syncbn.rst

Implementing Synchronized Multi-GPU Batch Normalization
=======================================================

In this tutorial, we discuss the implementation detail of Multi-GPU Batch Normalization (BN) :class:`encoding.nn.BatchNorm2d` and compatible :class:`encoding.parallel.SelfDataParallel`. We will provide the training example in a later version.

How BN works?
-------------

BN layer was introduced in the paper `Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift <https://arxiv.org/abs/1502.03167>`_, which dramatically speed up the training process of the network (enables larger learning rate) and makes the network less sensitive to the weight initialization. 

- Forward Pass: 
    For the input data :math:`X={x_1, ...x_N}`, the data are normalized to be zero-mean and unit-variance, then scale and shit:

    .. math::
        y_i = \gamma\cdot\frac{x_i-\mu}{\sigma} + \beta ,

    where :math:`\mu=\frac{\sum_i^N x_i}{N} , \sigma = \sqrt{\frac{\sum_i^N (x_i-\mu)^2}{N}+\epsilon}` and :math:`\gamma, \beta` are the learnable parameters.
        
- Backward Pass:
    For calculating the gradient :math:`\frac{d_\ell}{d_{x_i}}`, we need to consider the gradient from :math:`\frac{d_\ell}{d_y}` and the gradients from :math:`\frac{d_\ell}{d_\mu}` and :math:`\frac{d_\ell}{d_\sigma}`, since the :math:`\mu \text{ and } \sigma` are the function of the input :math:`x_i`. We use patial direvative in the notations:

    .. math::

        \frac{d_\ell}{d_{x_i}} = \frac{d_\ell}{d_{y_i}}\cdot\frac{d_{y_i}}{d_{x_i}} + \frac{d_\ell}{d_\mu}\cdot\frac{d_\mu}{d_{x_i}} + \frac{d_\ell}{d_\sigma}\cdot\frac{d_\sigma}{d_{x_i}}

    where :math:`\frac{d_\ell}{d_{x_i}}=\frac{\gamma}{\sigma}, \frac{d_\ell}{d_\mu}=-\frac{\gamma}{\sigma}\sum_i^N\frac{d_\ell}{d_{y_i}} 
    \text{ and } \frac{d_\sigma}{d_{x_i}}=-\frac{1}{\sigma}(\frac{x_i-\mu}{N})`.

Why Synchronize BN?
-------------------

- Standard Implementations of BN in public frameworks (suck as Caffe, MXNet, Torch, TF, PyTorch) are unsynchronized, which means that the data are normalized within each GPU. Therefore the `working batch-size` of the BN layer is `BatchSize/nGPU` (batch-size in each GPU). 

- Since the `working batch-size` is typically large enough for standard vision tasks, such as classification and detection, there is no need to synchronize BN layer during the training. The synchronization will slow down the training.

- However, for the Semantic Segmentation task, the state-of-the-art approaches typically adopt dilated convoluton, which is very memory consuming. The `working bath-size` can be too small for BN layers (2 or 4 in each GPU) when using larger/deeper pre-trained networks, such as :class:`encoding.dilated.ResNet` or :class:`encoding.dilated.DenseNet`. 

How to Synchronize?
-------------------

Suppose we have :math:`K` number of GPUs, :math:`sum(x)_k` and :math:`sum(x^2)_k` denotes the sum of elements and sum of element squares in :math:`k^{th}` GPU.

- Forward Pass:
    We can calculate the sum of elements :math:`sum(x)=\sum x_i \text{ and sum of squares } sum(x^2)=\sum x_i^2` in each GPU, then apply :class:`encoding.parallel.AllReduce` operation to sum accross GPUs. Then calculate the global mean :math:`\mu=\frac{sum(x)}{N} \text{ and global variance } \sigma=\sqrt{\frac{sum(x^2)}{N}-\mu^2+\epsilon}`. 

- Backward Pass:
    * :math:`\frac{d_\ell}{d_{x_i}}=\frac{\gamma}{\sigma}` can be calculated locally in each GPU.
    * Calculate the gradient of :math:`sum(x)` and :math:`sum(x^2)` individually in each GPU :math:`\frac{d_\ell}{d_{sum(x)_k}}` and :math:`\frac{d_\ell}{d_{sum(x^2)_k}}`. 

    * Then Sync the gradient (automatically handled by :class:`encoding.parallel.AllReduce`) and continue the backward.

- Synchronized DataParallel:
    Standard DataParallel pipeline of public frameworks (MXNet, PyTorch...) in each training iters: 

        * duplicate the network (weights) to all the GPUs,
        * split the training batch to each GPU,
        * forward and backward to calculate gradient,
        * update network parameters (weights) then go to next iter.

    Therefore, communicattion accross different GPUs are not supported. To address this problem, we introduce a :class:`encoding.parallel.SelfDataParallel` mode, which enables each layer to accept mutli-GPU inputs directly. Those self-parallel layers are provide in :class:`encoding.nn`.

- Cross GPU Autograd:
    Due to the BN layers are frequently used in the networks, the PyTorch autograd engine will be messed up by such a complicated backward graph. To address this problem, we provide an aotograd function :class:`encoding.parallel.AllReduce` to handle the cross GPU gradient calculation.

Comparing Performance 
---------------------

- Training Time:

- Segmentation Performance:


Citation
--------

.. note::

    This code is provided together with the paper (coming soon), please cite our work.