xref: /foundation/ai/mindspore/patches/0002-train-and-build.patch

From 9b1f17ffcb18cd58f796c9f30c0bc37e31347450 Mon Sep 17 00:00:00 2001
From: zhangyanhui <zhangyanhui17@huawei.com>
Date: Fri, 21 Jun 2024 16:08:21 +0800
Subject: [PATCH] auto-apply 0002-train-and-build.patch

---
 .../plugin/device/cpu/kernel/nnacl/BUILD.gn   |  16 +-
 mindspore/lite/BUILD.gn                       | 229 +++++++++++++++++-
 .../include/registry/opencl_runtime_wrapper.h |   2 +-
 mindspore/lite/src/litert/kernel/cpu/BUILD.gn |   1 +
 .../litert/kernel/cpu/fp32/oneslike_fp32.cc   |  51 ++++
 .../litert/kernel/cpu/fp32/oneslike_fp32.h    |  46 ++++
 mindspore/lite/src/litert/lite_model.h        |   2 +-
 mindspore/lite/src/litert/lite_session.cc     |   1 +
 mindspore/lite/src/litert/weight_decoder.h    |   2 +-
 mindspore/lite/src/tensor.h                   |   2 +-
 mindspore/lite/src/tensorlist.h               |   2 +-
 mindspore/lite/src/train/train_session.cc     |   6 +-
 .../lite/tools/benchmark_train/net_train.h    |   2 +-
 .../tools/converter/converter_metagraph.cc    |  14 +-
 .../tools/converter/graphdef_transform.cc     |   4 +
 .../legacy_optimizer/graph/CMakeLists.txt     |   1 +
 .../legacy_optimizer/graph/node_name_pass.cc  |  96 ++++++++
 .../legacy_optimizer/graph/node_name_pass.h   |  35 +++
 18 files changed, 493 insertions(+), 19 deletions(-)
 create mode 100644 mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.cc
 create mode 100644 mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.h
 create mode 100644 mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.cc
 create mode 100644 mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.h

diff --git a/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/BUILD.gn b/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/BUILD.gn
index a1e7908e..7bbc3782 100644
--- a/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/BUILD.gn
+++ b/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/BUILD.gn
@@ -714,7 +714,17 @@ arm32_assembly_sources = [
   "assembly/arm32/WinogradTransRight.S",
 ]
 
-fp16_assembly_sources = [
+arm32_fp16_assembly_sources = [
+  "assembly/arm82_aarch32_fp16/Float16Tofloat32.S",
+  "assembly/arm82_aarch32_fp16/Float32ToFloat16.S",
+  "assembly/arm82_aarch32_fp16/Matmul12x8Fp16.S",
+  "assembly/arm82_aarch32_fp16/MatVecMulFp16.S",
+  "assembly/arm82_aarch32_fp16/TiledC4MatmulFp16.S",
+  "assembly/arm82_aarch32_fp16/WinogradTransLeft.S",
+  "assembly/arm82_aarch32_fp16/WinogradTransRight.S",
+]
+
+arm64_fp16_assembly_sources = [
   "assembly/fp16/CalculateMinMaxFp16Count8.S",
   "assembly/fp16/ConvDwFp16Border.S",
   "assembly/fp16/ConvDwFp16Center.S",
@@ -839,11 +849,13 @@ nnacl_sources += infer_control_sources
 
 # source files on arm32
 arm_only_sources = arm32_assembly_sources
+#arm_only_sources += arm32_fp16_assembly_sources
+not_needed(arm32_fp16_assembly_sources)
 
 # source files on arm64
 arm64_only_sources = fp16_kernel_sources
 arm64_only_sources += fp16_grad_sources
