def main(args, local_rank):
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
vocabs = dict()
vocabs['src'] = Vocab(args.src_vocab, 0, [BOS, EOS])
vocabs['tgt'] = Vocab(args.tgt_vocab, 0, [BOS, EOS])
if args.world_size == 1 or (dist.get_rank() == 0):
logger.info(args)
for name in vocabs:
logger.info("vocab %s, size %d, coverage %.3f", name,
vocabs[name].size, vocabs[name].coverage)
set_seed(19940117)
#device = torch.device('cpu')
torch.cuda.set_device(local_rank)
device = torch.device('cuda', local_rank)
if args.resume_ckpt:
model = MatchingModel.from_pretrained(vocabs, args.resume_ckpt)
else:
model = MatchingModel.from_params(vocabs, args.layers, args.embed_dim,
args.ff_embed_dim, args.num_heads,
args.dropout, args.output_dim,
args.bow)
if args.world_size > 1:
set_seed(19940117 + dist.get_rank())
model = model.to(device)
if args.resume_ckpt:
dev_data = DataLoader(vocabs,
args.dev_data,
args.dev_batch_size,
addition=args.additional_negs)
acc = validate(model, dev_data, device)
logger.info("initialize from %s, initial acc %.2f", args.resume_ckpt,
acc)
optimizer = Adam(model.parameters(),
lr=args.lr,
betas=(0.9, 0.98),
eps=1e-9)
lr_schedule = get_linear_schedule_with_warmup(optimizer, args.warmup_steps,
args.total_train_steps)
train_data = DataLoader(vocabs,
args.train_data,
args.per_gpu_train_batch_size,
worddrop=args.worddrop,
addition=args.additional_negs)
global_step, step, epoch = 0, 0, 0
tr_stat = Statistics()
logger.info("start training")
model.train()
while global_step <= args.total_train_steps:
for batch in train_data:
batch = move_to_device(batch, device)
loss, acc, bsz = model(batch['src_tokens'], batch['tgt_tokens'],
args.label_smoothing)
tr_stat.update({
'loss': loss.item() * bsz,
'nsamples': bsz,
'acc': acc * bsz
})
tr_stat.step()
loss.backward()
step += 1
if not (step % args.gradient_accumulation_steps
== -1 % args.gradient_accumulation_steps):
continue
if args.world_size > 1:
average_gradients(model)
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
lr_schedule.step()
optimizer.zero_grad()
global_step += 1
if args.world_size == 1 or (dist.get_rank() == 0):
if global_step % args.print_every == -1 % args.print_every:
logger.info("epoch %d, step %d, loss %.3f, acc %.3f",
epoch, global_step,
tr_stat['loss'] / tr_stat['nsamples'],
tr_stat['acc'] / tr_stat['nsamples'])
tr_stat = Statistics()
if global_step > args.warmup_steps and global_step % args.eval_every == -1 % args.eval_every:
dev_data = DataLoader(vocabs,
args.dev_data,
args.dev_batch_size,
addition=args.additional_negs)
acc = validate(model, dev_data, device)
logger.info("epoch %d, step %d, dev, dev acc %.2f", epoch,
global_step, acc)
save_path = '%s/epoch%d_batch%d_acc%.2f' % (
args.ckpt, epoch, global_step, acc)
model.save(args, save_path)
model.train()
if global_step > args.total_train_steps:
break
epoch += 1
logger.info('rank %d, finish training after %d steps', local_rank,
global_step)
def run(hparams,
model,
train_dataloader,
valid_dataloader,
device,
out_dir='checkpoints'):
learning_rate = hparams['learning_rate']
accumulate_step = hparams['accumulate_step']
lr_schedule = hparams['lr_schedule']
warmup_steps = hparams['warmup_steps']
warmup_proportion = hparams['warmup_proportion']
n_embd = hparams['n_embd']
num_optim_steps = hparams['num_optim_steps']
train_batch_size = hparams['train_batch_size']
valid_step = hparams['valid_step']
no_token_id = hparams['no_token_id']
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
total_params = sum([np.prod(p.size()) for p in model_parameters])
logger.info('Number of parameter = {}'.format(total_params))
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'ln']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = Adam(optimizer_grouped_parameters,
learning_rate,
max_grad_norm=1.0)
step = 0
global_step = 0
epoch = 0
while True:
model.train()
