Dump operator names of a script module #30467

iseeyuan · 2019-11-26T17:45:20Z

Stack from ghstack:

Dump operator names of a script module #30467 Dump operator names of a script module

Introduce function jit.export_opnames(module), which returns a list of all operator names used in the module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size.

Example:
import torch
m = torch.jit.load("example.pt")
print(torch.jit.export_opnames(m))

The outputs are in alphabetical order:
['aten::convolution', 'aten::add.Tensor', 'aten::add.Tensor', 'aten::addmm', 'aten::append.Tensor', 'aten::cat', 'aten::dropout', 'aten::embedding', 'aten::matmul', 'aten::max.dim', 'aten::mul.Tensor', 'aten::permute', 'aten::relu', 'aten::t', 'aten::tanh', 'prim::ListConstruct', 'prim::TupleConstruct', 'prim::TupleUnpack']

Differential Revision: D18801619

ghstack-source-id: c5dad13 Pull Request resolved: #30467

ZolotukhinM

I think this doesn't have to be a method of Module, it can be a standalone pass. Please also find a couple of other comments inline.

ZolotukhinM · 2019-11-26T20:13:23Z

torch/csrc/jit/script/module.cpp

+std::vector<std::string> Module::opnames() const {
+  std::unordered_set<std::string> names;
+  export_opnames(*this, names);
+  return std::vector<std::string>(names.begin(), names.end());


Order of elements in the returned vector is non-deterministic since they come from std::unordered_set. That might result in hard to debug issues, so please make sure that the order is well-specified.

ZolotukhinM · 2019-11-26T20:14:47Z

torch/csrc/jit/script/module.h

      const std::string& filename,
      const ExtraFilesMap& extra_files = ExtraFilesMap()) const;

+  std::vector<std::string> opnames() const;


It doesn't have to be a method of torch::jit::script::Module class. I think it can and should live separately, and you actually already implemented it this way.

ZolotukhinM · 2019-11-26T20:16:32Z

torch/csrc/jit/script/module.cpp

+  auto methods = m.get_methods();
+  for (const auto& method : methods) {
+    const auto& func = method.function();
+    torch::jit::Code code(func.graph());


Why do we need to construct the Code object? Will traversing all nodes in the graph not be enough for some reason?

I think the code would contain the final ops for interpreter (when we bypass the optimizations in mobile). I'm not sure if emitting nodes would introduce more ops in future? Another reason is that I'm not sure how to know if a node is an operator node.

All exporters use nodes directly and hence this function should use that too. Also, Code doesn't do any optimizations, it just wraps some nodes with Instructions - it's an internal detail of how interpreter works and should not affect exporting.

ZolotukhinM · 2019-11-26T20:17:51Z

torch/csrc/jit/script/module.cpp

+namespace {
+void export_opnames(const script::Module& m, std::unordered_set<std::string>& opnames) {
+  auto methods = m.get_methods();
+  for (const auto& method : methods) {


Nit: for (const auto& method : m.get_methods()) {, no need in having methods variable.

Introduce function script.module.opnames(), which returns a list of all operator names used in that module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size. Example: import torch m = torch.jit.load("example.pt") print(m.opnames()) Outputs: ['aten::append.Tensor', 'aten::addmm', 'aten::tanh', 'aten::add_.Tensor', 'aten::matmul', 'aten::relu', 'aten::cat', 'prim::TupleUnpack', 'prim::ListConstruct', 'aten::t', 'aten::_convolution', 'aten::mul.Tensor', 'aten::permute', 'aten::add.Tensor', 'aten::max.dim', 'aten::dropout', 'aten::embedding', 'prim::TupleConstruct']

ghstack-source-id: bc4ffaa Pull Request resolved: #30467

Introduce function jit.opnames(module), which returns a list of all operator names used in the module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size. Example: import torch m = torch.jit.load("example.pt") print(torch.jit.opnames(m)) The outputs are in alphabetical order: ['aten::_convolution', 'aten::add.Tensor', 'aten::add_.Tensor', 'aten::addmm', 'aten::append.Tensor', 'aten::cat', 'aten::dropout', 'aten::embedding', 'aten::matmul', 'aten::max.dim', 'aten::mul.Tensor', 'aten::permute', 'aten::relu', 'aten::t', 'aten::tanh', 'prim::ListConstruct', 'prim::TupleConstruct', 'prim::TupleUnpack']

ghstack-source-id: 1911b10 Pull Request resolved: #30467

ZolotukhinM · 2019-12-02T18:55:18Z

torch/csrc/jit/export.cpp

+}
+} // namespace
+
+std::vector<std::string> opnames(const script::Module& m) {


Nit: It's better to use a verb for the function name (e.g. export_opnames).