-arm64_only_sources += fp16_assembly_sources
+arm64_only_sources += arm64_fp16_assembly_sources
 arm64_only_sources += arm64_assembly_sources
 arm64_only_sources += optimizing_assembly_sources
 arm64_only_sources += arm64_fp32_kernel_sources
diff --git a/mindspore/lite/BUILD.gn b/mindspore/lite/BUILD.gn
index 6d83e6f9..8a492bf9 100644
--- a/mindspore/lite/BUILD.gn
+++ b/mindspore/lite/BUILD.gn
@@ -74,10 +74,12 @@ import("//build/ohos.gni")
 ohos_group("mindspore") {
   deps = [
     ":mindspore_lib",
+    ":mindspore_train_lib",
     "mindir:mindir_lib",
   ]
 }
 
+# Inference library
 cxx_api_sources = [
   "src/litert/cxx_api/cell.cc",
   "src/litert/cxx_api/context.cc",
@@ -429,7 +431,6 @@ ohos_shared_library("mindspore_lib") {
   SUPPORT_NNRT = true
   if (SUPPORT_NNRT) {
     sources += [
-      # "mindir/src/mindir_nnrt_lite_graph.cc",
       "src/litert/delegate/nnrt/checker/primitive_check.cc",
       "src/litert/delegate/nnrt/nnrt_delegate.cc",
       "src/litert/delegate/nnrt/nnrt_model_kernel.cc",
@@ -444,8 +445,9 @@ ohos_shared_library("mindspore_lib") {
     external_deps += [ "neural_network_runtime:nnrt_target" ]
     deps += [ "mindir:mindir_lib" ]
     defines += [ "SUPPORT_NNRT" ]
-    defines += [ "MSLITE_ENABLE_EXPERIMENTAL_KERNEL" ]
   }
+  defines += [ "MSLITE_ENABLE_EXPERIMENTAL_KERNEL" ]
+  defines += [ "SUPPORT_TRAIN" ]
   cflags_cc = [
     "-Wno-ignored-qualifiers",
     "-Wunused-private-field",
@@ -458,6 +460,225 @@ ohos_shared_library("mindspore_lib") {
   subsystem_name = "thirdparty"
 }
 
+# Train library
+expression_cxx_api_sources = [
+  "src/litert/cxx_api/expression/net.cc",
+  "src/litert/cxx_api/expression/net_impl.cc",
+  "src/litert/cxx_api/expression/node_impl.cc",
+]
+
+expression_op_sources = [
+  "src/expression/ops/activation.cc",
+  "src/expression/ops/adam.cc",
+  "src/expression/ops/addn.cc",
+  "src/expression/ops/arithmetic.cc",
+  "src/expression/ops/arithmetic_self.cc",
+  "src/expression/ops/assign.cc",
+  "src/expression/ops/batchnorm.cc",
+  "src/expression/ops/biasadd.cc",
+  "src/expression/ops/conv.cc",
+  "src/expression/ops/dense.cc",
+  "src/expression/ops/depend.cc",
+  "src/expression/ops/dropout.cc",
+  "src/expression/ops/flatten.cc",
+  "src/expression/ops/pooling.cc",
+  "src/expression/ops/reduce.cc",
+  "src/expression/ops/reshape.cc",
+  "src/expression/ops/softmax.cc",
+  "src/expression/ops/softmaxCE.cc",
+  "src/expression/ops/tile.cc",
+  "src/expression/ops/transpose.cc",
+]
+
+all_expression_sources = [
+  "src/expression/export.cc",
+  "src/expression/expr.cc",
+  "src/expression/import.cc",
+  "src/expression/net.cc",
+  "src/expression/node.cc",
+  "src/expression/ops.cc",
+  "src/expression/ops_utils.cc",
+  "src/expression/param.cc",
+  "src/expression/sequential.cc",
+]
+
+all_expression_sources += expression_cxx_api_sources
+all_expression_sources += expression_op_sources
+
+all_train_sources = [
+  # ${API_TRAIN_SRC} is empty.
+  # ${TRAIN_SRC_WITH_MD} is empty.
