Merge branch 'master' of https://github.com/NVIDIA/apex

README.md

+# PSA:  Amp 1.0 API coming soon!  
+(as introduced by https://info.nvidia.com/webinar-mixed-precision-with-pytorch-reg-page.html.  The `amp` and `FP16_Optimizer` tools currently in master are separate prototypes, which will be unified by the Amp 1.0 API.)
+
+Branch `api_refactor` is tracking my progress.  I will merge to master, along with documentation and examples, by the end of February.
+
 # Introduction
 
 This repository holds NVIDIA-maintained utilities to streamline 

apex/optimizers/fp16_optimizer.py

@@ -2,17 +2,27 @@ import torch
 from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
 import ctypes
 
-lib = ctypes.cdll.LoadLibrary(None)
-lib.THCudaHalfTensor_normall.argtypes=[ctypes.c_void_p, ctypes.c_void_p]
-lib.THCudaHalfTensor_normall.restype = ctypes.c_float
-
-def fused_norm(input):
-    if input.type() == 'torch.cuda.HalfTensor':
-        # 16384 is half 2 if you stare at it long enough
-        return lib.THCudaHalfTensor_normall(torch.cuda._state_cdata,
-            input._cdata, 16384)
-    else:
-        return input.norm()
+stashed_err = None
+try:
+    lib = ctypes.cdll.LoadLibrary(None)
+    lib.THCudaHalfTensor_normall.argtypes=[ctypes.c_void_p, ctypes.c_void_p]
+    lib.THCudaHalfTensor_normall.restype = ctypes.c_float
+
+    def fused_norm(input):
+        if input.type() == 'torch.cuda.HalfTensor':
+            # 16384 is half 2 if you stare at it long enough
+            return lib.THCudaHalfTensor_normall(torch.cuda._state_cdata,
+                input._cdata, 16384)
+        else:
+            return input.norm()
+except TypeError as err:
+    stashed_err = err
+    def fused_norm(input):
+        raise RuntimeError("Failed to create fused_norm. This may happen on Windows "
+              "because of lib = ctypes.cdll.LoadLibrary(None):  you can't "
+              "LoadLibrary with None.  Original exception message was ",
+              stashed_err)
+
 
 class FP16_Optimizer(object):
     """

apex/parallel/__init__.py

@@ -51,3 +51,42 @@ def convert_syncbn_model(module, process_group=None, channel_last=False):
     # TODO(jie) should I delete model explicitly?
     del module
     return mod
+
+def create_syncbn_process_group(group_size):
+    '''
+    Creates process groups to be used for syncbn of a give ``group_size`` and returns
+    process group that current GPU participates in.
+
+    ``group_size`` must divide the total number of GPUs (world_size).
+
+    ``group_size`` of 0 would be considered as =world_size. In this case ``None`` will be returned.
+
+    ``group_size`` of 1 would be equivalent to using non-sync bn, but will still carry the overhead.
+
+    Args:
+        group_size (int): number of GPU's to collaborate for sync bn
+
+    Example::
+
+        >>> # model is an instance of torch.nn.Module
+        >>> import apex
+        >>> group = apex.parallel.create_syncbn_process_group(group_size)
+    '''
+
+    if group_size==0:
+        return None
+
+    world_size = torch.distributed.get_world_size()
+    assert(world_size >= group_size)
+    assert(world_size % group_size == 0)
+
+    group=None
+    for group_num in (range(world_size//group_size)):
+        group_ids = range(group_num*group_size, (group_num+1)*group_size)
+        cur_group = torch.distributed.new_group(ranks=group_ids)
+        if (torch.distributed.get_rank()//group_size == group_num):
+            group = cur_group
+            #can not drop out and return here, every process must go through creation of all subgroups
+
+    assert(group is not None)
+    return group

apex/parallel/optimized_sync_batchnorm.py

@@ -55,7 +55,7 @@ class SyncBatchNorm(_BatchNorm):
         >>> inp = torch.randn(10, 14, 14, 100).cuda()
     """
 
-    def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True, process_group=None, channel_last = False):
+    def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True, process_group=None, channel_last=False):
         super(SyncBatchNorm, self).__init__(num_features, eps=eps, momentum=momentum, affine=affine, track_running_stats=track_running_stats)
         self.process_group = process_group
         self.channel_last = channel_last

apex/parallel/sync_batchnorm.py

@@ -48,7 +48,9 @@ class SyncBatchNorm(_BatchNorm):
 
     warned = False
 
-    def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True, process_group=None):
+    def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True, process_group=None, channel_last=False):
+        if channel_last == True:
+            raise AttributeError("channel_last is not supported by primitive SyncBatchNorm implementation. Try install apex with `--cuda_ext` if channel_last is desired.")
 
