File size: 2,133 Bytes
d7e58f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from collections import OrderedDict

import torch.distributed as dist
from mmcv.runner import OptimizerHook
from torch._utils import (
    _flatten_dense_tensors,
    _take_tensors,
    _unflatten_dense_tensors,
)


def _allreduce_coalesced(tensors, world_size, bucket_size_mb=-1):
    if bucket_size_mb > 0:
        bucket_size_bytes = bucket_size_mb * 1024 * 1024
        buckets = _take_tensors(tensors, bucket_size_bytes)
    else:
        buckets = OrderedDict()
        for tensor in tensors:
            tp = tensor.type()
            if tp not in buckets:
                buckets[tp] = []
            buckets[tp].append(tensor)
        buckets = buckets.values()

    for bucket in buckets:
        flat_tensors = _flatten_dense_tensors(bucket)
        dist.all_reduce(flat_tensors)
        flat_tensors.div_(world_size)
        for tensor, synced in zip(
                bucket, _unflatten_dense_tensors(flat_tensors, bucket)):
            tensor.copy_(synced)


def allreduce_grads(params, coalesce=True, bucket_size_mb=-1):
    grads = [
        param.grad.data for param in params
        if param.requires_grad and param.grad is not None
    ]
    world_size = dist.get_world_size()
    if coalesce:
        _allreduce_coalesced(grads, world_size, bucket_size_mb)
    else:
        for tensor in grads:
            dist.all_reduce(tensor.div_(world_size))


class DistOptimizerHook(OptimizerHook):
    def __init__(self, grad_clip=None, coalesce=True, bucket_size_mb=-1):
        self.grad_clip = grad_clip
        self.coalesce = coalesce
        self.bucket_size_mb = bucket_size_mb

    def after_train_iter(self, runner):
        runner.optimizer.zero_grad()
        runner.outputs['loss'].backward()
        if self.grad_clip is not None:
            self.clip_grads(runner.model.parameters())
        runner.optimizer.step()


def reduce_mean(tensor):
    """"Obtain the mean of tensor on different GPUs."""
    if not (dist.is_available() and dist.is_initialized()):
        return tensor
    tensor = tensor.clone()
    dist.all_reduce(tensor.div_(dist.get_world_size()), op=dist.ReduceOp.SUM)
    return tensor