+  "src/common/quant_utils.cc",
+  "src/litert/cxx_api/metrics/accuracy.cc",
+  "src/litert/cxx_api/train/model_build.cc",
+  "src/litert/cxx_api/train/model_build_impl.cc",
+  "src/litert/cxx_api/train/converters.cc",
+  "src/litert/cxx_api/train/train_support.cc",
+  "src/train/train_populate_parameter.cc",
+  "src/train/train_session.cc",
+  "src/train/graph_fusion.cc",
+  "src/train/graph_dropout.cc",
+  "src/train/transfer_session.cc",
+  "src/train/train_utils.cc",
+  "src/train/loss_monitor.cc",
+  "src/train/lr_scheduler.cc",
+  "src/train/accuracy_metrics.cc",
+#  "src/train/accuracy_monitor.cc", # depends on minddata header, not compiled
+  "src/train/classification_train_accuracy_monitor.cc",
+  "src/train/train_export.cc",
+  "src/train/opt_allocator.cc",
+  "src/train/optimizer/common/fusion_utils.cc",
+  "src/train/optimizer/fusion/matmul_activation_fusion_pass.cc",
+  "src/train/optimizer/fusion/reshape_gather_reshape_fusion_pass.cc",
+  "src/train/optimizer/fusion/gru_fusion_pass.cc",
+  "src/common/storage.cc",
+  "tools/converter/optimizer.cc",
+  "tools/converter/legacy_optimizer/fusion/fusion_pass.cc",
+  "tools/converter/legacy_optimizer/fusion/fusion_pattern.cc",
+  "tools/common/meta_graph_utils.cc",
+  "tools/common/statistic_utils.cc",
+  "tools/converter/legacy_optimizer/fusion/matmul_biasadd_fusion_pass.cc",
+  "tools/converter/legacy_optimizer/graph/dropout_node_remove_pass.cc",
+  "tools/converter/legacy_optimizer/graph/isolated_node_remove_pass.cc",
+  "tools/converter/legacy_optimizer/graph/subgraph_node_pass.cc",
+]
+
+all_train_sources += all_expression_sources
+
+fp16_train_kernel_sources = [
+  "src/litert/kernel/cpu/fp16_grad/activation_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/arithmetic_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/arithmetic_fp16_self_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/bias_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/bn_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/convolution_fp16_grad_filter.cc",
+  "src/litert/kernel/cpu/fp16_grad/convolution_fp16_grad_input.cc",
+  "src/litert/kernel/cpu/fp16_grad/dropout_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/layernorm_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/neg_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/pooling_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/resize_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/strided_slice_fp16_grad.cc",
+  "src/litert/kernel/cpu/fp16_grad/unsorted_segment_sum_fp16.cc",
+]
+
+fp32_train_kernel_sources = [
+  "src/litert/kernel/cpu/fp32_grad/activation_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/adam.cc",
+  "src/litert/kernel/cpu/fp32_grad/adam_weight_decay.cc",
+  "src/litert/kernel/cpu/fp32_grad/apply_momentum.cc",
+  "src/litert/kernel/cpu/fp32_grad/arithmetic_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/arithmetic_self_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/assign.cc",
+  "src/litert/kernel/cpu/fp32_grad/bias_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/bn_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/convolution.cc",
+  "src/litert/kernel/cpu/fp32_grad/convolution_grad_filter.cc",
+  "src/litert/kernel/cpu/fp32_grad/convolution_grad_input.cc",
+  "src/litert/kernel/cpu/fp32_grad/deconvolution_grad_filter.cc",
+  "src/litert/kernel/cpu/fp32_grad/dropout.cc",
+  "src/litert/kernel/cpu/fp32_grad/dropout_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/layernorm_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/lstm_grad_data_fp32.cc",
+  "src/litert/kernel/cpu/fp32_grad/lstm_grad_fp32.cc",
