Support int8 Deconv (#1347)

intel · Oct 14, 2022 · 5ca9b4c · 5ca9b4c
1 parent 81a0e45
commit 5ca9b4c
Show file tree

Hide file tree

Showing 19 changed files with 1,012 additions and 176 deletions.
diff --git a/neural_compressor/adaptor/inteltensorflow.yaml b/neural_compressor/adaptor/inteltensorflow.yaml
@@ -24,8 +24,10 @@
 
   ops: &common_ops
     int8: ['Conv2D', 'Conv3D', 'DepthwiseConv2dNative', 'FusedBatchNorm', 'FusedBatchNormV2','FusedBatchNormV3',
-           'MatMul', 'BatchMatMul', 'BatchMatMulV2', 'ConcatV2', 'MaxPool', 'MaxPool3D', 'AvgPool', '_MklFusedInstanceNorm']
-    uint8: ['Conv2D', 'Conv3D', 'DepthwiseConv2dNative', 'MatMul', 'BatchMatMul', 'BatchMatMulV2', 'ConcatV2', 'MaxPool', 'MaxPool3D', 'AvgPool']
+           'MatMul', 'BatchMatMul', 'BatchMatMulV2', 'ConcatV2', 'MaxPool', 'MaxPool3D', 'AvgPool', '_MklFusedInstanceNorm',
+           'Conv2DBackpropInput', 'Conv3DBackpropInputV2']
+    uint8: ['Conv2D', 'Conv3D', 'DepthwiseConv2dNative', 'MatMul', 'BatchMatMul', 'BatchMatMulV2', 'ConcatV2',
+            'MaxPool', 'MaxPool3D', 'AvgPool', 'Conv2DBackpropInput', 'Conv3DBackpropInputV2']
     bf16: ["Conv2D", "Conv2DBackpropFilter", "Conv2DBackpropInput", "Conv3D", "Conv3DBackpropFilterV2", "Conv3DBackpropInputV2",
            "DepthwiseConv2dNative", "DepthwiseConv2dNativeBackpropFilter", "DepthwiseConv2dNativeBackpropInput", "GRUBlockCell",
            "AUGRUBlockCell", "MklGRU", "MklAUGRU", "MatMul", "BatchMatMul", "BatchMatMulV2", "Einsum", # allow_list
@@ -296,7 +298,9 @@
         'Dequantize + DepthwiseConv2dNative + BiasAdd + QuantizeV2',
         'Dequantize + FusedBatchNormV3 + Relu + QuantizeV2',
         'Dequantize + _MklFusedInstanceNorm + Relu + QuantizeV2',
-        'Dequantize + _MklFusedInstanceNorm + LeakyRelu + QuantizeV2'
+        'Dequantize + _MklFusedInstanceNorm + LeakyRelu + QuantizeV2',
+        'Dequantize + Conv2DBackpropInput + BiasAdd + QuantizeV2',
+        'Dequantize + Conv3DBackpropInputV2 + BiasAdd + QuantizeV2'
         ]
     uint8: [
         'Dequantize + Conv2D + BiasAdd + AddN + Relu + QuantizeV2',
@@ -370,8 +374,9 @@
         'Dequantize + Conv3D + Add + Relu + QuantizeV2',
         'Dequantize + Conv3D + Add + Relu6 + QuantizeV2',
         'Dequantize + Conv3D + Add + Eelu + QuantizeV2',
-        'Dequantize + Conv3D + Add + LeakyRelu + QuantizeV2'
-
+        'Dequantize + Conv3D + Add + LeakyRelu + QuantizeV2',
+        'Dequantize + Conv2DBackpropInput + BiasAdd + QuantizeV2',
+        'Dequantize + Conv3DBackpropInputV2 + BiasAdd + QuantizeV2'
   ]
 
   grappler_optimization: &common_grappler_optimization

diff --git a/neural_compressor/adaptor/tensorflow.py b/neural_compressor/adaptor/tensorflow.py
@@ -47,7 +47,9 @@ class TensorFlowAdaptor(Adaptor):
         "MatMul": "matmul",
         "BatchMatMul": "matmul",
         "BatchMatMulV2": "matmul",
-        "Pad": "pad"
+        "Pad": "pad",
+        "Conv2DBackpropInput": "deconv2d",
+        "Conv3DBackpropInputV2": "deconv3d"
     }
     def __init__(self, framework_specific_info):
         super().__init__(framework_specific_info)
@@ -607,11 +609,12 @@ def _dump_model_op_stats(self, model_graphdef):
         fp32_op_list=list(set(fp32_op_list_uint8).union(set(fp32_op_list_int8)))
 
