pytorch

1
/**
2
 * Copyright (c) 2016-present, Facebook, Inc.
3
 *
4
 * Licensed under the Apache License, Version 2.0 (the "License");
5
 * you may not use this file except in compliance with the License.
6
 * You may obtain a copy of the License at
7
 *
8
 *     http://www.apache.org/licenses/LICENSE-2.0
9
 *
10
 * Unless required by applicable law or agreed to in writing, software
11
 * distributed under the License is distributed on an "AS IS" BASIS,
12
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
 * See the License for the specific language governing permissions and
14
 * limitations under the License.
15
 */
16

17
#include <torch/csrc/jit/api/module.h>
18
#include <torch/csrc/jit/mobile/module.h>
19
#include <torch/csrc/jit/serialization/import.h>
20
#include <torch/csrc/jit/runtime/instruction.h>
21
#include <c10/util/Flags.h>
22

23
#include <fstream>
24

25
namespace torch {
26
namespace jit {
27
void dump_opnames(const Module& m, std::unordered_set<std::string>& opnames) {
28
  auto methods = m.get_methods();
29
  for (const auto& method : methods) {
30
    const auto& func = method.function();
31
    std::cout << "function name: " << func.name() << std::endl;
32
    auto graph = toGraphFunction(func).graph()->copy();
33
    torch::jit::Code code(graph, "");
34
    for (size_t i = 0; i < code.instructions().size(); ++i) {
35
      auto ins = code.instructions()[i];
36
      auto node = code.instructions_source()[i];
37
      if (ins.op == OpCode::OP) {
38
        auto opname = node->schema().operator_name();
39
        std::string namestr = opname.name;
40
        if (!opname.overload_name.empty())
41
          namestr += "." + opname.overload_name;
42
        std::cout << "    " << namestr << std::endl;
43
        opnames.emplace(namestr);
44
      }
45
    }
46
  }
47
  for (const auto& sub_m : m.children()) {
48
    std::cout << "sub module name: " << sub_m.type()->name()->qualifiedName() << std::endl;
49
    dump_opnames(sub_m, opnames);
50
  }
51
}
52
}
53
}
54

55
C10_DEFINE_string(model, "", "The given torch script model.");
56
C10_DEFINE_string(output, "", "The output yaml file of operator list.");
57

58
int main(int argc, char** argv) {
59
  c10::SetUsageMessage(
60
    "Dump operators in a script module and its sub modules.\n"
61
    "Example usage:\n"
62
    "./dump_operator_names"
63
    " --model=<model_file>"
64
    " --output=<output.yaml>");
65

66
  if (!c10::ParseCommandLineFlags(&argc, &argv)) {
67
    std::cerr << "Failed to parse command line flags!" << std::endl;
68
    return 1;
69
  }
70

71
  CAFFE_ENFORCE_GE(FLAGS_model.size(), 0, "Model file must be specified.");
72
  CAFFE_ENFORCE_GE(FLAGS_output.size(), 0, "Output yaml file must be specified.");
73

74
  auto m = torch::jit::load(FLAGS_model);
75
  std::unordered_set<std::string> opnames;
76
  torch::jit::dump_opnames(m, opnames);
77
  std::ofstream ofile(FLAGS_output);
78
  std::cout << "-- Final List --" << std::endl;
79
  for (const auto& name : opnames) {
80
    std::cout << name << std::endl;
81
    ofile << "- " << name << std::endl;
82
  }
83
  ofile.close();
84
}
85