[Feature]: Support empty tensor in MMSyncBN (#1205)

* [Feature]: support empty tensor in MMSyncBN

* refine code

* resolve comments

* clean unnecessary comments

* fix inaccurate statistics when empty tensor

* resolve comments and add docstrings

* update unit tests

* rephrase, ready for merge

mmcv/ops/sync_bn.py

@@ -20,7 +20,7 @@ class SyncBatchNormFunction(Function):
 
     @staticmethod
     def symbolic(g, input, running_mean, running_var, weight, bias, momentum,
-                 eps, group, group_size):
+                 eps, group, group_size, stats_mode):
         return g.op(
             'mmcv::MMCVSyncBatchNorm',
             input,
@@ -31,41 +31,82 @@ class SyncBatchNormFunction(Function):
             momentum_f=momentum,
             eps_f=eps,
             group_i=group,
-            group_size_i=group_size)
+            group_size_i=group_size,
+            stats_mode=stats_mode)
 
     @staticmethod
     def forward(self, input, running_mean, running_var, weight, bias, momentum,
-                eps, group, group_size):
+                eps, group, group_size, stats_mode):
         self.momentum = momentum
         self.eps = eps
         self.group = group
         self.group_size = group_size
+        self.stats_mode = stats_mode
 
         assert isinstance(
                    input, (torch.HalfTensor, torch.FloatTensor,
                            torch.cuda.HalfTensor, torch.cuda.FloatTensor)), \
                f'only support Half or Float Tensor, but {input.type()}'
-        output = torch.empty_like(input)
-        input3d = input.view(input.size(0), input.size(1), -1)
+        output = torch.zeros_like(input)
+        input3d = input.flatten(start_dim=2)
         output3d = output.view_as(input3d)
+        num_channels = input3d.size(1)
 
-        mean = torch.empty(
-            input3d.size(1), dtype=torch.float, device=input3d.device)
-        var = torch.empty(
-            input3d.size(1), dtype=torch.float, device=input3d.device)
-        norm = torch.empty_like(
+        # ensure mean/var/norm/std are initialized as zeros
+        # ``torch.empty()`` does not guarantee that
+        mean = torch.zeros(
+            num_channels, dtype=torch.float, device=input3d.device)
+        var = torch.zeros(
+            num_channels, dtype=torch.float, device=input3d.device)
+        norm = torch.zeros_like(
             input3d, dtype=torch.float, device=input3d.device)
-        std = torch.empty(
-            input3d.size(1), dtype=torch.float, device=input3d.device)
+        std = torch.zeros(
+            num_channels, dtype=torch.float, device=input3d.device)
 