+  "src/litert/kernel/cpu/fp32_grad/lstm_grad_weight_fp32.cc",
+  "src/litert/kernel/cpu/fp32_grad/neg_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/nllloss_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/pooling_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/power_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/resize_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/sgd.cc",
+  "src/litert/kernel/cpu/fp32_grad/sigmoid_cross_entropy_with_logits.cc",
+  "src/litert/kernel/cpu/fp32_grad/sigmoid_cross_entropy_with_logits_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/smooth_l1_loss.cc",
+  "src/litert/kernel/cpu/fp32_grad/smooth_l1_loss_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/softmax_cross_entropy_with_logits.cc",
+  "src/litert/kernel/cpu/fp32_grad/softmax_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/sparse_softmax_cross_entropy_with_logits.cc",
+  "src/litert/kernel/cpu/fp32_grad/strided_slice_grad.cc",
+  "src/litert/kernel/cpu/fp32_grad/unsorted_segment_sum.cc",
+  "src/litert/kernel/cpu/fp32_grad/binary_cross_entropy.cc",
+  "src/litert/kernel/cpu/fp32_grad/binary_cross_entropy_grad.cc",
+]
+
+#all_train_sources += fp16_train_kernel_sources
+not_needed(fp16_train_kernel_sources)
+all_train_sources += fp32_train_kernel_sources
+
+ohos_shared_library("mindspore_train_lib") {
+  deps = [
+    ":mindspore_lib",
+  ]
+
+  sources = all_train_sources
+
+  include_dirs = [
+    "//base/hiviewdfx/hilog/interfaces/native/innerkits/include",
+    "//third_party/flatbuffers/include",
+    "./",
+    "../",
+    "../../",
+    "../core",
+    "src",
+    "src/c_api/",
+    "../ccsrc/plugin/device/cpu/kernel/",
+    "../core/mindrt/src/",
+    "../core/mindrt/include/",
+    "../../third_party/",
+    "./schema/",
+    "../ccsrc/",
+  ]
+
+  defines = [
+    "ENABLE_MINDRT",
+    "MS_COMPILE_OHOS",
+    "PRIMITIVE_WRITEABLE",
+    "VERSION_STR=\"2.1.0\"",
+  ]
+
+  if (target_cpu == "arm") {
+    defines += [
+      "ENABLE_ARM",
+      "ENABLE_ARM32",
+      "ENABLE_NEON",
+    ]
+  } else if (target_cpu == "arm64") {
+    defines += [
+      "ENABLE_ARM",
+      "ENABLE_ARM64",
+      "ENABLE_NEON",
+      "ENABLE_FP16",
+      "USE_OPENCL_WRAPPER",
+      "MS_OPENCL_PROFILE=false",
+      "CL_TARGET_OPENCL_VERSION=200",
+      "CL_HPP_TARGET_OPENCL_VERSION=120",
+      "CL_HPP_MINIMUM_OPENCL_VERSION=120",
+    ]
+  }
+  configs = [
+    ":mindspore_api",
+    ":disable_android",
+    ":train_kernel_option",
+  ]
+
+  remove_configs = [ "//build/config/compiler:no_rtti" ]
+  external_deps = [ "hilog:libhilog" ]
+  innerapi_tags = [ "platformsdk" ]
+  output_name = "libmindspore-lite-train"
+  output_extension = "so"
+  defines += [ "SUPPORT_TRAIN" ]
+  cflags_cc = [
+    "-Wno-ignored-qualifiers",
+    "-Wunused-private-field",
+    "-Wno-unused-private-field",
+    "-Wno-inconsistent-missing-override",
+    "-Wno-macro-redefined",
+    "-Wno-constant-conversion",
+  ]
+  part_name = "mindspore"
+}
+
+# Build configurations
 config("opencl_option") {
   cflags_cc = [ "-Wno-missing-braces" ]
 }
@@ -482,3 +703,7 @@ config("disable_android") {
 config("secure_option") {
   cflags = [ "-fstack-protector-all" ]
 }
+
+config("train_kernel_option") {
+  cflags_cc = [ "-fno-finite-math-only" ]
+}
diff --git a/mindspore/lite/include/registry/opencl_runtime_wrapper.h b/mindspore/lite/include/registry/opencl_runtime_wrapper.h
index fb647d37..b55554e4 100644
--- a/mindspore/lite/include/registry/opencl_runtime_wrapper.h
+++ b/mindspore/lite/include/registry/opencl_runtime_wrapper.h
@@ -1,5 +1,5 @@
 /**
- * Copyright 2021 Huawei Technologies Co., Ltd
