feat(tflite): support bias use int64 dtype

Author: dingshaohua960303

bin/convert

@@ -94,7 +94,7 @@ complete_str = """_mgeconvert(){
                     return
                     ;;
                 tracedmodule_to_tflite)
-                    words="-i --input -o --output --input_data_type --input_scales --input_zero_points --require_quantize --param_fake_quant --quantize_file_path --graph_name --mtk --end_point  --outspec --remove_relu --prefer_same_pad_mode"
+                    words="-i --input -o --output --input_data_type --input_scales --input_zero_points --require_quantize --param_fake_quant --quantize_file_path --graph_name --mtk --end_point  --outspec --remove_relu --prefer_same_pad_mode --use_int64_bias"
                     COMPREPLY=( $(compgen -W "$words" -- $word) )
                     return
                     ;;
@@ -397,6 +397,7 @@ def init(subparsers):
                     outspec=args.outspec,
                     remove_relu=args.remove_relu,
                     prefer_same_pad_mode=args.prefer_same_pad_mode,
+                    use_int64_bias=args.use_int64_bias,
                 )
             else:
                 mgeconvert.mge_to_tflite(
@@ -406,6 +407,7 @@ def init(subparsers):
                     mtk=args.mtk,
                     outspec=args.outspec,
                     prefer_same_pad_mode=args.prefer_same_pad_mode,
+                    use_int64_bias=args.use_int64_bias,
                 )
 
         def tflite_parser(subparsers):
@@ -488,6 +490,12 @@ def init(subparsers):
                 help="whether prefer to use SAME pad mode for conv op",
             )
 
+            p.add_argument(
+                "--use_int64_bias",
+                action="store_true",
+                help="whether use int64 as dtype of bias",
+            )
+
         tflite_parser(subparsers)
 
 

mgeconvert/backend/ir_to_tflite/tflite_op.py

@@ -233,6 +233,7 @@ def get_shape_param(
     "int8": TensorType.INT8,
     "int16": TensorType.INT16,
     "int32": TensorType.INT32,
+    "int64": TensorType.INT64,
     "qint8_narrow": TensorType.INT8,
 }
 

mgeconvert/converter_ir/ir_transform.py

@@ -48,6 +48,7 @@
     TanHOpr,
     TransposeOpr,
     TrueDivOpr,
+    _ConvOpr,
     _PoolOpr,
 )
 from .ir_tensor import AxisOrder, IRTensor
@@ -121,6 +122,7 @@ class TransformerRule(Enum):
     TRANSPOSE_LINEAR_WEIGHT_TO_NHWC = 133
     # force fc with no trans for megengine
     FC_NO_TRANS = 134
+    BIAS_ASTYPE_INT64 = 135
 
 
 def cmp_rules(a, b):
@@ -1537,3 +1539,29 @@ def trans_tensor(tensor):
             if opr.transpose_b and tensor_b.owner_opr == None:
                 opr.transpose_b = False
                 trans_tensor(tensor_b)
+
+
+@_register_tranformation_rule(TransformerRule.BIAS_ASTYPE_INT64)
+def _bias_astype_int64(net: IRGraph):
+    for opr in net.all_oprs:
+        if not isinstance(opr, (MatMulOpr, _ConvOpr)):
+            continue
+        bias = None
+        if isinstance(opr, MatMulOpr) and len(opr.inp_tensors) == 3:
+            bias = opr.inp_tensors[2]
+        elif isinstance(opr, Deconv2dOpr) and len(opr.inp_tensors) > 3:
+            if (
+                opr.inp_tensors[0].shape == [4] and len(opr.inp_tensors) == 4
+            ):  # shape as input
+                bias = opr.inp_tensors[-1]
+            if len(opr.inp_tensors) == 4:
+                bias = opr.inp_tensors[-1]
+        elif isinstance(opr, (Conv2dOpr, ConvRelu2dOpr)) and len(opr.inp_tensors) == 3:
+            bias = opr.inp_tensors[-1]
+        if bias is not None and bias.scale is not None:
+            bias.set_qparams(
+                scale=bias.scale,
+                zero_point=bias.zero_point,
+                q_dtype="int64",
+                np_dtype="int64",
+            )

mgeconvert/converters/mge_to_tflite.py

@@ -24,6 +24,7 @@ def mge_to_tflite(
     disable_nhwc=False,
     outspec=None,
     prefer_same_pad_mode=False,
+    use_int64_bias=False,
 ):
     """
     Convert megengine model to TFLite,
@@ -70,6 +71,9 @@ def mge_to_tflite(
         transformer_options.append(TransformerRule.DECONV_ADD_ZERO_BIAS,)
         transformer_options.append(TransformerRule.FUSE_FOR_DECONV_BIAS,)
 
+    if use_int64_bias:
+        transformer_options.append(TransformerRule.BIAS_ASTYPE_INT64)
+
     transformer = IRTransform(transformer_options)
     transformed_irgraph = transformer.transform(irgraph)
 

mgeconvert/converters/tm_to_tflite.py

@@ -39,6 +39,7 @@ def tracedmodule_to_tflite(
     remove_relu=False,
     prefer_same_pad_mode=False,
     disable_nhwc=False,
+    use_int64_bias=False,
 ):
     """
 	Convert traced model to TFLite,
@@ -89,6 +90,8 @@ def tracedmodule_to_tflite(
         transformer_options.append(TransformerRule.DECONV_ADD_ZERO_BIAS,)
     if remove_relu:
         transformer_options.append(TransformerRule.REMOVE_TFLITE_RELU,)
+    if use_int64_bias:
+        transformer_options.append(TransformerRule.BIAS_ASTYPE_INT64)
 
     transformer = IRTransform(transformer_options)
     transformed_irgraph = transformer.transform(irgraph)