         if not SyncBatchNorm.warned:
             print("Warning:  using Python fallback for SyncBatchNorm, possibly because apex was installed without --cuda_ext.  The exception raised when attempting to import the cuda backend was: ", self.syncbn_import_error)

tests/synced_batchnorm/test_groups.py

+import torch
+import numpy as np
+import apex
+import syncbn
+import os
+import argparse
+import torch.optim as optim
+
+def compare(desc, inp1, inp2, error):
+    a = inp1.clone().detach().cpu().numpy()
+    b = inp2.clone().detach().cpu().numpy()
+    close = np.allclose(a,b, error, error)
+    if not close:
+        print(desc, close)
+        z = a - b
+        index = (np.abs(z) >= error + error * np.abs(b)).nonzero()
+        print("dif    : ", z[index])
+        print("inp1   : ", a[index])
+        print("inp2   : ", b[index])
+    return close
+
+feature_size = 10
+space_size = 40
+batch_size = 32
+
+
+from apex.parallel import DistributedDataParallel as DDP
+parser = argparse.ArgumentParser()
+parser.add_argument("--local_rank", default=0, type=int)
+parser.add_argument("--fp16", action='store_true', default=False)
+parser.add_argument("--fp64", action='store_true', default=False)
+parser.add_argument("--group_size", default=0, type=int)
+args = parser.parse_args()
+
+try:
+    args.world_size = int(os.environ['WORLD_SIZE'])
+except:
+    print("This is a multi-gpu test. To run it please use 'python -m torch.distributed.launch --nproc_per_node=<num gpus> test_groups.py <more options>'")
+    exit(1)
+
+torch.cuda.set_device(args.local_rank)
+torch.distributed.init_process_group(backend='nccl', init_method='env://')
+
+start = (args.local_rank%args.group_size) * batch_size//args.group_size
+finish = (args.local_rank%args.group_size + 1) * batch_size//args.group_size
+
+error = 1e-5
+dtype = np.float32
+if args.fp16:
+    error = 1e-3
+    dtype = np.float16
+elif args.fp64:
+    error = 1e-8
+    dtype = np.float64
+
+
+np.random.seed(18 + args.local_rank//args.group_size)
+
+inp = np.random.randn(batch_size, feature_size, space_size, space_size).astype(dtype)
+grad = np.random.randn(batch_size, feature_size, space_size, space_size).astype(dtype)
+weight = np.random.randn(feature_size).astype(dtype)
+bias = np.random.randn(feature_size).astype(dtype)
+
+
+type_tensor = torch.cuda.FloatTensor
+if args.fp16:
+    type_tensor = torch.cuda.HalfTensor
+if args.fp64:
+    type_tensor = torch.cuda.DoubleTensor
+
+ref_tensor = torch.cuda.DoubleTensor
+
+inp_t = type_tensor(inp)
+weight_t = type_tensor(weight)
+bias_t = type_tensor(bias)
+
+inp_r = ref_tensor(inp.transpose(1, 0, 2, 3).reshape(feature_size, -1))
+inp2_r = ref_tensor(inp)
+weight_r = ref_tensor(weight).view(-1, 1, 1)
+bias_r = ref_tensor(bias).view(-1, 1, 1)
+
+grad_output_t = type_tensor(grad)
+
+m = inp_r.mean(1)
+b_v = inp_r.var(1, unbiased=False)
+unb_v = inp_r.var(1, unbiased=True)
+
+eps = 1e-5
+
+mean, var_biased = syncbn.welford_mean_var(inp_t)
+inv_std = 1.0 / torch.sqrt(var_biased + eps)
+
+bn = torch.nn.BatchNorm2d(feature_size).cuda()
+bn.momentum = 1.0
+bn.weight.data = weight_t.clone()
+bn.bias.data = bias_t.clone()
+if args.fp16:
+    bn.half()
+if args.fp64:
+    bn.double()
+bn = DDP(bn)
+inp_bn = inp_t.clone().requires_grad_()
+grad_bn = grad_output_t.clone().detach()
+out_bn = bn(inp_bn)
+out_bn.backward(grad_bn)
+# compensating the averaging over processes done by DDP
+# in order to produce mathematically equivalent result
+# https://github.com/NVIDIA/apex/issues/134#issuecomment-458307368
+for param in bn.parameters():
+    param.grad = param.grad / args.group_size
+bn_opt = optim.SGD(bn.parameters(), lr=1.0)
+
+sbn = apex.parallel.SyncBatchNorm(feature_size, process_group=apex.parallel.create_syncbn_process_group(args.group_size)).cuda()