+ * Copyright 2021-2023 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
diff --git a/mindspore/lite/src/litert/kernel/cpu/BUILD.gn b/mindspore/lite/src/litert/kernel/cpu/BUILD.gn
index b34e0427..48308425 100644
--- a/mindspore/lite/src/litert/kernel/cpu/BUILD.gn
+++ b/mindspore/lite/src/litert/kernel/cpu/BUILD.gn
@@ -112,6 +112,7 @@ cpu_kernel_sources = [
     "fp32/uniform_real_fp32.cc",
     "fp32/unstack_fp32.cc",
     "fp32/where_fp32.cc",
+    "fp32/oneslike_fp32.cc",
     "fp32/online_fusion/cast_gather_reduce_fp32.cc",
     "fp32/online_fusion/reduce_concat_fp32.cc",
     "fp32/online_fusion/split_reduce_concat_fp32.cc",
diff --git a/mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.cc b/mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.cc
new file mode 100644
index 00000000..b4c3bf7e
--- /dev/null
+++ b/mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.cc
@@ -0,0 +1,51 @@
+/**
+ * Copyright 2022 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/litert/kernel/cpu/fp32/oneslike_fp32.h"
+#include "schema/model_generated.h"
+#include "src/litert/kernel_registry.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_OnesLike;
+
+namespace mindspore::kernel {
+int OnesLikeCPUKernel::Prepare() {
+  CHECK_LESS_RETURN(in_tensors_.size(), 1);
+  CHECK_LESS_RETURN(out_tensors_.size(), 1);
+  return RET_OK;
+}
+
+int OnesLikeCPUKernel::Run() {
+  auto output = out_tensors_[0];
+  CHECK_NULL_RETURN(output);
+  if (output->data_type() == kNumberTypeInt32) {
+    ApproximateOnesLike(static_cast<int *>(output->data()), output->ElementsNum());
+  } else if (output->data_type() == kNumberTypeFloat32) {
+    ApproximateOnesLike(static_cast<float *>(output->data()), output->ElementsNum());
+  }
+  return RET_OK;
+}
+
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_OnesLike, LiteKernelCreator<OnesLikeCPUKernel>)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_OnesLike, LiteKernelCreator<OnesLikeCPUKernel>)
+#ifdef ENABLE_FP16
+REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_OnesLike, LiteKernelCreator<OnesLikeCPUKernel>)
+#endif
+}  // namespace mindspore::kernel
\ No newline at end of file
diff --git a/mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.h b/mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.h
new file mode 100644
index 00000000..f90aebed
--- /dev/null
+++ b/mindspore/lite/src/litert/kernel/cpu/fp32/oneslike_fp32.h
@@ -0,0 +1,46 @@
+/**
+ * Copyright 2022 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_CPU_FP32_ONESLike_FP32_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_CPU_FP32_ONESLike_FP32_H_
+
+#include <vector>
+#include "src/litert/lite_kernel.h"
+
+namespace mindspore::kernel {
+class OnesLikeCPUKernel : public LiteKernel {
+ public:
+  OnesLikeCPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs,
+                    const std::vector<lite::Tensor *> &outputs, const lite::InnerContext *ctx)
+      : LiteKernel(parameter, inputs, outputs, ctx) {}
+
+  ~OnesLikeCPUKernel() = default;
+
+  int Prepare() override;
+  int ReSize() override { return lite::RET_OK; }
+  int Run() override;
+
+ private:
+  template <typename T>
+  void ApproximateOnesLike(T *output, int data_size) {
+    for (int i = 0; i < data_size; ++i) {
+      output[i] = 1;
+    }
+    return;
+  }
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_CPU_FP32_ONESLike_FP32_H_
\ No newline at end of file
diff --git a/mindspore/lite/src/litert/lite_model.h b/mindspore/lite/src/litert/lite_model.h
index 2b5422fa..635b529a 100644
--- a/mindspore/lite/src/litert/lite_model.h
+++ b/mindspore/lite/src/litert/lite_model.h
@@ -1,5 +1,5 @@
 /**