+        batch_size = input3d.size(0)
+        if batch_size > 0:
             ext_module.sync_bn_forward_mean(input3d, mean)
+            batch_flag = torch.ones([1], device=mean.device, dtype=mean.dtype)
+        else:
+            # skip updating mean and leave it as zeros when the input is empty
+            batch_flag = torch.zeros([1], device=mean.device, dtype=mean.dtype)
+
+        # synchronize mean and the batch flag
+        vec = torch.cat([mean, batch_flag])
+        if self.stats_mode == 'N':
+            vec *= batch_size
         if self.group_size > 1:
-            dist.all_reduce(mean, group=self.group)
-            mean /= self.group_size
+            dist.all_reduce(vec, group=self.group)
+        total_batch = vec[-1].detach()
+        mean = vec[:num_channels]
+
+        if self.stats_mode == 'default':
+            mean = mean / self.group_size
+        elif self.stats_mode == 'N':
+            mean = mean / total_batch.clamp(min=1)
+        else:
+            raise NotImplementedError
+
+        # leave var as zeros when the input is empty
+        if batch_size > 0:
             ext_module.sync_bn_forward_var(input3d, mean, var)
+
+        if self.stats_mode == 'N':
+            var *= batch_size
         if self.group_size > 1:
             dist.all_reduce(var, group=self.group)
+
+        if self.stats_mode == 'default':
             var /= self.group_size
+        elif self.stats_mode == 'N':
+            var /= total_batch.clamp(min=1)
+        else:
+            raise NotImplementedError
+
+        # if the total batch size over all the ranks is zero,
+        # we should not update the statistics in the current batch
+        update_flag = total_batch.clamp(max=1)
+        momentum = update_flag * self.momentum
         ext_module.sync_bn_forward_output(
             input3d,
             mean,
@@ -78,7 +119,7 @@ class SyncBatchNormFunction(Function):
             std,
             output3d,
             eps=self.eps,
-            momentum=self.momentum,
+            momentum=momentum,
             group_size=self.group_size)
         self.save_for_backward(norm, std, weight)
         return output
@@ -87,28 +128,67 @@ class SyncBatchNormFunction(Function):
     @once_differentiable
     def backward(self, grad_output):
         norm, std, weight = self.saved_tensors
-        grad_weight = torch.empty_like(weight)
-        grad_bias = torch.empty_like(weight)
-        grad_input = torch.empty_like(grad_output)
-        grad_output3d = grad_output.view(
-            grad_output.size(0), grad_output.size(1), -1)
+        grad_weight = torch.zeros_like(weight)
+        grad_bias = torch.zeros_like(weight)
+        grad_input = torch.zeros_like(grad_output)
+        grad_output3d = grad_output.flatten(start_dim=2)
         grad_input3d = grad_input.view_as(grad_output3d)
+
+        batch_size = grad_input3d.size(0)
+        if batch_size > 0:
             ext_module.sync_bn_backward_param(grad_output3d, norm, grad_weight,
                                               grad_bias)
+
         # all reduce
         if self.group_size > 1:
             dist.all_reduce(grad_weight, group=self.group)
             dist.all_reduce(grad_bias, group=self.group)
             grad_weight /= self.group_size
             grad_bias /= self.group_size
-        ext_module.sync_bn_backward_data(grad_output3d, weight, grad_weight,
-                                         grad_bias, norm, std, grad_input3d)
+
+        if batch_size > 0:
+            ext_module.sync_bn_backward_data(grad_output3d, weight,
+                                             grad_weight, grad_bias, norm, std,
+                                             grad_input3d)
+
         return grad_input, None, None, grad_weight, grad_bias, \
-            None, None, None, None
+            None, None, None, None, None
 
 
 @NORM_LAYERS.register_module(name='MMSyncBN')
 class SyncBatchNorm(Module):
+    """Synchronized Batch Normalization.
+
+    Args:
+        num_features (int): number of features/chennels in input tensor
+        eps (float, optional): a value added to the denominator for numerical
+            stability. Defaults to 1e-5.
+        momentum (float, optional): the value used for the running_mean and
+            running_var computation. Defaults to 0.1.
+        affine (bool, optional): whether to use learnable affine parameters.
+            Defaults to True.
+        track_running_stats (bool, optional): whether to track the running
+            mean and variance during training. When set to False, this
+            module does not track such statistics, and initializes statistics
+            buffers ``running_mean`` and ``running_var`` as ``None``. When
+            these buffers are ``None``, this module always uses batch
+            statistics in both training and eval modes. Defaults to True.
+        group (int, optional): synchronization of stats happen within
+            each process group individually. By default it is synchronization
+            across the whole world. Defaults to None.
+        stats_mode (str, optional): The statistical mode. Available options
+            includes ``'default'`` and ``'N'``. Defaults to 'default'.
+            When ``stats_mode=='default'``, it computes the overall statistics
+            using those from each worker with equal weight, i.e., the
+            statistics are synchronized and simply divied by ``group``. This
+            mode will produce inaccurate statistics when empty tensors occur.
+            When ``stats_mode=='N'``, it compute the overall statistics using
+            the total number of batches in each worker ignoring the number of
+            group, i.e., the statistics are synchronized and then divied by
+            the total batch ``N``. This mode is beneficial when empty tensors
+            occur during training, as it average the total mean by the real
+            number of batch.
+    """
 
