File: dedup_module_uses.h

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (28 lines) | stat: -rw-r--r-- 826 bytes parent folder | download | duplicates (2)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
#pragma once

#include <torch/csrc/jit/api/module.h>

namespace torch {
namespace jit {

/** Recursively deduplicate multiple uses of the same module by
 *  creating an instance clone for each use of the module, which means
 *  the type will be the same as before and all the attributes will be
 *  copied, then we'll change the use of the original module to the use
 *  of cloned module in the Graph.
 *
 *  This is done to ensure that modules can survive destructive passes
 *  without changing model behavior. For example, here:
 *
 *    x = self.conv1(x)
 *    x = self.relu(x)
 *    x = self.conv2(x)
 *    x = self.relu(x)
 *
 *  self.relu needs to be deduplicated for potential future destructive passes
 *  to work properly.
 */
TORCH_API void DedupModuleUses(Module& module);

} // namespace jit
} // namespace torch