- * Copyright 2020 Huawei Technologies Co., Ltd
+ * Copyright 2020-2023 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
diff --git a/mindspore/lite/src/litert/lite_session.cc b/mindspore/lite/src/litert/lite_session.cc
index ded4d761..8f54879e 100644
--- a/mindspore/lite/src/litert/lite_session.cc
+++ b/mindspore/lite/src/litert/lite_session.cc
@@ -2022,6 +2022,7 @@ int lite::LiteSession::LoadModelAndCompileByPath(const std::string &model_path,
     delete model;
     return RET_ERROR;
   }
+  model->Free();
   set_model(model);
   return RET_OK;
 }
diff --git a/mindspore/lite/src/litert/weight_decoder.h b/mindspore/lite/src/litert/weight_decoder.h
index 9afaca55..9fbcefde 100644
--- a/mindspore/lite/src/litert/weight_decoder.h
+++ b/mindspore/lite/src/litert/weight_decoder.h
@@ -1,5 +1,5 @@
 /**
- * Copyright 2020-2022 Huawei Technologies Co., Ltd
+ * Copyright 2020-2023 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
diff --git a/mindspore/lite/src/tensor.h b/mindspore/lite/src/tensor.h
index 838892cf..f2eb4d1a 100644
--- a/mindspore/lite/src/tensor.h
+++ b/mindspore/lite/src/tensor.h
@@ -69,7 +69,7 @@ enum CompressType {
   kFSEInfer = 6
 };
 
-class Tensor {
+class MS_API Tensor {
  public:
   Tensor() { tensor_c_ = {false, kTypeUnknown, NHWC, VarTensor, nullptr, 0}; }
 
diff --git a/mindspore/lite/src/tensorlist.h b/mindspore/lite/src/tensorlist.h
index bdfdda02..6925cc95 100644
--- a/mindspore/lite/src/tensorlist.h
+++ b/mindspore/lite/src/tensorlist.h
@@ -56,7 +56,7 @@ namespace mindspore::lite {
  *
  *  See the code for other constructors.
  */
-class TensorList : public Tensor {
+class MS_API TensorList : public Tensor {
  public:
   TensorList() { tensor_list_c_ = {false, kObjectTypeTensorType, DEFAULT_FORMAT, 0, kTypeUnknown, -1, nullptr, 0, 0}; }
 
diff --git a/mindspore/lite/src/train/train_session.cc b/mindspore/lite/src/train/train_session.cc
index ef3c71f3..ce6a6437 100644
--- a/mindspore/lite/src/train/train_session.cc
+++ b/mindspore/lite/src/train/train_session.cc
@@ -248,8 +248,8 @@ static int ReshapeWeightTensor(Tensor *orig_tensor, lite::Tensor *new_tensor) {
 
 int TrainSession::UpdateWeights(std::vector<lite::Tensor *> modify_tensors) {
   unsigned int num_of_found_tensors = 0;
-  for (auto tensor : tensors_) {
-    for (auto modify : modify_tensors) {
+  for (auto modify : modify_tensors) {
+    for (auto tensor : tensors_) {
       if (modify == nullptr) {
         MS_LOG(ERROR) << "Tensor is nullptr";
         return RET_PARAM_INVALID;
@@ -1346,7 +1346,7 @@ int TrainSession::ExportWeightsCollaborateWithMicro(const std::string &file_name
 std::vector<lite::Tensor *> TrainSession::GetFeatureMaps() const {
   std::vector<lite::Tensor *> features;
   for (auto cur_tensor : this->tensors_) {
-    if (cur_tensor->IsConst() && cur_tensor->data_type() == kNumberTypeFloat32) {
+    if (cur_tensor->category() == lite::Category::CONST_TENSOR && cur_tensor->data_type() == kNumberTypeFloat32) {
       features.push_back(cur_tensor);
     }
   }
diff --git a/mindspore/lite/tools/benchmark_train/net_train.h b/mindspore/lite/tools/benchmark_train/net_train.h
index 43853e99..67e58a04 100644
--- a/mindspore/lite/tools/benchmark_train/net_train.h
+++ b/mindspore/lite/tools/benchmark_train/net_train.h
@@ -1,5 +1,5 @@
 /**
- * Copyright 2020 Huawei Technologies Co., Ltd
+ * Copyright 2020-2023 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
diff --git a/mindspore/lite/tools/converter/converter_metagraph.cc b/mindspore/lite/tools/converter/converter_metagraph.cc
index 6ffff71c..46a66128 100644
--- a/mindspore/lite/tools/converter/converter_metagraph.cc
