraphBase):

api/intel_quantization/quantize_graph/quantize_graph_matmul.py

@@ -3,8 +3,8 @@
 from tensorflow.core.framework import node_def_pb2
 from tensorflow.python.framework import dtypes
 
-from intel_quantization.quantize_graph.quantize_graph_common import QuantizeGraphHelper as helper
-from intel_quantization.quantize_graph.quantize_graph_base import QuantizeNodeBase
+from .quantize_graph_common import QuantizeGraphHelper as helper
+from .quantize_graph_base import QuantizeNodeBase
 
 import logging
 

api/intel_quantization/quantize_graph/quantize_graph_pad.py

@@ -3,8 +3,8 @@
 from tensorflow.core.framework import node_def_pb2
 from tensorflow.python.framework import tensor_util
 
-from intel_quantization.quantize_graph.quantize_graph_base import QuantizeNodeBase
-from intel_quantization.quantize_graph.quantize_graph_common import QuantizeGraphHelper as helper
+from .quantize_graph_base import QuantizeNodeBase
+from .quantize_graph_common import QuantizeGraphHelper as helper
 
 
 class FuseNodeStartWithPad(QuantizeNodeBase):

api/intel_quantization/quantize_graph/quantize_graph_pooling.py

@@ -2,8 +2,8 @@
 from tensorflow.core.framework import node_def_pb2
 from tensorflow.python.framework import dtypes
 
-from intel_quantization.quantize_graph.quantize_graph_base import QuantizeNodeBase
-from intel_quantization.quantize_graph.quantize_graph_common import QuantizeGraphHelper as helper
+from .quantize_graph_base import QuantizeNodeBase
+from .quantize_graph_common import QuantizeGraphHelper as helper
 
 
 class FuseNodeStartWithPooling(QuantizeNodeBase):
@@ -24,7 +24,8 @@ def _add_pool_function(self, original_node, quantized_op_node):
 
     def _apply_pool_quantization(self):
         for _, v in self.node_name_mapping.items():
-            if v.node.op in ("AvgPool", "MaxPool"):
+            if v.node.op in ("AvgPool", "MaxPool") and self._find_relu_node(
+                    v.node):
                 self.eightbitize_single_input_tensor_node(
                     v.node, self._add_pool_function)
             else: