Merge branch 'sangoly/debug_tools' into 'incubate/lite'

add lite model debug tools

See merge request inference/paddlelite!49

paddle/fluid/lite/CMakeLists.txt

@@ -187,6 +187,7 @@ add_subdirectory(model_parser)
 add_subdirectory(utils)
 add_subdirectory(api)
 add_subdirectory(gen_code)
+add_subdirectory(tools)
 
 
 if (WITH_TESTING)

paddle/fluid/lite/api/cxx_api.cc

@@ -50,18 +50,22 @@ const lite::Tensor *Predictor::GetOutput(size_t offset) {
 }
 
 void Predictor::Build(const std::string &model_path, const Place &prefer_place,
-                      const std::vector<Place> &valid_places) {
+                      const std::vector<Place> &valid_places,
+                      const std::vector<std::string> &passes) {
   LoadModel(model_path, scope_.get(), &program_desc_);
-  Build(program_desc_, prefer_place, valid_places);
+  Build(program_desc_, prefer_place, valid_places, passes);
 }
 
 const framework::proto::ProgramDesc &Predictor::program_desc() const {
   return program_desc_;
 }
 
+const RuntimeProgram &Predictor::runtime_program() const { return *program_; }
+
 void Predictor::Build(const framework::proto::ProgramDesc &desc,
                       const Place &prefer_place,
-                      const std::vector<Place> &valid_places) {
+                      const std::vector<Place> &valid_places,
+                      const std::vector<std::string> &passes) {
   program_desc_ = desc;
   Program program(desc, scope_, valid_places);
 
@@ -69,7 +73,7 @@ void Predictor::Build(const framework::proto::ProgramDesc &desc,
   core::KernelPickFactor factor;
   factor.ConsiderTarget();
   factor.ConsiderPrecision();
-  optimizer_.Run(std::move(program), valid_places, factor);
+  optimizer_.Run(std::move(program), valid_places, factor, passes);
   program_ = optimizer_.GenRuntimeProgram();
 }
 

paddle/fluid/lite/api/cxx_api.h

@@ -39,10 +39,12 @@ class Predictor {
 
   // Build from a model, with places set for hardware config.
   void Build(const std::string& model_path, const Place& prefer_place,
-             const std::vector<Place>& valid_places);
+             const std::vector<Place>& valid_places,
+             const std::vector<std::string>& passes = {});
 
   void Build(const framework::proto::ProgramDesc& desc,
-             const Place& prefer_place, const std::vector<Place>& valid_places);
+             const Place& prefer_place, const std::vector<Place>& valid_places,
+             const std::vector<std::string>& passes = {});
 
   // Run the predictor for a single batch of data.
   void Run() { program_->Run(); }
@@ -53,9 +55,9 @@ class Predictor {
   // Get offset-th col of fetch results.
   const lite::Tensor* GetOutput(size_t offset);
 
-  // Return the program desc for debug.
   const framework::proto::ProgramDesc& program_desc() const;
   const lite::Tensor* GetTensor(const std::string& name) const;
+  const RuntimeProgram& runtime_program() const;
 
   // This method is disabled in mobile, for unnecessary dependencies required.
   void SaveModel(const std::string& dir);

paddle/fluid/lite/core/program.h

@@ -154,6 +154,8 @@ class RuntimeProgram {
 
   size_t num_instructions() const { return instructions_.size(); }
 
+  const std::vector<Instruction>& instructions() const { return instructions_; }
+
  protected:
   std::string SerializeProgram(const framework::proto::ProgramDesc& desc);
   void SaveParams(const std::string& dir,

paddle/fluid/lite/kernels/use_kernels.h

@@ -37,6 +37,7 @@ USE_LITE_KERNEL(pool2d, kARM, kFloat, kNCHW, def);
 USE_LITE_KERNEL(relu, kARM, kFloat, kNCHW, def);
 USE_LITE_KERNEL(transpose, kARM, kFloat, kNCHW, def);
 USE_LITE_KERNEL(transpose2, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(batch_norm, kARM, kFloat, kNCHW, def);
 #endif
 
 #ifdef LITE_WITH_X86

paddle/fluid/lite/tools/CMakeLists.txt

+add_subdirectory(debug)

paddle/fluid/lite/tools/build.sh

@@ -10,10 +10,17 @@ NUM_CORES_FOR_COMPILE=8
 
 # for code gen, a source file is generated after a test, but is dependended by some targets in cmake.
 # here we fake an empty file to make cmake works.
-function prepare_for_codegen {
+function prepare_workspace {
     # in build directory
-    mkdir -p ./paddle/fluid/lite/gen_code
-    touch ./paddle/fluid/lite/gen_code/__generated_code__.cc
+    # 1. Prepare gen_code file
+    GEN_CODE_PATH_PREFIX=paddle/fluid/lite/gen_code
+    mkdir -p ./${GEN_CODE_PATH_PREFIX}
+    touch ./${GEN_CODE_PATH_PREFIX}/__generated_code__.cc
+
+    # 2.Prepare debug tool
+    DEBUG_TOOL_PATH_PREFIX=paddle/fluid/lite/tools/debug
+    mkdir -p ./${DEBUG_TOOL_PATH_PREFIX}
+    cp ../${DEBUG_TOOL_PATH_PREFIX}/analysis_tool.py ./${DEBUG_TOOL_PATH_PREFIX}/
 }
 
 function check_need_ci {
@@ -21,7 +28,7 @@ function check_need_ci {
 }
 
 function cmake_x86 {
-    prepare_for_codegen
+    prepare_workspace
     cmake ..  -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DLITE_WITH_X86=ON ${common_flags}
 }
 
@@ -44,7 +51,7 @@ function cmake_opencl {
 
 # This method is only called in CI.
 function cmake_x86_for_CI {
-    prepare_for_codegen # fake an empty __generated_code__.cc to pass cmake.
+    prepare_workspace # fake an empty __generated_code__.cc to pass cmake.
     cmake ..  -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DLITE_WITH_X86=ON ${common_flags} -DLITE_WITH_PROFILE=ON
 
     # Compile and execute the gen_code related test, so it will generate some code, and make the compilation reasonable.
@@ -56,7 +63,7 @@ function cmake_x86_for_CI {
 }
 
 function cmake_gpu {
-    prepare_for_codegen
+    prepare_workspace
     cmake .. " -DWITH_GPU=ON {common_flags} -DLITE_WITH_GPU=ON"
 }
 
@@ -164,6 +171,7 @@ function test_arm_model {
 }
 
 function cmake_arm {
+    prepare_workspace
     # $1: ARM_TARGET_OS in "android" , "armlinux"
     # $2: ARM_TARGET_ARCH_ABI in "armv8", "armv7" ,"armv7hf"
     # $3: ARM_TARGET_LANG in "gcc" "clang"

paddle/fluid/lite/tools/debug/CMakeLists.txt

+cc_library(debug_utils_lite SRCS debug_utils.cc)
+
+lite_cc_binary(lite_model_debug_tool SRCS model_debug_tool.cc
+    DEPS
+    cxx_api_lite
+    debug_utils_lite
+    model_parser_lite
+    target_wrapper_host
+    mir_passes
+    ${ops_lite} ${host_kernels}
+    X86_DEPS ${x86_kernels}
+    ARM_DEPS ${arm_kernels})

paddle/fluid/lite/tools/debug/analysis_tool.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+'''
+Fluid model analysis tools 
+'''
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import logging
+import os
+import subprocess
+import sys
+from collections import OrderedDict
+from operator import mul
+
+# Simple logging config
+logging.basicConfig(
+    level=logging.DEBUG,
+    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import paddle.fluid as fluid
+from paddle.fluid import debugger
+from paddle.fluid import core
+
+# Command arguments
+parser = argparse.ArgumentParser()
+parser.add_argument(
+    "--model_dir", type=str, required=True, help="Model dir path")
+parser.add_argument(
+    "--input_file", default="", type=str, help="Input datas file path")
+parser.add_argument(
+    "--topo_file",
+    type=str,
+    required=True,
+    help="Runtime topology order output file path")
+parser.add_argument(
+    "--tensor_file",
+    default="",
+    type=str,
+    required=True,
+    help="Tensor file path")
+parser.add_argument(
+    "--tensor_names",
+    default="",
+    type=str,
+    help="If tensor_names is not empty, then only this tensors will be compare")
+parser.add_argument(
+    "--separator",
+    default=",",
+    type=str,
+    help="Deafult separator, use in string split")
+parser.add_argument(
+    "--output_tensor",
+    default=0,
+    type=int,
+    help="dump fluid runntime tensors or not")
+parser.add_argument(
+    "--tensor_output_file",
+    default="./tensor_output_py",
+    type=str,
+    help="dump fluid runntime tensors filepath")
+parser.add_argument(
+    "--tensor_output_length",
+    default=-1,
+    type=int,
+    help="Output tensor data length, dims size will be used if tensor_output_length < 0"
+)
+parser.add_argument(
+    "--only_first",
+    default=1,
+    type=int,
+    help="If only output the first mismatch vars info or not")
+parser.add_argument(
+    "--output_file",
+    default="./diff.txt",
+    type=str,
+    help="dump diff info filepath")
+parser.add_argument(
+    "--threshold", default=1e-5, type=float, help="float value diff threshold")
+
+
+# Help functions
+def load_file(filename, delim=None):
+    """
+    Load file help function
+    """
+    with open(filename) as fd:
+        for line in fd:
+            line = line.strip()
+            assert len(line) != ""
+            if delim:
+                line = line.split(delim)
+            yield line
+
+
+class FluidModelExecutor(object):
+    """
+    A fluid inference model executeor
+    """
+
+    def __init__(self, model_dir, input_file):
+        self.model_dir = model_dir
+        self.place = fluid.CPUPlace()
+        self.exe = fluid.Executor(self.place)
+        self.scope = fluid.core.Scope()
+        self.input_data = self._load_input_file(input_file)
+
+        self.program, self.feed_target_names, self.fetch_targets = self._load_inference_model(
+        )
+
+    def infer_var_list(self,
+                       arg_names=None,
+                       out_data_len=-1,
+                       dump_tensor=False,
+                       dump_tensor_file=''):
+        """
+        Get variables' tensor in var_list
+        """
+        with fluid.scope_guard(self.scope):
+            global_block = self.program.global_block()
+            feed_list = self._prepare_feed_data(global_block,
+                                                self.feed_target_names)
+            fetch_targets = self._fetch_tmp_vars(global_block, arg_names)
+            results = self.exe.run(program=self.program,
+                                   feed=feed_list,
+                                   fetch_list=fetch_targets,
+                                   return_numpy=False)
+            return self._get_results(
+                results,
+                fetch_targets,
+                arg_names=arg_names,
+                need_save=dump_tensor,
+                save_path=dump_tensor_file,
+                out_data_len=out_data_len)
+
+    def draw_graph(self, output_path='./', filename='debug'):
+        """
+        Draw graph with graphviz
+        """
+        dot_path = os.path.join([output_path, filename + '.dot'])
+        pdf_path = os.path.join([output_path, filename + '.pdf'])
+        debugger.draw_block_graphviz(self.program.global_block(), path=dot_path)
+        cmd = ["dot", "-Tpdf", dot_path, "-o", pdf_path]
+        subprocess.Popen(
+            cmd,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE)
+
+    def _prepare_feed_data(self, block, feed_target_names):
+        feed_dict = dict()
+
+        def fill_data(np_dtype, col, shape):
+            if self.input_data:
+                input_size = reduce(mul, shape)
+                assert len(self.input_data[0]) > col
+                data = self.input_data[0][col].split(' ')
+                assert len(data) == input_size
+                return np.array(
+                    map(np_dtype, data), dtype=np_dtype).reshape(shape)
+            else:
+                return np.ones(shape, dtype=np_dtype)
+
+        # TODO(sangoly): support multiple feed fields 
+        assert len(feed_target_names) == 1
+        for idx, name in enumerate(feed_target_names):
+            var = block.var(name)
+            np_shape = list(var.shape)
+            # TODO(sangoly): support batch
+            if np_shape[0] == -1:
+                np_shape[0] = 1
+            if var.dtype == core.VarDesc.VarType.INT32:
+                feed_dict[name] = fill_data(np.int32, idx, np_shape)
+            elif var.dtype == core.VarDesc.VarType.INT64:
+                feed_dict[name] = fill_data(np.int64, idx, np_shape)
+            elif var.dtype == core.VarDesc.VarType.FP16:
+                feed_dict[name] = fill_data(np.float16, idx, np_shape)
+            elif var.dtype == core.VarDesc.VarType.FP32:
+                feed_dict[name] = fill_data(np.float32, idx, np_shape)
+            elif var.dtype == core.VarDesc.VarType.FP64:
+                feed_dict[name] = fill_data(np.float64, idx, np_shape)
+            else:
+                raise TypeError("Data type is not supported")
+        return feed_dict
+
+    def _load_input_file(self, input_file=None):
+        input_data = []
+        if not input_file:
+            return input_data
+        logger.info("Loading input file %s ..." % input_file)
+        for line in load_file(input_file, "\t"):
+            input_data.append(line)
+        return input_data
+
+    def _load_inference_model(self):
+        with fluid.scope_guard(self.scope):
+            model_abs_path = os.path.join(self.model_dir, 'model')
+            param_abs_path = os.path.join(self.model_dir, 'params')
+            if os.path.exists(model_abs_path) and os.path.exists(
+                    param_abs_path):
+                return fluid.io.load_inference_model(self.model_dir, exe,
+                                                     'model', 'params')
+            else:
+                return fluid.io.load_inference_model(self.model_dir, self.exe)
+
+    def _fetch_tmp_vars(self, block, var_names_list=None):
+        fetch_var = block.var('fetch')
+        old_fetch_names = set([var.name for var in self.fetch_targets])
+        new_fetch_vars = [block.var(name) for name in old_fetch_names]
+        i = len(new_fetch_vars)
+        if var_names_list is None:
+            var_names_list = block.vars.keys()
+        for var_name in var_names_list:
+            if var_name in old_fetch_names: continue
+            new_fetch_vars.append(block.var(var_name))
+            block.append_op(
+                type='fetch',
+                inputs={'X': [var_name]},
+                outputs={'Out': [fetch_var]},
+                attrs={'col': i})
+            i = i + 1
+        return new_fetch_vars
+
+    def _get_results(self,
+                     results,
+                     new_fetch_targets,
+                     need_save=False,
+                     arg_names=None,
+                     save_path='',
+                     out_data_len=10):
+        res = OrderedDict()
+        old_fetch_names = set([var.name for var in self.fetch_targets])
+        if need_save:
+            out_fd = open(save_path, 'w')
+        for result in results:
+            idx = results.index(result)
+            name = new_fetch_targets[idx].name
+            dim = [v if v >= 0 else 1 for v in new_fetch_targets[idx].shape]
+            size = min(reduce(mul, dim),
+                       out_data_len) if out_data_len > 0 else reduce(mul, dim)
+            values = list(np.array(result).flatten())[:size]
+            res[name] = {"dim": dim, "values": values}
+            if need_save:
+                if arg_names and name not in arg_names: continue
+                dim_str = '{' + ','.join(map(str, dim)) + '}'
+                out_fd.write('\t'.join(
+                    [name, dim_str, ' '.join(map(str, values))]) + '\n')
+        if need_save:
+            out_fd.close()
+        return res
+
+
+class Analyser(object):
+    """
+    A FLuid model analysis tool
+    """
+
+    def __init__(self, args):
+        self.args = args
+        self.tensors = OrderedDict()
+        self.topo = {}
+        self.input = []
+        logger.info("Loading fluid inference model %s ..." % args.model_dir)
+        self.predictor = FluidModelExecutor(args.model_dir, args.input_file)
+
+    def analysis(self):
+        """
+        Analyser work function
+        """
+        self._load_topo_file()
+        self._load_tensor_file()
+        arg_names = self.args.tensor_names.split(',') if self.args.tensor_names != "" \
+                                           else self.tensors.keys()
+        infer_results = self.predictor.infer_var_list(
+            out_data_len=self.args.tensor_output_length,
+            arg_names=arg_names,
+            dump_tensor=self.args.output_tensor,
+            dump_tensor_file=self.args.tensor_output_file)
+        if self.args.tensor_names == "":
+            self._check_diff_nodes(infer_results)
+
+    def _parse_topo_field(self, field):
+        params = [item.split(':')[1].strip() for item in field[1:-1].split(' ')]
+        params = [item.split('#') for item in params if item != ""]
+        return [item for lst in params for item in lst]
+
+    def _load_topo_file(self):
+        if self.args.topo_file == "":
+            raise ValueError("Topo file path in empty")
+        logger.info("Loading topo file %s ..." % self.args.topo_file)
+        for line in load_file(self.args.topo_file, '\t'):
+            op_type, inputs, outputs = line
+            for name in self._parse_topo_field(outputs):
+                if name not in self.topo:
+                    self.topo[name] = []
+                self.topo[name].append(line)
+
+    def _load_tensor_file(self):
+        if self.args.tensor_file == "":
+            raise ValueError("Tensor file path in empty")
+        logger.info("Loading tensor file %s ..." % args.tensor_file)
+        for line in load_file(args.tensor_file, "\t"):
+            name, dim, values = line
+            dim = map(int, dim[1:-1].split(','))
+            values = map(float, values.split(' '))
+
+            dim_size = reduce(mul, dim)
+            value_size = len(values)
+            assert dim_size == value_size, \
+                        "Dim size mismatch with data: %d vs %d" % (dim_size, value_size)
+
+            self.tensors[name] = {"dim": dim, "values": values}
+
+    def _check_diff_nodes(self, results):
+        """
+        NOTE: The tensor output by c++ debug tool is according to runtime topology order,
+              so we can find the first ops (may be one of them) with error results
+        """
+        assert len(self.tensors) == len(results), \
+                "FLuid output tensor'size mismatch with `tensor_file`"
+        diff_vars = []
+        flag = False
+        for k in self.tensors:
+            if k not in results:
+                raise KeyError("Have not found infer result for `%s`" % k)
+            if len(self.tensors[k]['values']) != len(results[k]['values']):
+                raise ValueError(
+                    "Argname: %s size mismatch with `tensor_file`: %d vs %d" %
+                    (k, len(self.tensors[k]['values']),
+                     len(results[k]['values'])))
+            for i in range(len(self.tensors[k]['values'])):
+                if abs(self.tensors[k]['values'][i] - results[k]['values'][
+                        i]) > args.threshold:
+                    diff_vars.append(k)
+                    if args.only_first:
+                        flag = True
+                    break
+            if flag: break
+        self._output_diff_nodes(results, diff_vars)
+
+    def _output_diff_nodes(self, results, diff_vars):
+        logger.info('is here')
+
+        def output_param_info(inputs, outputs, infos, fd):
+            def tensor_repr(name):
+                return '\t'.join([
+                    name, '{' + ','.join(map(str, infos[name]['dim'])) + '}',
+                    ' '.join(map(str, infos[name]['values']))
+                ])
+
+            for name in self._parse_topo_field(inputs):
+                if name not in infos: continue
+                fd.write(tensor_repr(name) + '\n')
+            for name in self._parse_topo_field(outputs):
+                if name not in infos: continue
+                fd.write(tensor_repr(name) + '\n')
+
+        if len(diff_vars) == 0:
+            logger.info("No diff found. Congratulation!")
+            return
+        logger.info("Total diff vars: %d" % len(diff_vars))
+        with open(self.args.output_file, 'w') as fd:
+            for var in diff_vars:
+                if var not in self.topo:
+                    raise KeyError("%s not in any op's output params, " % var +
+                                   "please check your model and input")
+                fd.write(
+                    '>>>>>>>>>>>>>>>>>>DIFF VARIABLE: %s<<<<<<<<<<<<<<<<<<<\n' %
+                    var)
+                for idx, (op_type, inputs,
+                          outputs) in enumerate(self.topo[var]):
+                    op_repr = '\t'.join([op_type, inputs, outputs])
+                    logger.info("dump diff info: ------------ %s" % op_repr)
+                    fd.write(op_repr + '\n')
+                    fd.write(
+                        "--------------- Tensor File info ---------------\n")
+                    output_param_info(inputs, outputs, self.tensors, fd)
+                    fd.write(
+                        "--------------- Fluid Tensor info ---------------\n")
+                    output_param_info(inputs, outputs, results, fd)
+                    fd.write("\n\n")
+
+
+if __name__ == "__main__":
+    args = parser.parse_args()
+    analyser = Analyser(args)
+    analyser.analysis()

paddle/fluid/lite/tools/debug/check_model.sh

+#!/bin/bash
+
+############################# Arguments ############################
+# For both cpp & python
+BUILD_ROOT_DIR=""                 # Cmake build root path, for LD_LIBRARY_PATH
+MODEL_DIR=""                      # Model dir path
+INPUT_FILE=""                     # Input data file, only the first record will be used. 
+                                  # If the path is empty, then all-ones input will be used.
+CPP_TOPO_FILE=./topo_file.txt     # Runtime program topology info. Write by Cpp-debug-tool and Read by Py-debug-tool
+CPP_TENSOR_FILE=./tensor_cpp.txt  # Store Cpp-debug-tool's tensor outputs int runtime topology order.
+                                  # Write by Cpp-debug-tool and Read by Py-debug-tool 
+TENSOR_NAMES=""                   # If is not empty, then only dump the tensor fo arguments whoes name is 
+                                  # in tensor names. Separate by ','.
+TENSOR_OUTPUT_LENGTH=-1           # Output tensor data length. Tensor's dim size will be used if this value < 0.
+
+# For Cpp debug tools
+CPP_OUTPUT_TOPO=1                 # If output topology info or not.
+CPP_OUTPUT_VARS=1                 # If output TmpVar' tensor or not.
+CPP_OUTPUT_WEIGHTS=1              # If output WeightVar' tensor or not.
+CPP_ARM_THREAD_NUM=1              # ARM thread num. Used by ARM device info. 
+                                  # Only be used by compile option - LITE_WITH_ARM
+
+# For python debug tools
+PY_THRESHOLD=0.00001              # The numerical lower bound  be used to judge [Cpp vs Py] runtime model diff.
+PY_TENSOR_FILE=./tensor_py.txt    # Store Py-debug-tool's tensor outputs.
+PY_OUTPUT_FILE=./diff.txt         # Store model different op/var info for debug.
+PY_ONLY_OUTPUT_FIRST_DIFF=1       # If only output the first different var's info in runtime topology order or not.
+PY_OUTPUT_TENSOR=1                # If output var' tensor in CPP_TENSOR_FILE/TENSOR_NAMES or not.
+
+############################# MAIN #################################
+function print_usage {
+    echo -e "\nUSAGE:"
+    echo -e "debug_cpp_stage -> debug_py_stage"
+    echo
+    echo "----------------------------------------"
+    echo -e "debug_cpp_stage:"
+    echo -e "run_debug.sh [--option=value]* debug_cpp_stage"
+    echo -e "See run_debug.sh#run_cpp_debug_tool for detail"
+    echo
+    echo -e "debug_py_stage:"
+    echo -e "run_debug.sh [--option=value]* debug_py_stage"
+    echo -e "See run_debug.sh#run_py_debug_tool for detail"
+    echo "----------------------------------------"
+}
+
+function check_enviroment {
+    if [ "X${BUILD_ROOT_DIR}" == "X" ]; then
+	echo -e "\nOption: --build_root_dir=xxx is required.\n";
+	exit 1
+    fi 
+    if [ "X${MODEL_DIR}" == "X" ]; then
+	echo -e "\nOption: --model_dir=xxx is required.\n";
+	exit 1
+    fi 
+}
+
+function run_cpp_debug_tool {
+    check_enviroment
+
+    local tool_name="lite_model_debug_tool"
+    local tool_path=$(find ${BUILD_ROOT_DIR} -type f -name ${tool_name})
+    if [ "X${tool_path}" == "X" ]; then
+	echo -e "\nERROR: ${tool_name} not found in ${BUILD_ROOT_DIR}.\n"
+	exit 1
+    fi
+    echo "Find Cpp-debug-tool path: ${tool_path}"
+    export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$BUILD_ROOT_DIR/third_party/install/mklml/lib"
+    ${tool_path} \
+        --model_dir=$MODEL_DIR                         \
+        --input_file=$INPUT_FILE                       \
+        --topo_output_file=$CPP_TOPO_FILE              \
+        --output_topo=$CPP_OUTPUT_TOPO                 \
+        --tensor_output_file=$CPP_TENSOR_FILE          \
+        --output_vars=$CPP_OUTPUT_VARS                 \
+        --output_weights=$CPP_OUTPUT_WEIGHTS           \
+        --tensor_names=$TENSOR_NAMES                   \
+        --tensor_output_length=$TENSOR_OUTPUT_LENGTH   \
+        --arm_thread_num=$CPP_ARM_THREAD_NUM
+}
+
+function run_py_debug_tool {
+    check_enviroment
+
+    local tool_name="analysis_tool.py"
+    local tool_path=$(find ${BUILD_ROOT_DIR} -type f -name ${tool_name})
+    if [ "X${tool_path}" == "X" ]; then
+	echo -e "\nERROR: ${tool_name} not found in ${BUILD_ROOT_DIR}.\n"
+	return
+    fi
+    echo "Find Py-debug-tool path: ${tool_path}"
+    python ${tool_path} \
+        --model_dir=$MODEL_DIR                         \
+        --input_file=$INPUT_FILE                       \
+        --topo_file=$CPP_TOPO_FILE                     \
+        --tensor_file=$CPP_TENSOR_FILE                 \
+        --tensor_names=$TENSOR_NAMES                   \
+        --output_tensor=$PY_OUTPUT_TENSOR              \
+        --tensor_output_file=$PY_TENSOR_FILE           \
+        --tensor_output_length=$TENSOR_OUTPUT_LENGTH   \
+        --only_first=$PY_ONLY_OUTPUT_FIRST_DIFF        \
+        --output_file=$PY_OUTPUT_FILE                  \
+        --threshold=$PY_THRESHOLD
+}
+
+function main {
+    # Parse command line.
+    for i in "$@"; do
+        case $i in
+            --model_dir=*)
+                MODEL_DIR="${i#*=}"
+                shift
+                ;;
+            --input_file=*)
+                INPUT_FILE="${i#*=}"
+                shift
+                ;;
+            --cpp_topo_file=*)
+                CPP_TOPO_FILE="${i#*=}"
+                shift
+                ;;
+            --cpp_tensor_file=*)
+                CPP_TENSOR_FILE="${i#*=}"
+                shift
+                ;;
+            --tensor_names=*)
+                TENSOR_NAMES="${i#*=}"
+                shift
+                ;;
+            --tensor_output_length=*)
+                TENSOR_OUTPUT_LENGTH="${i#*=}"
+                shift
+                ;;
+            --cpp_output_vars=*)
+                CPP_OUTPUT_VARS="${i#*=}"
+                shift
+                ;;
+            --cpp_output_weights=*)
+                CPP_OUTPUT_WEIGHTS="${i#*=}"
+                shift
+                ;;
+            --py_threshold=*)
+                PY_THRESHOLD="${i#*=}"
+                shift
+                ;;
+            --py_tensor_file=*)
+                PY_TENSOR_FILE="${i#*=}"
+                shift
+                ;;
+            --py_output_file=*)
+                PY_OUTPUT_FILE="${i#*=}"
+                shift
+                ;;
+            --py_only_output_first_diff=*)
+                PY_ONLY_OUTPUT_FIRST_DIFF="${i#*=}"
+                shift
+                ;;
+            --py_output_tensor=*)
+                PY_OUTPUT_TENSOR="${i#*=}"
+                shift
+                ;;
+	    --build_root_dir=*)
+		BUILD_ROOT_DIR="${i#*=}"
+		shift
+		;;
+            debug_cpp_stage)
+                run_cpp_debug_tool
+                shift
+                ;;
+            debug_py_stage)
+                run_py_debug_tool
+                shift
+                ;;
+            *)
+                # unknown option
+                print_usage
+                exit 1
+                ;;
+        esac
+    done
+}
+
+main $@

paddle/fluid/lite/tools/debug/debug_utils.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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 "paddle/fluid/lite/tools/debug/debug_utils.h"

paddle/fluid/lite/tools/debug/debug_utils.h

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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.
+
+#pragma once
+#include <gflags/gflags.h>
+#include <algorithm>
+#include <fstream>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/lite/api/cxx_api.h"
+#include "paddle/fluid/lite/core/compatible_tensor.h"
+#include "paddle/fluid/lite/utils/string.h"
+
+DEFINE_string(model_dir, "", "Model dir path");
+DEFINE_string(input_file, "", "Input datas file path");
+DEFINE_string(topo_output_file, "", "Runtime topology order output file path");
+DEFINE_bool(output_topo, true, "Dump runtime topology or not");
+DEFINE_string(tensor_output_file, "", "Tensor output file path");
+DEFINE_bool(output_vars, true, "Dump vars or not");
+DEFINE_bool(output_weights, true, "Dump weight tensors or not");
+DEFINE_string(
+    tensor_names, "",
+    "If tensor_names is not empty, then only this tensors will be dump");
+DEFINE_int32(tensor_output_length, -1,
+             "Output tensor data length, dims size will be used if "
+             "output_tensor_length < 0");
+DEFINE_int32(arm_thread_num, 1, "Arm thread nums, 1 as default");
+DEFINE_string(separator, ",", "Deafult separator, use in string split");
+
+namespace paddle {
+namespace lite {
+namespace tools {
+namespace debug {
+
+struct DebugConfig {
+  // arguments
+  std::string model_dir;
+  std::string topo_output_file;
+  std::string tensor_output_file;
+  std::string input_file;
+  std::vector<std::string> tensor_names;
+  bool output_weights;
+  bool output_topo;
+  bool output_vars;
+  int tensor_output_length;
+  int arm_thread_num;
+
+  std::unordered_map<std::string, lite::VarDesc> var_descs;
+  std::vector<std::vector<std::string>> input_values;
+};
+
+template <typename T>
+std::vector<T> Split2Vector(const std::string& input,
+                            const std::string& separator) {
+  std::vector<T> tgt;
+  std::vector<std::string> inputs = Split(input, separator);
+  tgt.resize(inputs.size());
+  std::stringstream ss;
+  for (int i = 0; i < inputs.size(); ++i) {
+    ss << inputs[i] << " ";
+  }
+  for (int i = 0; i < inputs.size(); ++i) {
+    ss >> tgt[i];
+  }
+  return tgt;
+}
+
+void CollectFeedVarsInfo(std::unordered_map<int, std::string>* feed_vars_info,
+                         const framework::proto::ProgramDesc& desc) {
+  CHECK(feed_vars_info);
+  for (const auto& proto_op_desc : desc.blocks(0).ops()) {
+    lite::OpDesc op_desc(proto_op_desc);
+    auto op_type = op_desc.Type();
+    if (op_type == "feed") {
+      (*feed_vars_info)
+          .emplace(op_desc.GetAttr<int>("col"), op_desc.Output("Out").front());
+    }
+  }
+}
+template <typename T>
+void FillTensorData(lite::Tensor* tensor, const DebugConfig& conf, int col) {
+  CHECK(tensor);
+  auto dim_size = tensor->dims().production();
+  auto* data = tensor->mutable_data<T>();
+  if (conf.input_values.size() > 0) {
+    CHECK(col < conf.input_values[0].size())
+        << "Input data fields out of index. field_len: "
+        << conf.input_values[0].size() << " col: " << col;
+    std::vector<T> input_data(
+        std::move(Split2Vector<T>(conf.input_values[0][col], " ")));
+    CHECK(input_data.size() == dim_size)
+        << "Input data field[" << col
+        << "] mismatch TensorDim: " << input_data.size() << " vs " << dim_size;
+    for (int i = 0; i < dim_size; i++) {
+      data[i] = input_data[i];
+    }
+  } else {
+    LOG(INFO) << "------------> Use all-ones input";
+    for (int i = 0; i < dim_size; i++) {
+      data[i] = 1;
+    }
+  }
+}
+
+void CheckDim(std::vector<DDim::value_type>* dim) {
+  CHECK(dim);
+  for (int i = 0; i < dim->size(); ++i) {
+    if ((*dim)[i] < 0) (*dim)[i] = -(*dim)[i];
+  }
+}
+
+void PrepareModelInputTensor(const DebugConfig& conf, lite::Scope* scope,
+                             const framework::proto::ProgramDesc& desc) {
+  CHECK(scope);
+
+  std::unordered_map<int, std::string> feed_vars_info;
+  CollectFeedVarsInfo(&feed_vars_info, desc);
+  auto* feed_var =
+      scope->FindVar("feed")->GetMutable<std::vector<lite::Tensor>>();
+  feed_var->resize(feed_vars_info.size());
+
+  for (auto& item : feed_vars_info) {
+    auto& var_desc = conf.var_descs.at(item.second);
+    auto val_type = var_desc.GetDataType();
+    auto dim = var_desc.GetShape();
+    CheckDim(&dim);
+    auto* input_tensor = &feed_var->at(item.first);
+    input_tensor->Resize(DDim(dim));
+    switch (val_type) {
+#define FILL_TENSOR_BY_TYPE_ONCE(pb_type__, type__)         \
+  case framework::proto::VarType::pb_type__:                \
+    FillTensorData<type__>(input_tensor, conf, item.first); \
+    break
+
+      FILL_TENSOR_BY_TYPE_ONCE(UINT8, uint8_t);
+      FILL_TENSOR_BY_TYPE_ONCE(INT8, int8_t);
+      FILL_TENSOR_BY_TYPE_ONCE(INT16, int16_t);
+      FILL_TENSOR_BY_TYPE_ONCE(INT32, int32_t);
+      FILL_TENSOR_BY_TYPE_ONCE(INT64, int64_t);
+      FILL_TENSOR_BY_TYPE_ONCE(FP32, float);
+      FILL_TENSOR_BY_TYPE_ONCE(FP64, double);
+
+      default:
+        LOG(FATAL) << "Unsupported data type: " << static_cast<int>(val_type);
+#undef FILL_TENSOR_BY_TYPE_ONCE
+    }
+  }
+}
+
+void ParseInputFile(DebugConfig* conf) {
+  CHECK(conf);
+  if (conf->input_file.empty()) return;
+  auto& inputs = conf->input_values;
+  std::ifstream fd(conf->input_file);
+  CHECK(fd.is_open()) << "Open input file: " << conf->input_file << " failed!";
+  std::string line;
+  while (std::getline(fd, line)) {
+    inputs.emplace_back(std::move(Split(line, FLAGS_separator)));
+  }
+  LOG(INFO) << "Load data:" << inputs.size() << " items";
+}
+
+void ParseConfig(DebugConfig* conf) {
+  CHECK(conf);
+#define CHECK_NON_EMPTY(name__) \
+  CHECK(!FLAGS_##name__.empty()) << "Option " << #name__ << " can't be empty."
+  CHECK_NON_EMPTY(model_dir);
+  if (FLAGS_output_topo) {
+    CHECK_NON_EMPTY(topo_output_file);
+  }
+  if (FLAGS_output_vars || FLAGS_output_weights) {
+    CHECK_NON_EMPTY(tensor_output_file);
+  }
+#undef CHECK_NON_EMPTY
+  conf->model_dir = FLAGS_model_dir;
+  conf->topo_output_file = FLAGS_topo_output_file;
+  conf->tensor_output_file = FLAGS_tensor_output_file;
+  conf->input_file = FLAGS_input_file;
+  conf->output_weights = FLAGS_output_weights;
+  conf->output_vars = FLAGS_output_vars;
+  conf->output_topo = FLAGS_output_topo;
+  conf->tensor_output_length = FLAGS_tensor_output_length;
+  conf->arm_thread_num = FLAGS_arm_thread_num;
+
+  if (!FLAGS_tensor_names.empty()) {
+    conf->tensor_names = Split(FLAGS_tensor_names, FLAGS_separator);
+  }
+
+  ParseInputFile(conf);
+}
+
+void CollectAndDumpTopoInfo(const std::vector<Instruction>& instructions,
+                            const DebugConfig& conf) {
+  if (!conf.output_topo) return;
+  LOG(INFO) << "----------------- dump topo file";
+  std::ofstream os(conf.topo_output_file);
+  CHECK(os.is_open());
+  for (auto& inst : instructions) {
+    auto* op_info = inst.op()->op_info();
+    CHECK(op_info);
+    os << op_info->Type() << "\t";
+    os << "(";
+#define DUMP_TOPO_INFO_ONCE(name__)                   \
+  {                                                   \
+    auto argnames = op_info->name__##ArgumentNames(); \
+    for (int i = 0; i < argnames.size(); ++i) {       \
+      os << argnames[i] << ":";                       \
+      auto vars = op_info->name__(argnames[i]);       \
+      for (int j = 0; j < vars.size(); ++j) {         \
+        os << vars[j];                                \
+        if (j != vars.size() - 1) os << "#";          \
+      }                                               \
+      if (i != argnames.size() - 1) os << " ";        \
+    }                                                 \
+  }
+    DUMP_TOPO_INFO_ONCE(Input);
+    os << ")\t(";
+    DUMP_TOPO_INFO_ONCE(Output);
+    os << ")\n";
+#undef DUMP_TOPO_INFO_ONCE
+  }
+  os.close();
+}
+
+void CollectVarDescs(std::unordered_map<std::string, lite::VarDesc>* var_descs,
+                     const framework::proto::ProgramDesc& desc) {
+  CHECK(var_descs);
+  CHECK(!desc.blocks().empty());
+  std::unordered_set<std::string> weights;
+  for (auto proto_var_desc : desc.blocks(0).vars()) {
+    lite::VarDesc var_desc(proto_var_desc);
+    (*var_descs).emplace(var_desc.Name(), std::move(var_desc));
+  }
+}
+
+std::unordered_set<std::string> CollectUnusedVars(
+    const std::vector<Instruction>& instructions) {
+  std::unordered_set<std::string> unused;
+  std::unordered_set<std::string> all_inputs;
+  for (auto& inst : instructions) {
+    for (const auto& name : inst.op()->op_info()->input_names()) {
+      all_inputs.insert(name);
+    }
+  }
+
+  for (auto& inst : instructions) {
+    for (const auto& name : inst.op()->op_info()->output_names()) {
+      if (all_inputs.count(name) == 0) unused.insert(name);
+    }
+  }
+
+  return unused;
+}
+
+std::string GetTensorRepr(const lite::Tensor& tensor, int out_data_len) {
+  std::stringstream ss;
+  auto size = tensor.dims().production();
+  if (out_data_len >= 0) {
+    size = std::min(size, static_cast<DDim::value_type>(out_data_len));
+  }
+  for (int i = 0; i < size; i++) {
+    ss << tensor.template data<float>()[i];
+    if (i != size - 1) ss << " ";
+  }
+  return ss.str();
+}
+
+void CollectAndDumpTensorInfo(const std::vector<Instruction>& instructions,
+                              const framework::proto::ProgramDesc& desc,
+                              const DebugConfig& conf) {
+  CHECK(instructions.size() > 0) << "No instruction found";
+  const auto* scope = const_cast<lite::OpLite*>(instructions[0].op())->scope();
+  std::ofstream os(conf.tensor_output_file);
+  CHECK(os.is_open());
+
+  std::unordered_set<std::string> dump_vars;
+#define DUMP_TENSOR_ONCE(name__)                                  \
+  LOG(INFO) << "----------------- dump tensor: " << name__;       \
+  auto& tensor = scope->FindVar(name__)->Get<lite::Tensor>();     \
+  os << name__ << "\t" << tensor.dims() << "\t"                   \
+     << GetTensorRepr(tensor, conf.tensor_output_length) << "\n"; \
+  dump_vars.insert(name__)
+
+#define DUMP_OP_TENSOR_ONCE(name__, skip__)                              \
+  for (const auto& name : inst.op()->op_info()->name__##_names()) {      \
+    bool is_weight = conf.var_descs.at(name).Persistable();              \
+    if (unused.count(name) != 0 || name == #skip__ ||                    \
+        (!conf.output_weights && is_weight) ||                           \
+        (!conf.output_vars && !is_weight) || dump_vars.count(name) != 0) \
+      continue;                                                          \
+    DUMP_TENSOR_ONCE(name);                                              \
+  }
+
+  if (conf.tensor_names.size() == 0) {
+    std::unordered_set<std::string> unused(
+        std::move(CollectUnusedVars(instructions)));
+
+    for (auto& inst : instructions) {
+      DUMP_OP_TENSOR_ONCE(input, feed);
+      DUMP_OP_TENSOR_ONCE(output, fetch);
+    }
+  } else {
+    for (const auto& name : conf.tensor_names) {
+      DUMP_TENSOR_ONCE(name);
+    }
+  }
+#undef DUMP_OP_TENSOR_ONCE
+#undef DUMP_TENSOR_ONCE
+  os.close();
+}
+
+}  // namespace debug
+}  // namespace tools
+}  // namespace lite
+}  // namespace paddle

paddle/fluid/lite/tools/debug/model_debug_tool.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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 <sstream>
+#include <string>
+#include <vector>
+#include "paddle/fluid/lite/api/cxx_api.h"
+#include "paddle/fluid/lite/core/mir/use_passes.h"
+#include "paddle/fluid/lite/core/op_registry.h"
+#include "paddle/fluid/lite/kernels/use_kernels.h"
+#include "paddle/fluid/lite/operators/use_ops.h"
+#include "paddle/fluid/lite/tools/debug/debug_utils.h"
+
+namespace paddle {
+namespace lite {
+namespace tools {
+namespace debug {
+
+void Run(DebugConfig* conf) {
+  CHECK(conf);
+#ifdef LITE_WITH_ARM
+  DeviceInfo::Init();
+  DeviceInfo::Global().SetRunMode(LITE_POWER_HIGH, conf->arm_thread_num);
+#endif
+  lite::Predictor predictor;
+  std::vector<Place> valid_places({
+      Place{TARGET(kHost), PRECISION(kFloat)},
+#ifdef LITE_WITH_ARM
+      Place{TARGET(kARM), PRECISION(kFloat)},
+#endif
+#ifdef LITE_WITH_X86
+      Place{TARGET(kX86), PRECISION(kFloat)},
+#endif
+  });
+
+  std::vector<std::string> passes{{
+      "static_kernel_pick_pass", "variable_place_inference_pass",
+      "type_target_transform_pass", "variable_place_inference_pass",
+      "io_copy_kernel_pick_pass", "variable_place_inference_pass",
+      "runtime_context_assign_pass",
+  }};
+
+  predictor.Build(conf->model_dir,
+#ifdef LITE_WITH_ARM
+                  Place{TARGET(kARM), PRECISION(kFloat)},
+#endif
+#ifdef LITE_WITH_X86
+                  Place{TARGET(kX86), PRECISION(kFloat)},
+#endif
+                  valid_places, passes);
+
+  auto& instructions = predictor.runtime_program().instructions();
+  auto& program_desc = predictor.program_desc();
+  auto* scope = const_cast<lite::OpLite*>(instructions[0].op())->scope();
+
+  CollectVarDescs(&(conf->var_descs), program_desc);
+  PrepareModelInputTensor(*conf, scope, program_desc);
+  predictor.Run();
+
+  CollectAndDumpTopoInfo(instructions, *conf);
+  CollectAndDumpTensorInfo(instructions, program_desc, *conf);
+
+  // TODO(sangoly): Maybe add some profile info here
+  auto* out = predictor.GetOutput(0);
+  LOG(INFO) << out << " memory size " << out->data_size();
+  LOG(INFO) << "out " << out->data<float>()[0];
+  LOG(INFO) << "dims " << out->dims();
+  LOG(INFO) << "out data size: " << out->data_size();
+}
+
+}  // namespace debug
+}  // namespace tools
+}  // namespace lite
+}  // namespace paddle
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  paddle::lite::tools::debug::DebugConfig conf;
+  paddle::lite::tools::debug::ParseConfig(&conf);
+  paddle::lite::tools::debug::Run(&conf);
+
+  return 0;
+}

paddle/fluid/lite/utils/string.h

@@ -74,14 +74,21 @@ static std::string Repr(const std::vector<std::string>& v) {
   return "{" + Join(tmp, ",") + "}";
 }
 
-static std::vector<std::string> Split(const std::string& s, char delim) {
-  std::stringstream ss(s);
-  std::string line;
-  std::vector<std::string> res;
-  while (std::getline(ss, line, delim)) {
-    res.push_back(line);
+static std::vector<std::string> Split(const std::string& original,
+                                      const std::string& separator) {
+  std::vector<std::string> results;
+  std::string::size_type pos1, pos2;
+  pos2 = original.find(separator);
+  pos1 = 0;
+  while (std::string::npos != pos2) {
+    results.push_back(original.substr(pos1, pos2 - pos1));
+    pos1 = pos2 + separator.size();
+    pos2 = original.find(separator, pos1);
   }
-  return res;
+  if (pos1 != original.length()) {
+    results.push_back(original.substr(pos1));
+  }
+  return results;
 }
 
 }  // namespace lite