 
-        int8_op_prefix_list = ['QuantizedConv2D', '_QuantizedConv3D', 'QuantizedDepthwise',
+        int8_op_prefix_list = ['QuantizedConv2D', '_FusedQuantizedConv3D', 'QuantizedDepthwise',
                                'QuantizedMaxPool', 'QuantizedAvgPool',
                                'QuantizedConcatV2', 'QuantizedMatMul',
                                '_QuantizedFusedBatchNorm', '_QuantizedMatMul',
-                               '_QuantizedBatchMatMul', '_QuantizedFusedInstanceNorm']
+                               '_QuantizedBatchMatMul', '_QuantizedFusedInstanceNorm',
+                               '_FusedQuantizedDeconv2D', '_FusedQuantizedDeconv3D']
         from tensorflow.python.framework import dtypes
 
         res = {}
@@ -636,6 +639,10 @@ def _dump_model_op_stats(self, model_graphdef):
                     origin_op_type = 'DepthwiseConv2dNative'
                 if origin_op_type == 'BatchMatMul':
                     origin_op_type = 'BatchMatMulV2'
+                if origin_op_type == 'Deconv2D':
+                    origin_op_type = 'Conv2DBackpropInput'
+                if origin_op_type == 'Deconv3D':
+                    origin_op_type = 'Conv3DBackpropInputV2'
                 res[origin_op_type]['INT8'] += 1
 
             if i.op in fp32_op_list:

diff --git a/neural_compressor/adaptor/tf_utils/graph_rewriter/generic/insert_print_node.py b/neural_compressor/adaptor/tf_utils/graph_rewriter/generic/insert_print_node.py
@@ -46,10 +46,12 @@ def do_transformation(self):
         if top_node.op == 'ConcatV2':
             for i in range(top_node.attr['N'].i):
                 insert_node_pairs.append([top_node.input[i], self.post_node_name])
-        elif top_node.op == 'BatchMatMulV2':
+        elif top_node.op in ('BatchMatMul', 'BatchMatMulV2'):
             insert_node_pairs.append([top_node.input[0], self.post_node_name])
             if graph_info[top_node.input[1]].node.op != 'Const':
-                insert_node_pairs.append([top_node.input[1], self.post_node_name])  
+                insert_node_pairs.append([top_node.input[1], self.post_node_name])
+        elif top_node.op in ('Conv2DBackpropInput', 'Conv3DBackpropInputV2'):
+            insert_node_pairs.append([top_node.input[2], self.post_node_name])
         else:
             refresh_pre_node_name = graph_info[self.pre_node_name].node.input[0]
             # Check the Conv2D could be fused with previous Pad or not.

diff --git a/neural_compressor/adaptor/tf_utils/graph_rewriter/int8/fuse_conv_redundant_dequantize.py b/neural_compressor/adaptor/tf_utils/graph_rewriter/int8/fuse_conv_redundant_dequantize.py
@@ -23,12 +23,14 @@
 from neural_compressor.adaptor.tf_utils.graph_util import GraphRewriterHelper as Helper
 
 class FuseConvRedundantDequantizeTransformer(GraphRewriterBase):
-    """Fuse _QuantizedConv with the successor Dequantize Op.
-       _QuantizedConv2D only supports quantizedtype out_type currently
+    """Fuse _QuantizedConv/_QuantizedDeConv with the successor Dequantize Op.
     """
     fuse_patterns = [[
-        "_QuantizedConv3D",
-        "_QuantizedDepthwiseConv2D"
+        "_FusedQuantizedConv3D",
+        "_FusedQuantizedConv2D",
+        "_FusedQuantizedDepthwiseConv2D",
+        "_FusedQuantizedDeconv2D",
+        "_FusedQuantizedDeconv3D"
     ], ['Dequantize']]
 
     def __init__(self, model, device='cpu'):