[Feature] Add revert_sync_batchnorm (#1253)

* [Feature] Add revert_sync_batchnorm

* support mmsyncbn (to be tested)

* Test passed

* Update docstring, rename the test file

* remove test_sync_bn

* add comment

* add mmcv.ops check

* Add a comment

* Add notes and relax test req

Co-authored-by: gaotongxiao <gaotongxiao@gmail.con>

mmcv/cnn/utils/__init__.py

@@ -1,6 +1,7 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from .flops_counter import get_model_complexity_info
 from .fuse_conv_bn import fuse_conv_bn
+from .sync_bn import revert_sync_batchnorm
 from .weight_init import (INITIALIZERS, Caffe2XavierInit, ConstantInit,
                           KaimingInit, NormalInit, PretrainedInit,
                           TruncNormalInit, UniformInit, XavierInit,
@@ -14,5 +15,5 @@ __all__ = [
     'uniform_init', 'xavier_init', 'fuse_conv_bn', 'initialize',
     'INITIALIZERS', 'ConstantInit', 'XavierInit', 'NormalInit',
     'TruncNormalInit', 'UniformInit', 'KaimingInit', 'PretrainedInit',
-    'Caffe2XavierInit'
+    'Caffe2XavierInit', 'revert_sync_batchnorm'
 ]

mmcv/cnn/utils/sync_bn.py

@@ -0,0 +1,59 @@
+import torch
+
+import mmcv
+
+
+class _BatchNormXd(torch.nn.modules.batchnorm._BatchNorm):
+    """A general BatchNorm layer without input dimension check.
+
+    Reproduced from @kapily's work:
+    (https://github.com/pytorch/pytorch/issues/41081#issuecomment-783961547)
+    The only difference between BatchNorm1d, BatchNorm2d, BatchNorm3d, etc
+    is `_check_input_dim` that is designed for tensor sanity checks.
+    The check has been bypassed in this class for the convenience of converting
+    SyncBatchNorm.
+    """
+
+    def _check_input_dim(self, input):
+        return
+
+
+def revert_sync_batchnorm(module):
+    """Helper function to convert all `SyncBatchNorm` (SyncBN) and
+    `mmcv.ops.sync_bn.SyncBatchNorm`(MMSyncBN) layers in the model to
+    `BatchNormXd` layers.
+
+    Adapted from @kapily's work:
+    (https://github.com/pytorch/pytorch/issues/41081#issuecomment-783961547)
+
+    Args:
+        module (nn.Module): The module containing `SyncBatchNorm` layers.
+
+    Returns:
+        module_output: The converted module with `BatchNormXd` layers.
+    """
+    module_output = module
+    module_checklist = [torch.nn.modules.batchnorm.SyncBatchNorm]
+    if hasattr(mmcv, 'ops'):
+        module_checklist.append(mmcv.ops.SyncBatchNorm)
+    if isinstance(module, tuple(module_checklist)):
+        module_output = _BatchNormXd(module.num_features, module.eps,
+                                     module.momentum, module.affine,
+                                     module.track_running_stats)
+        if module.affine:
+            # no_grad() may not be needed here but
+            # just to be consistent with `convert_sync_batchnorm()`
+            with torch.no_grad():
+                module_output.weight = module.weight
+                module_output.bias = module.bias
+        module_output.running_mean = module.running_mean
+        module_output.running_var = module.running_var
+        module_output.num_batches_tracked = module.num_batches_tracked
+        module_output.training = module.training
+        # qconfig exists in quantized models
+        if hasattr(module, 'qconfig'):
+            module_output.qconfig = module.qconfig
+    for name, child in module.named_children():
+        module_output.add_module(name, revert_sync_batchnorm(child))
+    del module
+    return module_output

tests/test_cnn/test_revert_syncbn.py

@@ -0,0 +1,58 @@
+import os
+import platform
+
+import numpy as np
+import pytest
+import torch
+import torch.distributed as dist
+
+from mmcv.cnn.bricks import ConvModule
+from mmcv.cnn.utils import revert_sync_batchnorm
+
+if platform.system() == 'Windows':
+    import regex as re
+else:
+    import re
+
+
+def test_revert_syncbn():
+    conv = ConvModule(3, 8, 2, norm_cfg=dict(type='SyncBN'))
+    x = torch.randn(1, 3, 10, 10)
+    # Expect a ValueError prompting that SyncBN is not supported on CPU
+    with pytest.raises(ValueError):
+        y = conv(x)
+    conv = revert_sync_batchnorm(conv)
+    y = conv(x)
+    assert y.shape == (1, 8, 9, 9)
+
+
+def test_revert_mmsyncbn():
+    if 'SLURM_NTASKS' not in os.environ or int(os.environ['SLURM_NTASKS']) < 2:
+        print('Must run on slurm with more than 1 process!\n'
+              'srun -p test --gres=gpu:2 -n2')
+        return
+    rank = int(os.environ['SLURM_PROCID'])
+    world_size = int(os.environ['SLURM_NTASKS'])
+    local_rank = int(os.environ['SLURM_LOCALID'])
+    node_list = str(os.environ['SLURM_NODELIST'])
+
+    node_parts = re.findall('[0-9]+', node_list)
+    os.environ['MASTER_ADDR'] = (f'{node_parts[1]}.{node_parts[2]}' +
+                                 f'.{node_parts[3]}.{node_parts[4]}')
+    os.environ['MASTER_PORT'] = '12341'
+    os.environ['WORLD_SIZE'] = str(world_size)
+    os.environ['RANK'] = str(rank)
+
+    dist.init_process_group('nccl')
+    torch.cuda.set_device(local_rank)
+    x = torch.randn(1, 3, 10, 10).cuda()
+    dist.broadcast(x, src=0)
+    conv = ConvModule(3, 8, 2, norm_cfg=dict(type='MMSyncBN')).cuda()
+    conv.eval()
+    y_mmsyncbn = conv(x).detach().cpu().numpy()
+    conv = revert_sync_batchnorm(conv)
+    y_bn = conv(x).detach().cpu().numpy()
+    assert np.all(np.isclose(y_bn, y_mmsyncbn, 1e-3))
+    conv, x = conv.to('cpu'), x.to('cpu')
+    y_bn_cpu = conv(x).detach().numpy()
+    assert np.all(np.isclose(y_bn, y_bn_cpu, 1e-3))