     def __init__(self,
                  num_features,
@@ -116,7 +196,8 @@ class SyncBatchNorm(Module):
                  momentum=0.1,
                  affine=True,
                  track_running_stats=True,
-                 group=None):
+                 group=None,
+                 stats_mode='default'):
         super(SyncBatchNorm, self).__init__()
         self.num_features = num_features
         self.eps = eps
@@ -126,6 +207,9 @@ class SyncBatchNorm(Module):
         group = dist.group.WORLD if group is None else group
         self.group = group
         self.group_size = dist.get_world_size(group)
+        assert stats_mode in ['default', 'N'], \
+            f'"stats_mode" only accepts "default" and "N", got "{stats_mode}"'
+        self.stats_mode = stats_mode
         if self.affine:
             self.weight = Parameter(torch.Tensor(num_features))
             self.bias = Parameter(torch.Tensor(num_features))
@@ -174,12 +258,10 @@ class SyncBatchNorm(Module):
                     exponential_average_factor = self.momentum
 
         if self.training or not self.track_running_stats:
-            return SyncBatchNormFunction.apply(input, self.running_mean,
-                                               self.running_var, self.weight,
-                                               self.bias,
-                                               exponential_average_factor,
-                                               self.eps, self.group,
-                                               self.group_size)
+            return SyncBatchNormFunction.apply(
+                input, self.running_mean, self.running_var, self.weight,
+                self.bias, exponential_average_factor, self.eps, self.group,
+                self.group_size, self.stats_mode)
         else:
             return F.batch_norm(input, self.running_mean, self.running_var,
                                 self.weight, self.bias, False,
@@ -192,5 +274,6 @@ class SyncBatchNorm(Module):
         s += f'momentum={self.momentum}, '
         s += f'affine={self.affine}, '
         s += f'track_running_stats={self.track_running_stats}, '
-        s += f'group_size={self.group_size})'
+        s += f'group_size={self.group_size},'
+        s += f'stats_mode={self.stats_mode})'
         return s

tests/data/config/code.py

 from mmcv import Config  # isort:skip
+
 cfg = Config.fromfile('./tests/data/config/a.py')
 item5 = cfg.item1[0] + cfg.item2.a

tests/test_ops/test_syncbn.py

@@ -2,6 +2,7 @@ import os
 import platform
 
 import numpy as np
+import pytest
 import torch
 import torch.distributed as dist
 import torch.nn as nn
@@ -141,6 +142,121 @@ class TestSyncBN(object):
         assert np.allclose(x_grad.data.cpu().numpy(),
                            sx_grad.data.cpu().numpy(), 1e-2)
 
