def get_model(text_proc, args):
sent_vocab = text_proc.vocab
model = ActionPropDenseCap(d_model=args.d_model,
d_hidden=args.d_hidden,
n_layers=args.n_layers,
n_heads=args.n_heads,
vocab=sent_vocab,
in_emb_dropout=args.in_emb_dropout,
attn_dropout=args.attn_dropout,
vis_emb_dropout=args.vis_emb_dropout,
cap_dropout=args.cap_dropout,
nsamples=0,
kernel_list=args.kernel_list,
stride_factor=args.stride_factor,
learn_mask=args.learn_mask)
# Initialize the networks and the criterion
if len(args.start_from) > 0:
print("Initializing weights from {}".format(args.start_from))
model.load_state_dict(
torch.load(args.start_from,
map_location=lambda storage, location: storage))
# Ship the model to GPU, maybe
if args.cuda:
model.cuda()
return model
def get_model(text_proc, args):
sent_vocab = text_proc.vocab
model = ActionPropDenseCap(d_model=args.d_model,
d_hidden=args.d_hidden,
n_layers=args.n_layers,
n_heads=args.n_heads,
vocab=sent_vocab,
in_emb_dropout=args.in_emb_dropout,
attn_dropout=args.attn_dropout,
vis_emb_dropout=args.vis_emb_dropout,
cap_dropout=args.cap_dropout,
nsamples=args.train_sample,
kernel_list=args.kernel_list,
stride_factor=args.stride_factor,
learn_mask=args.mask_weight > 0)
# Initialize the networks and the criterion
if len(args.start_from) > 0:
print("Initializing weights from {}".format(args.start_from))
model.load_state_dict(
torch.load(args.start_from,
map_location=lambda storage, location: storage))
# Ship the model to GPU, maybe
if args.cuda:
if args.distributed:
model.cuda()
model = torch.nn.parallel.DistributedDataParallel(model)
else:
model = torch.nn.DataParallel(model).cuda()
# elif torch.cuda.device_count() > 1:
# model = torch.nn.DataParallel(model).cuda()
# else:
# model.cuda()
return model
def get_model(text_proc, args):
sent_vocab = text_proc.vocab # 字典
model = ActionPropDenseCap(
d_model=args.d_model,
d_hidden=args.d_hidden,
n_layers=args.n_layers,
n_heads=args.n_heads,
vocab=sent_vocab, # 字典对象
in_emb_dropout=args.in_emb_dropout, # 0.1
attn_dropout=args.attn_dropout, # 0.2
vis_emb_dropout=args.vis_emb_dropout, # 0.1
cap_dropout=args.cap_dropout, # 0.2
nsamples=args.train_sample, # 20
kernel_list=args.kernel_list,
stride_factor=args.stride_factor,
learn_mask=args.mask_weight > 0)
# Initialize the networks and the criterion
if len(args.start_from) > 0:
print("Initializing weights from {}".format(args.start_from))
model.load_state_dict(
torch.load(args.start_from,
map_location=lambda storage, location: storage))
# Ship the model to GPU, maybe
if args.cuda:
if args.distributed:
model.cuda()
"""
在多机多卡情况下分布式训练数据的读取也是一个问题，不同的卡读取到的数据应该是不同的。
dataparallel(单机多卡)的做法是直接将batch切分到不同的卡，这种方法对于多机来说不可取，因为多
机之间直接进行数据传输会严重影响效率。于是有了利用sampler确保dataloader只会load到
整个数据集的一个特定子集的做法。DistributedSampler(多机多卡)就是做这件事的。它为每一个子进程
划分出一部分数据集，以避免不同进程之间数据重复。
"""
model = torch.nn.parallel.DistributedDataParallel(model)
else:
model = torch.nn.DataParallel(model).cuda() # 单机多卡
# elif torch.cuda.device_count() > 1:
# model = torch.nn.DataParallel(model).cuda()
# else:
# model.cuda()
return model