ZolotukhinM · 2019-12-02T19:01:37Z

torch/csrc/jit/script/module.cpp

+  auto methods = m.get_methods();
+  for (const auto& method : methods) {
+    const auto& func = method.function();
+    torch::jit::Code code(func.graph());


All exporters use nodes directly and hence this function should use that too. Also, Code doesn't do any optimizations, it just wraps some nodes with Instructions - it's an internal detail of how interpreter works and should not affect exporting.

ZolotukhinM

Thanks, it looks better! I think we still should switch to use nodes directly (without going through Code - please see the comment inline) and also we need some tests for this.

Introduce function jit.opnames(module), which returns a list of all operator names used in the module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size. Example: import torch m = torch.jit.load("example.pt") print(torch.jit.opnames(m)) The outputs are in alphabetical order: ['aten::_convolution', 'aten::add.Tensor', 'aten::add_.Tensor', 'aten::addmm', 'aten::append.Tensor', 'aten::cat', 'aten::dropout', 'aten::embedding', 'aten::matmul', 'aten::max.dim', 'aten::mul.Tensor', 'aten::permute', 'aten::relu', 'aten::t', 'aten::tanh', 'prim::ListConstruct', 'prim::TupleConstruct', 'prim::TupleUnpack']

ghstack-source-id: e9a13fd Pull Request resolved: #30467

Introduce function jit.opnames(module), which returns a list of all operator names used in the module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size. Example: import torch m = torch.jit.load("example.pt") print(torch.jit.opnames(m)) The outputs are in alphabetical order: ['aten::_convolution', 'aten::add.Tensor', 'aten::add_.Tensor', 'aten::addmm', 'aten::append.Tensor', 'aten::cat', 'aten::dropout', 'aten::embedding', 'aten::matmul', 'aten::max.dim', 'aten::mul.Tensor', 'aten::permute', 'aten::relu', 'aten::t', 'aten::tanh', 'prim::ListConstruct', 'prim::TupleConstruct', 'prim::TupleUnpack']

ghstack-source-id: 45a7cec Pull Request resolved: #30467

iseeyuan · 2019-12-03T19:05:55Z

Updated to use nodes in graph directly. @ZolotukhinM

Introduce function jit.export_opnames(module), which returns a list of all operator names used in the module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size. Example: import torch m = torch.jit.load("example.pt") print(torch.jit.export_opnames(m)) The outputs are in alphabetical order: ['aten::_convolution', 'aten::add.Tensor', 'aten::add_.Tensor', 'aten::addmm', 'aten::append.Tensor', 'aten::cat', 'aten::dropout', 'aten::embedding', 'aten::matmul', 'aten::max.dim', 'aten::mul.Tensor', 'aten::permute', 'aten::relu', 'aten::t', 'aten::tanh', 'prim::ListConstruct', 'prim::TupleConstruct', 'prim::TupleUnpack']

ZolotukhinM

This is good to go once you add tests, thanks!

ZolotukhinM · 2019-12-03T23:57:50Z

torch/csrc/jit/export.cpp

+void export_opnames(const script::Module& m, std::set<std::string>& opnames) {
+  for (const auto& method : m.get_methods()) {
+    const auto& func = method.function();
+    torch::jit::Code code(func.graph());


This is unused now.

ZolotukhinM · 2019-12-03T23:59:13Z

torch/csrc/jit/export.cpp

+      if (op) {
+        auto opname = node->schema().operator_name();
+        std::string namestr = opname.name;
+        if (!opname.overload_name.empty())


Nit: use curly-braces even for one-line ifs.

