File size: 4,266 Bytes
3ce19a2 | 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 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 | """
-*- coding: utf-8 -*-
File : replicate.py
Author : Jiayuan Mao
Email : maojiayuan@gmail.com
Date : 27/01/2018
This file is part of Synchronized-BatchNorm-PyTorch.
https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
Distributed under MIT License.
MIT License
Copyright (c) 2018 Jiayuan MAO
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""
import functools
from torch.nn.parallel.data_parallel import DataParallel
__all__ = ['CallbackContext', 'execute_replication_callbacks', 'DataParallelWithCallback', 'patch_replication_callback']
class CallbackContext(object):
pass
def execute_replication_callbacks(modules):
"""
Execute an replication callback `__data_parallel_replicate__` on each module created by original replication.
The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)`
Note that, as all modules are isomorphism, we assign each sub-module with a context
(shared among multiple copies of this module on different devices).
Through this context, different copies can share some information.
We guarantee that the callback on the master copy (the first copy) will be called ahead of calling the callback
of any slave copies.
"""
master_copy = modules[0]
nr_modules = len(list(master_copy.modules()))
ctxs = [CallbackContext() for _ in range(nr_modules)]
for i, module in enumerate(modules):
for j, m in enumerate(module.modules()):
if hasattr(m, '__data_parallel_replicate__'):
m.__data_parallel_replicate__(ctxs[j], i)
class DataParallelWithCallback(DataParallel):
"""
Data Parallel with a replication callback.
An replication callback `__data_parallel_replicate__` of each module will be invoked after being created by
original `replicate` function.
The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)`
Examples:
> sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False)
> sync_bn = DataParallelWithCallback(sync_bn, device_ids=[0, 1])
# sync_bn.__data_parallel_replicate__ will be invoked.
"""
def replicate(self, module, device_ids):
modules = super(DataParallelWithCallback, self).replicate(module, device_ids)
execute_replication_callbacks(modules)
return modules
def patch_replication_callback(data_parallel):
"""
Monkey-patch an existing `DataParallel` object. Add the replication callback.
Useful when you have customized `DataParallel` implementation.
Examples:
> sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False)
> sync_bn = DataParallel(sync_bn, device_ids=[0, 1])
> patch_replication_callback(sync_bn)
# this is equivalent to
> sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False)
> sync_bn = DataParallelWithCallback(sync_bn, device_ids=[0, 1])
"""
assert isinstance(data_parallel, DataParallel)
old_replicate = data_parallel.replicate
@functools.wraps(old_replicate)
def new_replicate(module, device_ids):
modules = old_replicate(module, device_ids)
execute_replication_callbacks(modules)
return modules
data_parallel.replicate = new_replicate
|