def main():
data_holder, task2id, id2task, num_feat, num_voc, num_char, tgt_dict, embeddings = Dataloader_elmo1.multitask_dataloader(
pkl_path, num_task=num_task, batch_size=BATCH_SIZE)
para = model_para
task2label = {"conll2000": "chunk", "unidep": "POS", "conll2003": "NER"}
#task2label = {"conll2000": "chunk", "wsjpos": "POS", "conll2003": "NER"}
#logger = Logger('./logs/'+str(args.gpu))
para["id2task"] = id2task
para["n_feats"] = num_feat
para["n_vocs"] = num_voc
para["n_tasks"] = num_task
para["out_size"] = [
len(tgt_dict[task2label[id2task[ids]]]) for ids in range(num_task)
]
para["n_chars"] = num_char
model = Model_s.build_model_cnn(para)
model.Word_embeddings.apply_weights(embeddings)
params = list(filter(lambda p: p.requires_grad, model.parameters()))
num_params = sum(p.numel() for p in model.parameters())
print(model)
print("Num of paras:", num_params)
print(model.concat_flag)
def lr_decay(optimizer, epoch, decay_rate=0.05, init_lr=0.015):
lr = init_lr / (1 + decay_rate * epoch)
print(" Learning rate is set as:", lr)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
return optimizer
def exp_lr_decay(optimizer, epoch, decay_rate=0.05, init_lr=0.015):
lr = init_lr * decay_rate**epoch
print(" Learning rate is set as:", lr)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
return optimizer
if args.optim == "noam":
model_optim = optim_custorm.NoamOpt(
para["d_hid"], 1, 1000,
torch.optim.Adam(params,
lr=0.0015,
betas=(0.9, 0.98),
eps=1e-9,
weight_decay=L2))
args.decay = None
elif args.optim == "sgd":
model_optim = optim.SGD(params,
lr=0.015,
momentum=args.momentum,
weight_decay=1e-8)
elif args.optim == "adam":
model_optim = optim.Adam(params,
lr=0.0,
betas=(0.9, 0.98),
eps=1e-9,
weight_decay=1e-8)
if args.mode == "train":
best_F1 = 0
if not para["crf"]:
calculate_loss = nn.NLLLoss()
else:
calculate_loss = None
#calculate_loss = [CRFLoss_vb(len(tgt_dict[task2label[id2task[idx]]])+2, len(tgt_dict[task2label[id2task[idx]]]), len(tgt_dict[task2label[id2task[idx]]])+1) for idx in range(num_task)]
#if USE_CUDA:
# for x in calculate_loss:
# x = x.cuda()
print("Start training...")
print('-' * 60)
KLLoss = None #nn.KLDivLoss()
start_point = time.time()
for epoch_idx in range(NUM_EPOCH):
if args.optim == "sgd":
if args.decay == "exp":
model_optim = exp_lr_decay(model_optim, epoch_idx)
elif args.decay == "normal":
model_optim = lr_decay(model_optim, epoch_idx)
Pre, Rec, F1, loss_list = run_epoch(model, data_holder,
model_optim, calculate_loss,
KLLoss, para, epoch_idx,
id2task)
use_time = time.time() - start_point
print("Time using: %f mins" % (use_time / 60))
if not best_F1 or best_F1 < F1:
best_F1 = F1
Model_s.save_model(model_path, model, para)
print('*' * 60)
print(
"Save model with average Pre: %f, Rec: %f, F1: %f on dev set."
% (Pre, Rec, F1))
save_idx = epoch_idx
print('*' * 60)
print("save model at epoch:", save_idx)
else:
para_path = os.path.join(path, 'para.pkl')
with open(para_path, "wb") as f:
para_save = pickle.load(f)
model = Model_s.build_model(para_save)
model = Model_s.read_model(model_path, model)
prec_list, rec_list, f1_list = infer(model, data_holder, "test")
def main():
data_holder, task2id, id2task, num_feat, num_voc, num_voc, tgt_dict, embeddings = DataLoader.multitask_dataloader(
pkl_path, num_task=num_task, batch_size=BATCH_SIZE)
para = model_para
task2label = {"conll2000": "chunk", "unidep": "POS", "conll2003": "NER"}
#task2label = {"conll2000": "chunk", "wsjpos": "POS", "conll2003": "NER"}
#logger = Logger('./logs/'+str(args.gpu))
para["id2task"] = id2task
para["n_feats"] = num_feat
para["n_vocs"] = num_voc
para["n_tasks"] = num_task
para["out_size"] = [
len(tgt_dict[task2label[id2task[ids]]]) for ids in range(num_task)
]
model = Model.build_model(para)
model.Word_embeddings.apply_weights(embeddings)
params = model.parameters()
num_params = sum(p.numel() for p in model.parameters())
print(model)
print("Num of paras:", num_params)
print(model.concat_flag)
model_optim = optim_custorm.NoamOpt(
para["d_hid"], 1, 1000,
torch.optim.Adam(params,
lr=0.0,
betas=(0.9, 0.98),
eps=1e-9,
weight_decay=L2))
#model_optim = optim_custorm.NoamOpt(para["d_hid"], 1, 1000, torch.optim.SGD(params, lr=0.001, momentum=0.9))
if args.mode == "train":
best_F1 = 0
if not para["crf"]:
calculate_loss = nn.NLLLoss()
else:
calculate_loss = None
print("Start training...")
