!10328 [MSLITE][Develop] Keep constant input tensor

From: @sunsuodong
Reviewed-by: 
Signed-off-by:

mindspore/lite/schema/ops.fbs

@@ -244,7 +244,7 @@ table Conv2DGradFilter {
     dilateW: int;
     dilateH: int;
     hasBias: bool = false;  // DEPRECATED
-    filter_shape: [int];
+    filter_shape: [int];  // DEPRECATED
     activationType: ActivationType = 0;
 }
 
@@ -265,7 +265,7 @@ table Conv2DGradInput {
     dilateW: int;
     dilateH: int;
     hasBias: bool = false;  // DEPRECATED
-    input_shape: [int];
+    input_shape: [int];  // DEPRECATED
     activationType: ActivationType = 0;
 }
 

mindspore/lite/src/ops/conv2d_grad_filter.cc

@@ -219,11 +219,11 @@ Registry conv2DGradFilterRegistry(schema::PrimitiveType_Conv2DGradFilter, Conv2D
 #endif
 
 int Conv2DGradFilter::InferShape(std::vector<Tensor *> inputs, std::vector<Tensor *> outputs) {
-  if (2 != inputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad Filter should have 2 inputs, but it got " << inputs.size();
+  if (inputs.size() < 2) {
+    MS_LOG(ERROR) << "Conv2d Grad Filter should be at least two input, but it got " << inputs.size();
     return RET_ERROR;
   }
-  if (1 != outputs.size()) {
+  if (outputs.size() != 1) {
     MS_LOG(ERROR) << "Conv2d Grad Filter should have one output but it got " << outputs.size();
     return RET_ERROR;
   }

mindspore/lite/src/ops/conv2d_grad_input.cc

@@ -220,11 +220,11 @@ Registry Conv2DGradInputRegistry(schema::PrimitiveType_Conv2DGradInput, Conv2DGr
 #endif
 
 int Conv2DGradInput::InferShape(std::vector<Tensor *> inputs, std::vector<Tensor *> outputs) {
-  if (2 != inputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad Input should have 2 inputs";
+  if (inputs.size() < 2) {
+    MS_LOG(ERROR) << "Conv2d Grad Input should be at least two input";
     return RET_ERROR;
   }
-  if (1 != outputs.size()) {
+  if (outputs.size() != 1) {
     MS_LOG(ERROR) << "Conv2d Grad output should have one output";
     return RET_ERROR;
   }

mindspore/lite/src/ops/gather.cc

@@ -98,8 +98,8 @@ Registry GatherRegistry(schema::PrimitiveType_Gather, GatherCreator);
 
 int Gather::InferShape(std::vector<Tensor *> inputs_, std::vector<Tensor *> outputs_) {
   MS_ASSERT(this->primitive_ != nullptr);
-  if (inputs_.size() != kDoubleNum) {
-    MS_LOG(DEBUG) << "Gather should have two inputs";
+  if (inputs_.size() < kDoubleNum) {
+    MS_LOG(DEBUG) << "Gather should be at least two inputs";
   }
   if (outputs_.size() != kSingleNum) {
     MS_LOG(ERROR) << "Gather should have one outputs";

mindspore/lite/src/ops/group_conv2d_grad_input.cc

@@ -146,8 +146,8 @@ Registry GroupConv2DGradInputRegistry(schema::PrimitiveType_GroupConv2DGradInput
 #endif
 
 int GroupConv2DGradInput::InferShape(std::vector<Tensor *> inputs, std::vector<Tensor *> outputs) {
-  if (2 != inputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad input should have 2 inputs";
+  if (inputs.size() < 2) {
+    MS_LOG(ERROR) << "Conv2d Grad input should be at least two input";
     return RET_ERROR;
   }
   if (1 != outputs.size()) {

mindspore/lite/test/models_mindspore.cfg

@@ -1,6 +1,6 @@
 ssd.mindir 1.5
 mobilenetv2_438.mindir 1.5
-gate_u_net_small-1_110.mindir 1.5
+#gate_u_net_small-1_110.mindir 1.5
 shufflenetv2.mindir 1.5
 #inceptionv3.mindir 1.5
 googlenet.mindir 1.5

mindspore/lite/tools/anf_exporter/anf_exporter.cc

@@ -571,8 +571,34 @@ int AnfExporter::ConvertInputValueNode(const std::shared_ptr<AnfNode> &input_ano
     output_cnode->inputIndex.emplace_back(meta_graphT->allTensors.size());
     meta_graphT->allTensors.emplace_back(std::move(paramTensor));
   } else if (value->isa<mindspore::ValueSequeue>()) {
-    MS_LOG(DEBUG) << "Value type is ValueSequence.";
-    return RET_OK;
+    auto valueAbstract = valueNode->abstract();
+    auto abstractSequnce = utils::cast<abstract::AbstractSequeuePtr>(valueAbstract);
+    if (abstractSequnce->isa<abstract::AbstractTuple>()) {
+      auto abstractTuple = utils::cast<abstract::AbstractTuplePtr>(valueAbstract);
+      auto x_shape_data = abstractTuple->elements();
+      std::vector<int32_t> shape;
+      for (std::size_t i = 0; i < abstractTuple->size(); ++i) {
+        auto value_track = x_shape_data[i]->GetValueTrack();
+        MS_ASSERT(value_track != nullptr);
+        if (value_track->isa<Int32Imm>()) {
+          shape.push_back((GetValue<int>(value_track)));
+        } else if (value_track->isa<Int64Imm>()) {
+          shape.push_back((GetValue<int64_t>(value_track)));
+        } else {
+          MS_LOG(ERROR) << "Value type is ValueSequence is not integer, it is " << value_track->ToString() << ".";
+          return RET_ERROR;
+        }
+      }
+      auto typePtr = abstractTuple->elements()[0]->GetTypeTrack();
+      paramTensor->dataType = kNumberTypeInt32;
+      paramTensor->dims = {static_cast<int32_t>(shape.size())};
+      paramTensor->nodeType = schema::NodeType_ValueNode;
+      paramTensor->data.resize(shape.size() * sizeof(int));
+      memcpy(paramTensor->data.data(), shape.data(), shape.size() * sizeof(int));
+      node_id_map_[valueNode->fullname_with_scope()] = meta_graphT->allTensors.size();
+      output_cnode->inputIndex.emplace_back(meta_graphT->allTensors.size());
+      meta_graphT->allTensors.emplace_back(std::move(paramTensor));
+    }
   } else if (value->isa<mindspore::BoolImm>()) {
     auto valueAbstract = valueNode->abstract();
     auto abstractScalar = utils::cast<abstract::AbstractScalarPtr>(valueAbstract);

mindspore/lite/tools/converter/legacy_optimizer/graph/infer_quant_param_pass.cc

@@ -36,8 +36,7 @@ STATUS InferQuantParamPass::Run(schema::MetaGraphT *graph) {
     if (node->quantType == schema::QuantType_AwareTraining) {
       continue;
     }
-    if (GetCNodeTType(*node) == schema::PrimitiveType_FakeQuantWithMinMax ||
-        GetCNodeTType(*node) == schema::PrimitiveType_FakeQuantWithMinMaxVars) {
+    if (GetCNodeTType(*node) == schema::PrimitiveType_FakeQuantWithMinMaxVars) {
       MS_ASSERT(false);
     }
     auto quantParamCalcer = quantParamRegister->GetQuantParamCalcer(GetCNodeTType(*node));