+sbn.momentum = 1.0
+sbn.weight.data = weight_t.clone()
+sbn.bias.data = bias_t.clone()
+if args.fp16:
+    sbn.half()
+if args.fp64:
+    sbn.double()
+sbn = DDP(sbn)
+sbn_opt = optim.SGD(sbn.parameters(), lr=1.0)
+inp_sbn = inp_t.clone().requires_grad_()
+grad_sbn = grad_output_t.clone().detach()
+out_sbn = sbn(inp_sbn[start:finish])
+out_sbn.backward(grad_sbn[start:finish])
+
+sbn_result = True
+bn_result = True
+
+if args.local_rank == 0:
+    sbn_result = compare("comparing mean: ", mean, m, error) and sbn_result
+    sbn_result = compare("comparing biased variance: ", var_biased, b_v, error) and sbn_result
+
+out = syncbn.batchnorm_forward(inp_t, mean, inv_std, weight_t, bias_t)
+out_r = weight_r * (inp2_r - m.view(-1, 1, 1)) * torch.rsqrt(b_v.view(-1,1,1) + eps) + bias_r
+
+if args.local_rank == 0:
+    sbn_result = compare("comparing output: ", out, out_r, error) and sbn_result
+    compare("comparing bn output: ", out_bn, out_r, error)
+
+grad_output_t = type_tensor(grad)
+
+grad_output_r = ref_tensor(grad.transpose(1, 0, 2, 3).reshape(feature_size, -1))
+grad_output2_r = ref_tensor(grad)
+
+grad_bias_r = grad_output_r.sum(1)
+grad_weight_r = ((inp2_r - m.view(-1, 1, 1)) * torch.rsqrt(b_v.view(-1,1,1) + eps) * grad_output2_r).transpose(1,0).contiguous().view(feature_size, -1).sum(1)
+
+mean_dy_r = grad_output_r.mean(1)
+mean_dy_xmu_r = ((inp2_r - m.view(-1, 1, 1)) * grad_output2_r).transpose(1,0).contiguous().view(feature_size, -1).mean(1)
+
+grad_input_r = (grad_output2_r - mean_dy_r.view(-1, 1, 1) - (inp2_r - m.view(-1, 1, 1)) / (b_v.view(-1,1,1) + eps) * mean_dy_xmu_r.view(-1, 1, 1) ) * torch.rsqrt(b_v.view(-1,1,1) + eps) * weight_r.view(-1,1,1)
+
+mean_dy, mean_dy_xmu, grad_weight, grad_bias = syncbn.reduce_bn(grad_output_t, inp_t, mean, inv_std, weight_t)
+grad_input = syncbn.batchnorm_backward(grad_output_t, inp_t, mean, inv_std, weight_t, mean_dy, mean_dy_xmu)
+
+if args.local_rank == 0:
+    sbn_result = compare("comparing bias grad: ", grad_bias, grad_bias_r, error) and sbn_result
+    sbn_result = compare("comparing weight grad: ", grad_weight, grad_weight_r, error) and sbn_result
+    sbn_result = compare("comparing mean_dy grad: ", mean_dy, mean_dy_r, error) and sbn_result
+    sbn_result = compare("comparing mean_dy_xmu grad: ", mean_dy_xmu, mean_dy_xmu_r, error) and sbn_result
+    sbn_result = compare("comparing input grad: ", grad_input, grad_input_r, error) and sbn_result
+    compare("comparing bn input grad: ", inp_bn.grad, grad_input_r, error)
+
+if args.local_rank == 0:
+    sbn_result = compare("comparing running_mean: ", bn.module.running_mean.data, sbn.module.running_mean.data, error) and sbn_result
+    sbn_result = compare("comparing running_variance: ", bn.module.running_var.data, sbn.module.running_var.data, error) and sbn_result
+
+# execute by both
+compare("comparing layers output: ", out_bn[start:finish], out_sbn, error) and sbn_result
+compare("comparing layers grad_input: ", inp_bn.grad[start:finish], inp_sbn.grad[start:finish], error) and sbn_result
+
+bn_opt.step()
+sbn_opt.step()
+
+if args.local_rank == 0:
+    compare("comparing bn vs sbn bias: ", bn.module.bias, sbn.module.bias, error)
+    compare("comparing bn vs sbn weight: ", bn.module.weight, sbn.module.weight, error)
+
+
+if sbn_result:
+    print("====SBN group test passed")
+else:
+    print("*SBN group test failed*")

tests/synced_batchnorm/unit_test.sh

 python single_gpu_unit_test.py
 python -m torch.distributed.launch --nproc_per_node=2 two_gpu_unit_test.py
 python -m torch.distributed.launch --nproc_per_node=2 two_gpu_unit_test.py --fp64
+#beware, you need a system with at least 4 gpus to test group_size<world_size
+python -m torch.distributed.launch --nproc_per_node=4 test_groups.py --group_size=2