SyncBatchNorm¶

class
torch.nn.
SyncBatchNorm
(num_features: int, eps: float = 1e05, momentum: float = 0.1, affine: bool = True, track_running_stats: bool = True, process_group: Optional[Any] = None)[source]¶ Applies Batch Normalization over a NDimensional input (a minibatch of [N2]D inputs with additional channel dimension) as described in the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift .
$y = \frac{x  \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta$The mean and standarddeviation are calculated perdimension over all minibatches of the same process groups. $\gamma$ and $\beta$ are learnable parameter vectors of size C (where C is the input size). By default, the elements of $\gamma$ are sampled from $\mathcal{U}(0, 1)$ and the elements of $\beta$ are set to 0. The standarddeviation is calculated via the biased estimator, equivalent to torch.var(input, unbiased=False).
Also by default, during training this layer keeps running estimates of its computed mean and variance, which are then used for normalization during evaluation. The running estimates are kept with a default
momentum
of 0.1.If
track_running_stats
is set toFalse
, this layer then does not keep running estimates, and batch statistics are instead used during evaluation time as well.Note
This
momentum
argument is different from one used in optimizer classes and the conventional notion of momentum. Mathematically, the update rule for running statistics here is $\hat{x}_\text{new} = (1  \text{momentum}) \times \hat{x} + \text{momemtum} \times x_t$ , where $\hat{x}$ is the estimated statistic and $x_t$ is the new observed value.Because the Batch Normalization is done for each channel in the
C
dimension, computing statistics on(N, +)
slices, it’s common terminology to call this Volumetric Batch Normalization or Spatiotemporal Batch Normalization.Currently
SyncBatchNorm
only supportsDistributedDataParallel
(DDP) with single GPU per process. Usetorch.nn.SyncBatchNorm.convert_sync_batchnorm()
to convertBatchNorm*D
layer toSyncBatchNorm
before wrapping Network with DDP. Parameters
num_features – $C$ from an expected input of size $(N, C, +)$
eps – a value added to the denominator for numerical stability. Default:
1e5
momentum – the value used for the running_mean and running_var computation. Can be set to
None
for cumulative moving average (i.e. simple average). Default: 0.1affine – a boolean value that when set to
True
, this module has learnable affine parameters. Default:True
track_running_stats – a boolean value that when set to
True
, this module tracks the running mean and variance, and when set toFalse
, this module does not track such statistics and uses batch statistics instead in both training and eval modes if the running mean and variance areNone
. Default:True
process_group – synchronization of stats happen within each process group individually. Default behavior is synchronization across the whole world
 Shape:
Input: $(N, C, +)$
Output: $(N, C, +)$ (same shape as input)
Examples:
>>> # With Learnable Parameters >>> m = nn.SyncBatchNorm(100) >>> # creating process group (optional) >>> # process_ids is a list of int identifying rank ids. >>> process_group = torch.distributed.new_group(process_ids) >>> # Without Learnable Parameters >>> m = nn.BatchNorm3d(100, affine=False, process_group=process_group) >>> input = torch.randn(20, 100, 35, 45, 10) >>> output = m(input) >>> # network is nn.BatchNorm layer >>> sync_bn_network = nn.SyncBatchNorm.convert_sync_batchnorm(network, process_group) >>> # only single gpu per process is currently supported >>> ddp_sync_bn_network = torch.nn.parallel.DistributedDataParallel( >>> sync_bn_network, >>> device_ids=[args.local_rank], >>> output_device=args.local_rank)

classmethod
convert_sync_batchnorm
(module, process_group=None)[source]¶ Helper function to convert all
BatchNorm*D
layers in the model totorch.nn.SyncBatchNorm
layers. Parameters
module (nn.Module) – module containing one or more attr:BatchNorm*D layers
process_group (optional) – process group to scope synchronization, default is the whole world
 Returns
The original
module
with the convertedtorch.nn.SyncBatchNorm
layers. If the originalmodule
is aBatchNorm*D
layer, a newtorch.nn.SyncBatchNorm
layer object will be returned instead.
Example:
>>> # Network with nn.BatchNorm layer >>> module = torch.nn.Sequential( >>> torch.nn.Linear(20, 100), >>> torch.nn.BatchNorm1d(100), >>> ).cuda() >>> # creating process group (optional) >>> # process_ids is a list of int identifying rank ids. >>> process_group = torch.distributed.new_group(process_ids) >>> sync_bn_module = torch.nn.SyncBatchNorm.convert_sync_batchnorm(module, process_group)