!3250 [quant]bug fixes when export geir

Merge pull request !3250 from vlne-v1/quant_export_geir_biasadd_after_depth_wise

mindspore/ccsrc/pipeline/jit/pipeline.cc

@@ -395,7 +395,7 @@ void ExecutorPy::GetGeBackendPolicy() const {
 
 bool IsPhaseExportGeir(const std::string &phase_s) {
   auto phase_to_export = "export.geir";
-  return phase_s.rfind(phase_to_export, 0) != std::string::npos;
+  return phase_s.rfind(phase_to_export) != std::string::npos;
 }
 
 std::vector<ActionItem> GetPipline(const ResourcePtr &resource, const std::string &phase_s, bool use_vm) {

mindspore/ccsrc/transform/graph_ir/op_declare.cc

@@ -757,7 +757,7 @@ ATTR_MAP(ExtractImagePatches) = {{"ksizes", ATTR_DESC(ksizes, AnyTraits<int>(),
 OUTPUT_MAP(ExtractImagePatches) = {{0, OUTPUT_DESC(y)}};
 
 // Conv2D
-INPUT_MAP(Conv2D) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(filter)}};
+INPUT_MAP(Conv2D) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(filter)}, {3, INPUT_DESC(bias)}};
 ATTR_MAP(Conv2D) = {
   {"stride", ATTR_DESC(strides, AnyTraits<std::vector<int64_t>>(), AnyTraits<std::vector<int64_t>>())},
   {"pad_list", ATTR_DESC(pads, AnyTraits<std::vector<int64_t>>(), AnyTraits<std::vector<int64_t>>())},
@@ -794,7 +794,7 @@ ATTR_MAP(Conv2DBackpropFilterD) = {
 OUTPUT_MAP(Conv2DBackpropFilterD) = {{0, OUTPUT_DESC(y)}};
 
 // DepthwiseConv2D
-INPUT_MAP(DepthwiseConv2D) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(filter)}};
+INPUT_MAP(DepthwiseConv2D) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(filter)}, {3, INPUT_DESC(bias)}};
 ATTR_MAP(DepthwiseConv2D) = {
   {"stride", ATTR_DESC(strides, AnyTraits<std::vector<int64_t>>(), AnyTraits<std::vector<int64_t>>())},
   {"pads", ATTR_DESC(pads, AnyTraits<std::vector<int64_t>>(), AnyTraits<std::vector<int64_t>>())},
@@ -826,7 +826,7 @@ ATTR_MAP(DepthwiseConv2DBackpropFilterD) = {
 OUTPUT_MAP(DepthwiseConv2DBackpropFilterD) = {{0, OUTPUT_DESC(filter_grad)}};
 
 // MatMulV2
-INPUT_MAP(MatMulV2) = {{1, INPUT_DESC(x1)}, {2, INPUT_DESC(x2)}};
+INPUT_MAP(MatMulV2) = {{1, INPUT_DESC(x1)}, {2, INPUT_DESC(x2)}, {3, INPUT_DESC(bias)}};
 ATTR_MAP(MatMulV2) = {{"transpose_a", ATTR_DESC(transpose_x1, AnyTraits<bool>())},
                       {"transpose_b", ATTR_DESC(transpose_x2, AnyTraits<bool>())}};
 OUTPUT_MAP(MatMulV2) = {{0, OUTPUT_DESC(y)}};
@@ -1347,7 +1347,8 @@ OUTPUT_MAP(AscendQuant) = {{0, OUTPUT_DESC(y)}};
 // AscendDequant
 INPUT_MAP(AscendDequant) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(deq_scale)}};
 ATTR_MAP(AscendDequant) = {{"sqrt_mode", ATTR_DESC(sqrt_mode, AnyTraits<bool>())},
-                           {"relu_flag", ATTR_DESC(relu_flag, AnyTraits<bool>())}};
+                           {"relu_flag", ATTR_DESC(relu_flag, AnyTraits<bool>())},
+                           {"dtype", ATTR_DESC(dtype, AnyTraits<GEType>())}};
 OUTPUT_MAP(AscendDequant) = {{0, OUTPUT_DESC(y)}};
 #ifdef ENABLE_GE
 // Print

mindspore/nn/layer/quant.py

@@ -28,8 +28,8 @@ from mindspore._checkparam import check_int_positive, check_bool, twice
 from mindspore._checkparam import Rel
 import mindspore.context as context
 
-from .normalization import BatchNorm2d
-from .activation import get_activation
+from .normalization import BatchNorm2d, BatchNorm1d
+from .activation import get_activation, ReLU
 from ..cell import Cell
 from . import conv, basic
 from ..._checkparam import ParamValidator as validator
@@ -206,7 +206,7 @@ class DenseBnAct(Cell):
         self.has_bn = validator.check_bool("has_bn", has_bn)
         self.has_act = activation is not None
         if has_bn:
-            self.batchnorm = BatchNorm2d(out_channels)
+            self.batchnorm = BatchNorm1d(out_channels)
         self.activation = get_activation(activation)
 
     def construct(self, x):
@@ -1156,13 +1156,18 @@ class QuantBlock(Cell):
         self.has_bias = bias is not None
         self.activation = activation
         self.has_act = activation is not None
+        if isinstance(activation, ReLU):
+            self.activation = None
+            self.has_act = False
+            self.dequant.add_prim_attr("relu_flag", True)
         self.bias_add = P.BiasAdd()
 