Introduce function jit.export_opnames(module), which returns a list of all operator names used in the module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size. Example: import torch m = torch.jit.load("example.pt") print(torch.jit.export_opnames(m)) The outputs are in alphabetical order: ['aten::_convolution', 'aten::add.Tensor', 'aten::add_.Tensor', 'aten::addmm', 'aten::append.Tensor', 'aten::cat', 'aten::dropout', 'aten::embedding', 'aten::matmul', 'aten::max.dim', 'aten::mul.Tensor', 'aten::permute', 'aten::relu', 'aten::t', 'aten::tanh', 'prim::ListConstruct', 'prim::TupleConstruct', 'prim::TupleUnpack']

ghstack-source-id: ac72910 Pull Request resolved: #30467

Summary: Pull Request resolved: pytorch#30467 Introduce function jit.export_opnames(module), which returns a list of all operator names used in the module and its submodules. One usage is to have mobile custom build to link only operators in the returned list to save the mobile size. Example: import torch m = torch.jit.load("example.pt") print(torch.jit.export_opnames(m)) The outputs are in alphabetical order: ['aten::_convolution', 'aten::add.Tensor', 'aten::add_.Tensor', 'aten::addmm', 'aten::append.Tensor', 'aten::cat', 'aten::dropout', 'aten::embedding', 'aten::matmul', 'aten::max.dim', 'aten::mul.Tensor', 'aten::permute', 'aten::relu', 'aten::t', 'aten::tanh', 'prim::ListConstruct', 'prim::TupleConstruct', 'prim::TupleUnpack'] Test Plan: Imported from OSS Differential Revision: D18801619 Pulled By: iseeyuan fbshipit-source-id: f9b198d3e82b095daf704ee595d8026ad889bb13

Dump operator names of a script module

019c353

iseeyuan requested a review from apaszke as a code owner November 26, 2019 17:45

facebook-github-bot added the oncall: jit Add this issue/PR to JIT oncall triage queue label Nov 26, 2019

Update on "Dump operator names of a script module"

2b8bbb4

iseeyuan pushed a commit that referenced this pull request Nov 26, 2019

Dump operator names of a script module

385fb57

ghstack-source-id: c5dad13 Pull Request resolved: #30467

iseeyuan requested review from ZolotukhinM, ljk53 and suo November 26, 2019 19:56

ZolotukhinM suggested changes Nov 26, 2019

View reviewed changes

iseeyuan pushed a commit that referenced this pull request Nov 27, 2019

Dump operator names of a script module

988e363

ghstack-source-id: bc4ffaa Pull Request resolved: #30467

iseeyuan requested a review from ZolotukhinM November 27, 2019 19:43

iseeyuan pushed a commit that referenced this pull request Nov 27, 2019

Dump operator names of a script module

3f7d93a

ghstack-source-id: 1911b10 Pull Request resolved: #30467

suo removed their request for review November 28, 2019 00:19

ZolotukhinM reviewed Dec 2, 2019

View reviewed changes

iseeyuan pushed a commit that referenced this pull request Dec 3, 2019

Dump operator names of a script module

7d37dfe

ghstack-source-id: e9a13fd Pull Request resolved: #30467

iseeyuan pushed a commit that referenced this pull request Dec 3, 2019

Dump operator names of a script module

590c21e

ghstack-source-id: 45a7cec Pull Request resolved: #30467

iseeyuan requested a review from ZolotukhinM December 3, 2019 15:53

ZolotukhinM approved these changes Dec 4, 2019

View reviewed changes

iseeyuan pushed a commit that referenced this pull request Dec 4, 2019

Dump operator names of a script module

eaaff97

ghstack-source-id: ac72910 Pull Request resolved: #30467

facebook-github-bot closed this in b26401f Dec 4, 2019

facebook-github-bot deleted the gh/iseeyuan/37/head branch December 10, 2019 15:19

Dump operator names of a script module #30467

Dump operator names of a script module #30467

Uh oh!

Conversation

iseeyuan commented Nov 26, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ZolotukhinM left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ZolotukhinM left a comment

Choose a reason for hiding this comment

Uh oh!

iseeyuan commented Dec 3, 2019

Uh oh!

ZolotukhinM left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

iseeyuan commented Nov 26, 2019 •

edited

Loading