+    def _test_syncbn_empty_train(self, size=1, half=False):
+
+        if 'SLURM_NTASKS' not in os.environ or int(
+                os.environ['SLURM_NTASKS']) != 4:
+            print('must run with slurm has 4 processes!\n'
+                  'srun -p test --gres=gpu:4 -n4')
+            return
+        else:
+            print('Running syncbn test')
+        from mmcv.ops import SyncBatchNorm
+
+        assert size in (1, 2, 4)
+        if not dist.is_initialized():
+            self.dist_init()
+        rank = dist.get_rank()
+
+        torch.manual_seed(9)
+        torch.cuda.manual_seed(9)
+
+        self.x = torch.rand(0, 3, 2, 3).cuda()
+        self.y_bp = torch.rand(0, 3, 2, 3).cuda()
+
+        if half:
+            self.x = self.x.half()
+            self.y_bp = self.y_bp.half()
+        dist.broadcast(self.x, src=0)
+        dist.broadcast(self.y_bp, src=0)
+
+        torch.cuda.synchronize()
+        if size == 1:
+            groups = [None, None, None, None]
+            groups[0] = dist.new_group([0])
+            groups[1] = dist.new_group([1])
+            groups[2] = dist.new_group([2])
+            groups[3] = dist.new_group([3])
+            group = groups[rank]
+        elif size == 2:
+            groups = [None, None, None, None]
+            groups[0] = groups[1] = dist.new_group([0, 1])
+            groups[2] = groups[3] = dist.new_group([2, 3])
+            group = groups[rank]
+        elif size == 4:
+            group = dist.group.WORLD
+
+        syncbn = SyncBatchNorm(3, group=group, stats_mode='N').cuda()
+        syncbn.weight.data[0] = 0.2
+        syncbn.weight.data[1] = 0.5
+        syncbn.weight.data[2] = 0.7
+        syncbn.train()
+
+        bn = nn.BatchNorm2d(3).cuda()
+        bn.weight.data[0] = 0.2
+        bn.weight.data[1] = 0.5
+        bn.weight.data[2] = 0.7
+        bn.train()
+
+        sx = self.x[rank * 4:rank * 4 + 4]
+        sx.requires_grad_()
+        sy = syncbn(sx)
+        sy.backward(self.y_bp[rank * 4:rank * 4 + 4])
+        smean = syncbn.running_mean
+        svar = syncbn.running_var
+        sx_grad = sx.grad
+        sw_grad = syncbn.weight.grad
+        sb_grad = syncbn.bias.grad
+
+        if size == 1:
+            x = self.x[rank * 4:rank * 4 + 4]
+            y_bp = self.y_bp[rank * 4:rank * 4 + 4]
+        elif size == 2:
+            x = self.x[rank // 2 * 8:rank // 2 * 8 + 8]
+            y_bp = self.y_bp[rank // 2 * 8:rank // 2 * 8 + 8]
+        elif size == 4:
+            x = self.x
+            y_bp = self.y_bp
+        x.requires_grad_()
+        y = bn(x)
+        y.backward(y_bp)
+
+        if size == 2:
+            y = y[rank % 2 * 4:rank % 2 * 4 + 4]
+        elif size == 4:
+            y = y[rank * 4:rank * 4 + 4]
+
+        mean = bn.running_mean
+        var = bn.running_var
+        if size == 1:
+            x_grad = x.grad
+            w_grad = bn.weight.grad
+            b_grad = bn.bias.grad
+        elif size == 2:
+            x_grad = x.grad[rank % 2 * 4:rank % 2 * 4 + 4]
+            w_grad = bn.weight.grad / 2
+            b_grad = bn.bias.grad / 2
+        elif size == 4:
+            x_grad = x.grad[rank * 4:rank * 4 + 4]
+            w_grad = bn.weight.grad / 4
+            b_grad = bn.bias.grad / 4
+
+        assert np.allclose(mean.data.cpu().numpy(),
+                           smean.data.cpu().numpy(), 1e-3)
+        assert np.allclose(var.data.cpu().numpy(),
+                           svar.data.cpu().numpy(), 1e-3)
+        assert np.allclose(y.data.cpu().numpy(), sy.data.cpu().numpy(), 1e-3)
+        assert np.allclose(w_grad.data.cpu().numpy(),
+                           sw_grad.data.cpu().numpy(), 1e-3)
+        assert np.allclose(b_grad.data.cpu().numpy(),
+                           sb_grad.data.cpu().numpy(), 1e-3)
+        assert np.allclose(x_grad.data.cpu().numpy(),
+                           sx_grad.data.cpu().numpy(), 1e-2)
+
+        # 'stats_mode' only allows 'default' and 'N'
+        with pytest.raises(AssertionError):
+            SyncBatchNorm(3, group=group, stats_mode='X')
+
     def test_syncbn_1(self):
         self._test_syncbn_train(size=1)
 
@@ -158,3 +274,21 @@ class TestSyncBN(object):
 
     def test_syncbn_4_half(self):
         self._test_syncbn_train(size=4, half=True)
+
+    def test_syncbn_empty_1(self):
+        self._test_syncbn_empty_train(size=1)
+
+    def test_syncbn_empty_2(self):
+        self._test_syncbn_empty_train(size=2)
+
+    def test_syncbn_empty_4(self):
+        self._test_syncbn_empty_train(size=4)
+
+    def test_syncbn_empty_1_half(self):
+        self._test_syncbn_empty_train(size=1, half=True)
+
+    def test_syncbn_empty_2_half(self):
+        self._test_syncbn_empty_train(size=2, half=True)
+
+    def test_syncbn_empty_4_half(self):
+        self._test_syncbn_empty_train(size=4, half=True)