+++ b/mindspore/lite/tools/converter/converter_metagraph.cc
@@ -104,12 +104,14 @@ schema::MetaGraphT *ConverterToMetaGraph::Build(const std::shared_ptr<ConverterP
     return nullptr;
   }
 
-  // output name will be modified by Transform
-  status = UpdateMetaGraphOutputName(meta_graph, output_tensor_name);
-  if (status != RET_OK) {
-    MS_LOG(ERROR) << "UpdateGraphOutputName failed.";
-    delete meta_graph;
-    return nullptr;
+  if (!param->train_model) {
+    // output name will be modified by Transform
+    status = UpdateMetaGraphOutputName(meta_graph, output_tensor_name);
+    if (status != RET_OK) {
+      MS_LOG(ERROR) << "UpdateGraphOutputName failed.";
+      delete meta_graph;
+      return nullptr;
+    }
   }
 
   return meta_graph;
diff --git a/mindspore/lite/tools/converter/graphdef_transform.cc b/mindspore/lite/tools/converter/graphdef_transform.cc
index d571b532..90b744e5 100644
--- a/mindspore/lite/tools/converter/graphdef_transform.cc
+++ b/mindspore/lite/tools/converter/graphdef_transform.cc
@@ -26,6 +26,7 @@
 #include "tools/converter/legacy_optimizer/graph/dropout_node_remove_pass.h"
 #include "tools/converter/legacy_optimizer/graph/topological_sort_pass.h"
 #include "tools/converter/legacy_optimizer/graph/tensor_name_pass.h"
+#include "tools/converter/legacy_optimizer/graph/node_name_pass.h"
 #include "tools/converter/legacy_optimizer/graph/set_unused_quant_param_to_default_pass.h"
 #include "tools/converter/legacy_optimizer/graph/convert_fp32_to_fp16_pass.h"
 #include "tools/converter/legacy_optimizer/graph/subgraph_node_pass.h"
@@ -136,6 +137,9 @@ int GraphDefTransform::Transform(const std::shared_ptr<ConverterPara> &param) {
     Optimizer forming_model_optimizer;
     forming_model_optimizer.AddPass(new (std::nothrow) InferShapePass(param->fmk_type));
     forming_model_optimizer.AddPass(new (std::nothrow) SetUnusedQuantParamToDefaultPass(param));
+    if (param->train_model) {
+      forming_model_optimizer.AddPass(new (std::nothrow) NodeNamePass());
+    }
     forming_model_optimizer.AddPass(new (std::nothrow) TensorNamePass());
     forming_model_optimizer.AddPass(new (std::nothrow) ConvertFP32ToFP16Pass(param->weight_fp16));
     status = forming_model_optimizer.Run(graph_defT_);
diff --git a/mindspore/lite/tools/converter/legacy_optimizer/graph/CMakeLists.txt b/mindspore/lite/tools/converter/legacy_optimizer/graph/CMakeLists.txt
index 9b16f4f8..30bccbde 100644
--- a/mindspore/lite/tools/converter/legacy_optimizer/graph/CMakeLists.txt
+++ b/mindspore/lite/tools/converter/legacy_optimizer/graph/CMakeLists.txt
@@ -9,6 +9,7 @@ file(GLOB GRAPH_PASS
         ${CMAKE_CURRENT_SOURCE_DIR}/convert_fp32_to_fp16_pass.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/set_unused_quant_param_to_default_pass.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/tensor_name_pass.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/node_name_pass.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/subgraph_node_pass.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/subgraph_tensor_pass.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/const_node_reorder_pass.cc
diff --git a/mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.cc b/mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.cc
new file mode 100644
index 00000000..712927b0
--- /dev/null
+++ b/mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.cc
@@ -0,0 +1,96 @@
+/**
+ * Copyright 2022 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "tools/converter/legacy_optimizer/graph/node_name_pass.h"
+#include <string>
+#include <vector>
+#include "tools/converter/converter_context.h"
+
