def testHogwildStability_RowAdagrad(self):
NE = 10000
model = nn.Embedding(NE, 100)
optimizer = RowAdagrad(model.parameters())
num_processes = mp.cpu_count() // 2 + 1
self._stress_optimizer(model, optimizer, num_processes=num_processes)
# This fails for Adagrad because it's not stable
self.assertLess(model.weight.abs().max(), 1000)
def make_optimizer(params: Iterable[torch.nn.Parameter], is_emb: bool) -> Optimizer:
params = list(params)
if len(params) == 0:
optimizer = DummyOptimizer()
elif is_emb:
optimizer = RowAdagrad(params, lr=config.lr)
else:
if config.relation_lr is not None:
lr = config.relation_lr
else:
lr = config.lr
optimizer = Adagrad(params, lr=lr)
optimizer.share_memory()
return optimizer
def do_one_job( # noqa
self,
lhs_types: Set[str],
rhs_types: Set[str],
lhs_part: Partition,
rhs_part: Partition,
lhs_subpart: SubPartition,
rhs_subpart: SubPartition,
next_lhs_subpart: Optional[SubPartition],
next_rhs_subpart: Optional[SubPartition],
model: MultiRelationEmbedder,
trainer: Trainer,
all_embs: Dict[Tuple[EntityName, Partition], FloatTensorType],
subpart_slices: Dict[Tuple[EntityName, Partition, SubPartition],
slice],
subbuckets: Dict[Tuple[int, int], Tuple[LongTensorType, LongTensorType,
LongTensorType]],
batch_size: int,
lr: float,
) -> Stats:
tk = TimeKeeper()
for embeddings in all_embs.values():
assert embeddings.is_pinned()
occurrences: Dict[Tuple[EntityName, Partition, SubPartition],
Set[Side]] = defaultdict(set)
for entity_name in lhs_types:
occurrences[entity_name, lhs_part, lhs_subpart].add(Side.LHS)
for entity_name in rhs_types:
occurrences[entity_name, rhs_part, rhs_subpart].add(Side.RHS)
if lhs_part != rhs_part: # Bipartite
assert all(len(v) == 1 for v in occurrences.values())
tk.start("copy_to_device")
for entity_name, part, subpart in occurrences.keys():
if (entity_name, part, subpart) in self.sub_holder:
continue
embeddings = all_embs[entity_name, part]
optimizer = trainer.partitioned_optimizers[entity_name, part]
subpart_slice = subpart_slices[entity_name, part, subpart]
# TODO have two permanent storages on GPU and move stuff in and out
# from them
# logger.info(f"GPU #{self.gpu_idx} allocating {(subpart_slice.stop - subpart_slice.start) * embeddings.shape[1] * 4:,} bytes")
gpu_embeddings = torch.empty(
(subpart_slice.stop - subpart_slice.start,
embeddings.shape[1]),
dtype=torch.float32,
device=self.my_device,
)
gpu_embeddings.copy_(embeddings[subpart_slice], non_blocking=True)
gpu_embeddings = torch.nn.Parameter(gpu_embeddings)
gpu_optimizer = RowAdagrad([gpu_embeddings], lr=lr)
(cpu_state, ) = optimizer.state.values()
(gpu_state, ) = gpu_optimizer.state.values()
# logger.info(f"GPU #{self.gpu_idx} allocating {(subpart_slice.stop - subpart_slice.start) * 4:,} bytes")
gpu_state["sum"].copy_(cpu_state["sum"][subpart_slice],
non_blocking=True)
self.sub_holder[entity_name, part, subpart] = (
gpu_embeddings,
gpu_optimizer,
)
logger.debug(
f"Time spent copying subparts to GPU: {tk.stop('copy_to_device'):.4f} s"
)
for (
(entity_name, part, subpart),
(gpu_embeddings, gpu_optimizer),
) in self.sub_holder.items():
for side in occurrences[entity_name, part, subpart]:
model.set_embeddings(entity_name, side, gpu_embeddings)
trainer.partitioned_optimizers[entity_name, part,
subpart] = gpu_optimizer
tk.start("translate_edges")
num_edges = subbuckets[lhs_subpart, rhs_subpart][0].shape[0]
edge_perm = torch.randperm(num_edges)
edges_lhs, edges_rhs, edges_rel = subbuckets[lhs_subpart, rhs_subpart]
_C.shuffle(edges_lhs, edge_perm, os.cpu_count())
_C.shuffle(edges_rhs, edge_perm, os.cpu_count())
_C.shuffle(edges_rel, edge_perm, os.cpu_count())
assert edges_lhs.is_pinned()
assert edges_rhs.is_pinned()
assert edges_rel.is_pinned()
gpu_edges = EdgeList(
EntityList.from_tensor(edges_lhs),
EntityList.from_tensor(edges_rhs),
edges_rel,
).to(self.my_device, non_blocking=True)
logger.debug(f"GPU #{self.gpu_idx} got {num_edges} edges")
logger.debug(
f"Time spent copying edges to GPU: {tk.stop('translate_edges'):.4f} s"
)
tk.start("processing")
stats = process_in_batches(batch_size=batch_size,
model=model,
batch_processor=trainer,
edges=gpu_edges)
logger.debug(f"Time spent processing: {tk.stop('processing'):.4f} s")
next_occurrences: Dict[Tuple[EntityName, Partition, SubPartition],
Set[Side]] = defaultdict(set)
if next_lhs_subpart is not None:
for entity_name in lhs_types:
next_occurrences[entity_name, lhs_part,
next_lhs_subpart].add(Side.LHS)
if next_rhs_subpart is not None:
for entity_name in rhs_types:
next_occurrences[entity_name, rhs_part,
next_rhs_subpart].add(Side.RHS)
tk.start("copy_from_device")
for (entity_name, part,
subpart), (gpu_embeddings,
gpu_optimizer) in list(self.sub_holder.items()):
if (entity_name, part, subpart) in next_occurrences:
continue
embeddings = all_embs[entity_name, part]
optimizer = trainer.partitioned_optimizers[entity_name, part]
subpart_slice = subpart_slices[entity_name, part, subpart]
embeddings[subpart_slice].data.copy_(gpu_embeddings.detach(),
non_blocking=True)
del gpu_embeddings
(cpu_state, ) = optimizer.state.values()
(gpu_state, ) = gpu_optimizer.state.values()
cpu_state["sum"][subpart_slice].copy_(gpu_state["sum"],
non_blocking=True)
del gpu_state["sum"]
del self.sub_holder[entity_name, part, subpart]
logger.debug(
f"Time spent copying subparts from GPU: {tk.stop('copy_from_device'):.4f} s"
)
logger.debug(
f"do_one_job: Time unaccounted for: {tk.unaccounted():.4f} s")
return stats