     def construct(self, x):
         x = self.quant(x)
-        x = self.core_op(x, self.weight)
         if self.has_bias:
-            x = self.bias_add(x, self.bias)
+            x = self.core_op(x, self.weight, self.bias)
+        else:
+            x = self.core_op(x, self.weight)
         if self.has_act:
             x = self.activation(x)
         x = self.dequant(x, self.dequant_scale)

mindspore/ops/operations/_inner_ops.py

@@ -380,6 +380,7 @@ class Dequant(PrimitiveWithInfer):
     def __init__(self, sqrt_mode=False, relu_flag=False):
         self.sqrt_mode = validator.check_value_type("sqrt_mode", sqrt_mode, [bool], self.name)
         self.relu_flag = validator.check_value_type("relu_flag", relu_flag, [bool], self.name)
+        self.add_prim_attr("dtype", mstype.float16)
 
     def infer_shape(self, x_shape, deq_scale_shape):
         return x_shape

mindspore/ops/operations/math_ops.py

@@ -596,7 +596,7 @@ class MatMul(PrimitiveWithInfer):
             raise ValueError('MatMul input x, y should be the same dimension size and should be '
                              + f'equal to 2, while x size = {len(x)}, y size= {len(y)}')
 
-    def infer_shape(self, x, y):
+    def infer_shape(self, x, y, bias=None):
         self.check_shape_size(x, y)
         cls_name = self.name
         # expected dimension of x, y, x:[...,a,b] y:[..., c,d], the dim size should be the same except the last two
@@ -621,7 +621,7 @@ class MatMul(PrimitiveWithInfer):
         ret_dims = x[: -2] + [x_last[self.transpose_a], y_last[not self.transpose_b]]
         return ret_dims
 
-    def infer_dtype(self, x, y):
+    def infer_dtype(self, x, y, bias=None):
         args = {"x": x, "y": y}
         validator.check_tensor_type_same(args, mstype.float_type + mstype.int_type, self.name)
         if x.element_type() == mstype.int8:

mindspore/ops/operations/nn_ops.py

@@ -842,7 +842,7 @@ class Conv2D(PrimitiveWithInfer):
         self.group = validator.check_integer('group', group, 0, Rel.GT, self.name)
         self.add_prim_attr('offset_a', 0)
 
-    def infer_shape(self, x_shape, w_shape):
+    def infer_shape(self, x_shape, w_shape, b_shape=None):
         validator.check_integer("weight rank", len(w_shape), 4, Rel.EQ, self.name)
         validator.check_integer("x rank", len(x_shape), 4, Rel.EQ, self.name)
         validator.check(f"x_shape[1] / group", x_shape[1] // self.group, "w_shape[1]", w_shape[1], Rel.EQ, self.name)
@@ -887,7 +887,7 @@ class Conv2D(PrimitiveWithInfer):
         out_shape = [x_shape[0], out_channel, h_out, w_out]
         return out_shape
 
-    def infer_dtype(self, x_dtype, w_dtype):
+    def infer_dtype(self, x_dtype, w_dtype, b_dtype=None):
         args = {'x': x_dtype, 'w': w_dtype}
         valid_types = [mstype.int8, mstype.int32, mstype.float16, mstype.float32]
         validator.check_tensor_type_same(args, valid_types, self.name)
@@ -968,7 +968,7 @@ class DepthwiseConv2dNative(PrimitiveWithInfer):
         self.group = validator.check_integer("group", group, 0, Rel.GT, self.name)
         self.add_prim_attr('offset_a', 0)
 
-    def infer_shape(self, x_shape, w_shape):
+    def infer_shape(self, x_shape, w_shape, b_shape=None):
         validator.check_integer("weight rank", len(w_shape), 4, Rel.EQ, self.name)
         validator.check_integer("x rank", len(x_shape), 4, Rel.EQ, self.name)
         validator.check("x_shape[1]", x_shape[1], "w_shape[1]", w_shape[1], Rel.EQ, self.name)
@@ -1011,7 +1011,7 @@ class DepthwiseConv2dNative(PrimitiveWithInfer):
         out_shape = [x_shape[0], out_channel, h_out, w_out]
         return out_shape
 
-    def infer_dtype(self, x_dtype, w_dtype):
+    def infer_dtype(self, x_dtype, w_dtype, b_dtype=None):
         args = {'x': x_dtype, 'w': w_dtype}
         validator.check_tensor_type_same(args, mstype.number_type, self.name)
         if x_dtype.element_type() == mstype.int8:

tests/ut/python/train/quant/test_quant.py

@@ -78,7 +78,7 @@ def test_qat_lenet():
 def test_qat_mobile_per_channel_tf():
     network = mobilenetV2(num_classes=1000)
     img = Tensor(np.ones((1, 3, 224, 224)).astype(np.float32))
-    network = qat.convert_quant_network(network, bn_fold=True, per_channel=[False, True], symmetric=[True, False])
+    network = qat.convert_quant_network(network, bn_fold=True, per_channel=[True, False], symmetric=[True, False])
     # should load the checkpoint. mock here
     for param in network.get_parameters():
         param.init_data()