+namespace mindspore::lite {
+std::string CutShortName(const std::string &fullname, const std::string &delimiter) {
+  size_t end_pos = fullname.find_last_of(delimiter);
+  std::string name = "";
+  if (end_pos != std::string::npos) {
+    name = fullname.substr(end_pos + 1);
+  }
+  if ((fullname.find("op") != std::string::npos) && (name.find("op") == std::string::npos) &&
+      (end_pos != std::string::npos)) {
+    size_t pos = fullname.rfind(delimiter, end_pos - 1);
+    if (pos != std::string::npos) {
+      name.insert(0, fullname.substr(pos + 1, end_pos - pos));
+    } else {
+      name.insert(0, fullname.substr(0, end_pos + 1));
+    }
+  }
+
+  const std::vector<std::string> loss_names = {"loss_fct", "_loss_fn", "SigmoidCrossEntropy"};
+  for (auto &s : loss_names) {
+    if (fullname.find(s) != std::string::npos) {
+      name.insert(0, s + "/");
+      break;
+    }
+  }
+
+  if (fullname.find("Gradients") != std::string::npos) {
+    size_t pos = fullname.find(delimiter);
+    if (pos != std::string::npos) {
+      name.insert(0, fullname.substr(0, pos + 1));
+    }
+  }
+  return name;
+}
+
+STATUS NodeNamePass::Run(schema::MetaGraphT *graph) {
+  if (graph == nullptr) {
+    MS_LOG(ERROR) << "graph is nullptr";
+    return RET_NULL_PTR;
+  }
+
+  std::string delimiter = "/";
+  for (auto &node : graph->nodes) {
+    if (node == nullptr || node->primitive == nullptr) {
+      MS_LOG(ERROR) << "node or node->primitive is nullptr";
+      return RET_NULL_PTR;
+    }
+    std::string node_name = CutShortName(node->name, delimiter);
+    node->name = node_name != "" ? node_name : node->name;
+
+    for (int i = 0; i < static_cast<int>(node->inputIndex.size()); i++) {
+      auto tensor_id = node->inputIndex.at(i);
+      auto &tensor = graph->allTensors.at(tensor_id);
+      if (tensor->name.empty()) {
+        MS_LOG(DEBUG) << "input tensor (id = " << tensor_id << ") name is null";
+        tensor->name = node->name + "/input-" + std::to_string(i);
+      } else {
+        std::string in_tensor_name = CutShortName(tensor->name, delimiter);
+        tensor->name = in_tensor_name != "" ? in_tensor_name : tensor->name;
+      }
+    }
+
+    for (int i = 0; i < static_cast<int>(node->outputIndex.size()); i++) {
+      auto tensor_id = node->outputIndex.at(i);
+      auto &tensor = graph->allTensors.at(tensor_id);
+      if (tensor->name.empty()) {
+        tensor->name = node->name + "/output-" + std::to_string(i);
+      } else {
+        std::string out_tensor_name = CutShortName(tensor->name, delimiter);
+        tensor->name = out_tensor_name != "" ? out_tensor_name : tensor->name;
+      }
+    }
+  }
+  return RET_OK;
+}
+}  // namespace mindspore::lite
\ No newline at end of file
diff --git a/mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.h b/mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.h
new file mode 100644
index 00000000..4e58e5c7
--- /dev/null
+++ b/mindspore/lite/tools/converter/legacy_optimizer/graph/node_name_pass.h
@@ -0,0 +1,35 @@
+/**
+ * Copyright 2022 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_TOOLS_CONVERTER_LEGACY_OPTIMIZER_GRAPH_NODE_NAME_PASS_H_
+#define MINDSPORE_LITE_TOOLS_CONVERTER_LEGACY_OPTIMIZER_GRAPH_NODE_NAME_PASS_H_
+
+#include <memory>
+#include "tools/converter/optimizer.h"
+#include "tools/common/graph_util.h"
+
+namespace mindspore {
+namespace lite {
+class NodeNamePass : public GraphPass {
+ public:
+  NodeNamePass() {}
+
+  ~NodeNamePass() override = default;
+
+  STATUS Run(schema::MetaGraphT *graph) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_TOOLS_CONVERTER_LEGACY_OPTIMIZER_GRAPH_NODE_NAME_PASS_H_
-- 
2.25.1