def test_fp16(self):
store = c10d.TCPStore('localhost', self.port, self.rank == 0)
process_group = c10d.ProcessGroupNCCL(store, self.rank,
self.world_size)
gpus = gpus_for_rank(self.world_size)[self.rank]
model = nn.Linear(1, 1, bias=False).cuda(gpus[0]).half()
nn.init.constant_(model.weight, 1)
ddp_model = distributed_c10d._DistributedDataParallelC10d(
model,
device_ids=[gpus[0]],
process_group=process_group,
bucket_cap_mb=1,
)
# Input 2**15, so that the gradients will overflow with a
# world_size of 2, unless we normalize the gradient by the
# world_size before the reduction
input = torch.Tensor([[2**15]]).cuda(gpus[0]).half()
# Step model
ddp_model.train()
output = ddp_model(input)
loss = output.sum()
loss.backward()
self.assertFalse(
any(torch.isinf(p.grad).any() for p in ddp_model.parameters()))
def test_allreduce_ops(self):
store = c10d.FileStore(self.file.name)
pg = c10d.ProcessGroupNCCL(store, self.rank, self.world_size)
def allreduce(tensors, op):
opts = c10d.AllreduceOptions()
opts.reduceOp = op
work = pg.allreduce(tensors, opts)
work.wait()
# Sum
tensors = []
for i in range(self.num_gpus):
tensors.append(torch.Tensor([i + 1]).cuda(i))
allreduce(tensors, c10d.ReduceOp.SUM)
for i in range(self.num_gpus):
self.assertEqual(
torch.Tensor([float(self.num_gpus * (self.num_gpus + 1) / 2)]),
tensors[i])
# Product
tensors = []
for i in range(self.num_gpus):
tensors.append(torch.Tensor([i + 1]).cuda(i))
allreduce(tensors, c10d.ReduceOp.PRODUCT)
for i in range(self.num_gpus):
self.assertEqual(
torch.Tensor([float(math.factorial(self.num_gpus))]),
tensors[i])
# Min
tensors = []
for i in range(self.num_gpus):
tensors.append(torch.Tensor([i + 1]).cuda(i))
allreduce(tensors, c10d.ReduceOp.MIN)
for i in range(self.num_gpus):
self.assertEqual(torch.Tensor([1.0]), tensors[i])
# Max
tensors = []
for i in range(self.num_gpus):
tensors.append(torch.Tensor([i + 1]).cuda(i))
allreduce(tensors, c10d.ReduceOp.MAX)
for i in range(self.num_gpus):
self.assertEqual(torch.Tensor([self.num_gpus]), tensors[i])
def test_broadcast_ops(self):
store = c10d.FileStore(self.file.name)
pg = c10d.ProcessGroupNCCL(store, self.rank, self.world_size)
def broadcast(xs, rootRank, rootTensor):
opts = c10d.BroadcastOptions()
opts.rootRank = rootRank
opts.rootTensor = rootTensor
work = pg.broadcast(xs, opts)
work.wait()
# for every root tensor
for rt in range(self.num_gpus):
tensors = []
for i in range(self.num_gpus):
tensors.append(torch.Tensor([i]).cuda(i))
broadcast(tensors, self.rank, rt)
for i in range(self.num_gpus):
self.assertEqual(tensors[i], tensors[rt])
def test_reduce_ops(self):
store = c10d.FileStore(self.file.name)
pg = c10d.ProcessGroupNCCL(store, self.rank, self.world_size)
def reduce(xs, rootRank, rootTensor):
opts = c10d.ReduceOptions()
opts.rootRank = rootRank
opts.rootTensor = rootTensor
work = pg.reduce(xs, opts)
work.wait()
# for every root tensor
for rt in range(self.num_gpus):
tensors = []
for i in range(self.num_gpus):
tensors.append(torch.Tensor([i + 1]).cuda(i))
reduce(tensors, self.rank, rt)
self.assertEqual(
torch.Tensor([float(self.num_gpus * (self.num_gpus + 1) / 2)]),
tensors[rt])
def test_allgather_ops(self):
store = c10d.FileStore(self.file.name)
pg = c10d.ProcessGroupNCCL(store, self.rank, self.world_size)
def allgather(output_ts, input_ts):
work = pg.allgather(output_ts, input_ts)
work.wait()
tensors = []
output_ts = [[] for _ in range(self.num_gpus)]
for idx, ls in enumerate(output_ts):
for _ in range(self.world_size * self.num_gpus):
ls.append(torch.Tensor([0]).cuda(idx))
for i in range(self.num_gpus):
tensors.append(torch.Tensor([i]).cuda(i))
allgather(output_ts, tensors)
# Verification
for device_ts in output_ts:
for s_idx, t in enumerate(device_ts):
self.assertEqual(torch.Tensor([s_idx]), t)
def main():
# is_chief indicates this machine will do shared tasks for the cluster
# such as logging and checkpointing
# is_chief must be true only for at most 1 process in training cluster
# $RANK is set by pytorch.distributed.launch
# https://github.com/pytorch/pytorch/blob/db6e4576dab097abf01d032c3326e4b285eb8499/torch/distributed/launch.py#L193
global is_chief, event_writer, global_example_count, last_recv_bytes, last_transmit_bytes, last_log_time
is_chief = (not args.distributed) or (int(os.environ['RANK'])==0)
global_example_count = 0
if is_chief:
print(f"Logging to {args.logdir}")
event_writer = SummaryWriter(args.logdir)
log_tb("first", time.time())
else:
event_writer = NoOp()
# baseline number for network bytes
last_recv_bytes, last_transmit_bytes = network_bytes()
last_log_time = time.time()
print(args)
print("~~epoch\thours\ttop1Accuracy\n")
# need to index validation directory before we start counting the time
dataloader.sort_ar(args.data+'/validation')
global reduce_function
if args.c10d:
print('Distributed: loading c10d process group')
# https://github.com/pytorch/pytorch/blob/master/torch/lib/c10d/TCPStore.hpp
torch.cuda.set_device(args.local_rank)
rank = int(os.environ['RANK'])
store = c10d.TCPStore(os.environ['MASTER_ADDR'], int(os.environ['MASTER_PORT']), rank==0) # (masterAddr, masterPort, isServer)
process_group = c10d.ProcessGroupNCCL(store, rank, args.world_size) # (store, rank, size)
reduce_function = lambda t: process_group.allreduce(t, c10d.AllreduceOptions().reduceOp)
elif args.distributed:
print('Distributed: initializing process group')
torch.cuda.set_device(args.local_rank)
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=args.world_size)
assert(args.world_size == dist.get_world_size())
reduce_function = lambda t: dist.all_reduce(t, op=dist.reduce_op.SUM)
print("Distributed: success (%d/%d)"%(args.local_rank, args.world_size))
if args.fp16: assert torch.backends.cudnn.enabled, "fp16 mode requires cudnn backend to be enabled."
print("Loading model")
if args.factorized_resnet: model = resnet.resnet50factorized(pretrained=args.pretrained)
else: model = resnet.resnet50(pretrained=args.pretrained)
model = model.cuda()
if args.init_bn0: resnet.init_dist_weights(model) # Sets batchnorm std to 0
if args.fp16: model = network_to_half(model)
best_prec5 = 93 # only save models over 92%. Otherwise it stops to save every time
# Load model from checkpoint. This must happen distributed as model is saved without it
if args.resume:
checkpoint = torch.load(args.resume, map_location = lambda storage, loc: storage.cuda(args.local_rank))
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch']
best_prec5 = checkpoint['best_prec5']
if args.c10d:
model = distributed_c10d._DistributedDataParallelC10d(model, process_group, device_ids=[args.local_rank], output_device=args.local_rank)
c10d_sanity_check()
elif args.distributed: model = nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank)
global model_params, master_params
if args.fp16: model_params, master_params = prep_param_lists(model)
else: model_params = master_params = model.parameters()
optim_params = experimental_utils.bnwd_optim_params(model, model_params, master_params) if args.no_bn_wd else master_params
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(optim_params, 0, momentum=args.momentum, weight_decay=args.weight_decay) # start with 0 lr. Scheduler will change this later
if args.resume: # we must resume optimizer params separately
checkpoint = torch.load(args.resume, map_location = lambda storage, loc: storage.cuda(args.local_rank))
optimizer.load_state_dict(checkpoint['optimizer'])
# Load data data manager and lr scheduler from phases
phases = eval(args.phases)
print("Creating data loaders (this could take 6-12 minutes)")
dm = DataManager([p for p in phases if 'bs' in p])
scheduler = Scheduler(optimizer, [p for p in phases if 'lr' in p], args.scale_lr)
start_time = datetime.now() # Loading start to after everything is loaded
if args.evaluate: return validate(dm.val_dl, model, criterion, 0, start_time)
if args.distributed:
print('Syncing machines before training')
sum_tensor(torch.tensor([1.0]).float().cuda())
print("Begin training")
estart = time.time()
for epoch in range(args.start_epoch, scheduler.tot_epochs):
estart = time.time()
dm.set_epoch(epoch)
train(dm.trn_dl, model, criterion, optimizer, scheduler, epoch)
if args.prof: break
prec5 = validate(dm.val_dl, model, criterion, epoch, start_time)
is_best = prec5 > best_prec5
best_prec5 = max(prec5, best_prec5)
if args.local_rank == 0:
if is_best: save_checkpoint(epoch, model, best_prec5, optimizer, is_best=True, filename='model_best.pth.tar')
phase = dm.get_phase(epoch)
if phase:save_checkpoint(epoch, model, best_prec5, optimizer, filename=f'sz{phase["bs"]}_checkpoint.path.tar')
event_writer.export_scalars_to_json(args.logdir+'/scalars.json')
event_writer.close()
def test_nccl_backend(self):
store = c10d.TCPStore('localhost', self.port, self.is_master)
process_group = c10d.ProcessGroupNCCL(store, self.rank, self.world_size)
self._test_ddp_with_process_group(process_group)