print('-' * 60)
KLLoss = nn.KLDivLoss()
start_point = time.time()
for epoch_idx in range(NUM_EPOCH):
Pre, Rec, F1 = run_epoch(model, data_holder, model_optim,
calculate_loss, KLLoss, para, epoch_idx,
id2task)
use_time = time.time() - start_point
print("Time using: %f mins" % (use_time / 60))
if not best_F1 or best_F1 < F1:
best_F1 = F1
Model.save_model(model_path, model, para)
print('*' * 60)
print(
"Save model with average Pre: %f, Rec: %f, F1: %f on dev set."
% (Pre, Rec, F1))
save_idx = epoch_idx
print('*' * 60)
print("save model at epoch:", save_idx)
else:
para_path = os.path.join(path, 'para.pkl')
with open(para_path, "wb") as f:
para_save = pickle.load(f)
model = Model.build_model(para_save)
model = Model.read_model(model_path, model)
prec_list, rec_list, f1_list = infer(model, data_holder, "test")
def main():
critorion = loss_custorm.loss_fuc(nn.NLLLoss, ignore_index=0)
model = Model.build_san_model(args.d_emb, args.d_hid, args.n_layers, args.dropout, n_voc, args.beam_num)
#logger = Logger('./logs/'+args.gpu)
if args.mode == "train":
params = model.parameters()
model_optim = optim_custorm.NoamOpt(args.d_hid, args.factor, args.warm, torch.optim.Adam(params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=args.L2))
model.embeddings.apply_weights(weight)
print("Begin training...")
start_time = time.time()
best_loss = 0
for epoch_idx in range(args.max_epoch):
val_loss, train_loss, step = run_epoch(model, critorion, model_optim, epoch_idx)#, logger)
print('-' * 70)
print('| val_loss: %4.4f | train_loss: %4.4f' %(val_loss, train_loss))
print('-' * 70)
if not best_loss or best_loss > val_loss:
best_loss = val_loss
is_best = True
#Model.save_model(args.model_path+args.gpu+"/", model)
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_prec1': val_loss,
'optimizer' : model_optim.optimizer.state_dict(),
'step' : step
}, is_best)
#if epoch_idx % 5 == 0 or not epoch_idx:
predict(model, epoch_idx)#, logger)
if args.mode == "resume":
path_save = args.model_path+args.gpu+"/"+"checkpoint.pth.tar"
if os.path.isfile(path_save):
print("=> loading checkpoint '{}'".format(path_save))
checkpoint = torch.load(path_save)
start_epoch = checkpoint['epoch']
best_loss = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
step = checkpoint['step']
print("=> loaded checkpoint '{}' (epoch {})"
.format(path_save, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
print("Continue training from %d epoch." %epoch)
model_optim = optim_custorm.NoamOpt(args.d_hid, args.factor, args.warm, optimizer, step=step)
for epoch_idx in range(start_epoch, args.max_epoch):
val_loss, train_loss, step = run_epoch(model, critorion, model_optim, epoch_idx)#, logger)
print('-' * 70)
print('| val_loss: %4.4f | train_loss: %4.4f' %(val_loss, train_loss))
print('-' * 70)
if not best_loss or best_loss > val_loss:
best_loss = val_loss
is_best = True
#Model.save_model(args.model_path+args.gpu+"/", model)
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_prec1': val_loss,
'optimizer' : model_optim.optimizer.state_dict(),
'step' : step
}, is_best)
#if epoch_idx % 5 == 0 or not epoch_idx:
predict(model, epoch_idx)#, logger)
else:
model = Model.read_model(args.model_path, model)
save_hyp = pridict(model)