(tr_loss, tr_ppl, mean_ppl, nb_tr_examples,
nb_tr_steps) = 0.0, 0.0, 0.0, 0, 0
n_token_real, n_token_total = 0, 0
pbar = tqdm.tqdm(enumerate(train_dataloader),
total=len(train_dataloader))
for i, batch in pbar:
batch = tuple(t.cuda() for t in batch)
input_ids, position_ids, token_type_ids, label_ids, *_ = batch
if no_token_id:
token_type_ids = None
loss, ppl = model(input_ids, position_ids, token_type_ids,
label_ids)
loss = loss.mean()
loss = loss / (train_batch_size / input_ids.shape[0])
loss.backward()
nb_tr_steps += 1
tr_loss += float(
loss.sum().item()) * (train_batch_size / input_ids.shape[0])
if ppl.sum().item() < 1000000:
tr_ppl += ppl.sum().item()
else:
tr_ppl += mean_ppl
mean_loss = tr_loss / nb_tr_steps
mean_ppl = tr_ppl / nb_tr_steps
n_token_total += input_ids.shape[0] * input_ids.shape[1]
n_token_real += (input_ids != 0).sum().item()
#gradient update
step += 1
if step % accumulate_step == 0:
set_lr(optimizer, global_step, lr_schedule, learning_rate,
warmup_steps, warmup_proportion, n_embd,
num_optim_steps)
optimizer.step()
optimizer.zero_grad()
global_step += 1
print(
'epoch: {}, global_step: {}, step: {}, mean_loss: {}, mean_ppl:{}'
.format(epoch + 1, global_step + 1, step + 1, mean_loss,
mean_ppl),
file=train_logger)
if global_step % valid_step == 0:
print('Saving model...')
torch.save(
{
'model': model.state_dict(),
'epoch': epoch,
'hparams': hparams,
},
os.path.join(out_dir,
f'GPT2-pretrain-step-{global_step}.pkl'))
eval_loss, eval_ppl = valid(model, valid_dataloader, epoch,
device)
print('{},{},{},{},{}'.format(epoch + 1, global_step + 1,
step + 1, eval_loss,
eval_ppl),
file=valid_logger)
logger.info('current learning rate: ' +
str(optimizer.param_groups[0]['lr']))
model.train()
if global_step >= num_optim_steps:
break
if (step + 1) % CACHE_EMPTY_STEP == 0:
torch.cuda.empty_cache()
if global_step >= num_optim_steps:
break
epoch += 1
train_logger.close()
valid_logger.close()
def main(args, local_rank):
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
vocabs = dict()
vocabs['src'] = Vocab(args.src_vocab, 0, [BOS, EOS])
vocabs['tgt'] = Vocab(args.tgt_vocab, 0, [BOS, EOS])
if args.world_size == 1 or (dist.get_rank() == 0):
logger.info(args)
for name in vocabs:
logger.info("vocab %s, size %d, coverage %.3f", name,
vocabs[name].size, vocabs[name].coverage)
set_seed(19940117)
#device = torch.device('cpu')
torch.cuda.set_device(local_rank)
device = torch.device('cuda', local_rank)
if args.arch == 'vanilla':
model = Generator(vocabs, args.embed_dim, args.ff_embed_dim,
args.num_heads, args.dropout, args.enc_layers,
args.dec_layers, args.label_smoothing)
elif args.arch == 'mem':
model = MemGenerator(vocabs, args.embed_dim, args.ff_embed_dim,
args.num_heads, args.dropout, args.mem_dropout,
args.enc_layers, args.dec_layers,
args.mem_enc_layers, args.label_smoothing,
args.use_mem_score)
elif args.arch == 'rg':
logger.info("start building model")
logger.info("building retriever")
retriever = Retriever.from_pretrained(
args.num_retriever_heads,
vocabs,
args.retriever,
args.nprobe,
args.topk,
local_rank,
use_response_encoder=(args.rebuild_every > 0))
logger.info("building retriever + generator")
model = RetrieverGenerator(vocabs, retriever, args.share_encoder,
args.embed_dim, args.ff_embed_dim,
args.num_heads, args.dropout,
args.mem_dropout, args.enc_layers,
args.dec_layers, args.mem_enc_layers,
args.label_smoothing)
if args.resume_ckpt:
model.load_state_dict(torch.load(args.resume_ckpt)['model'])
else:
global_step = 0
if args.world_size > 1:
set_seed(19940117 + dist.get_rank())
model = model.to(device)
retriever_params = [
v for k, v in model.named_parameters() if k.startswith('retriever.')
]
other_params = [
v for k, v in model.named_parameters()
if not k.startswith('retriever.')
]
optimizer = Adam([{
'params': retriever_params,
'lr': args.embed_dim**-0.5 * 0.1
}, {
'params': other_params,
'lr': args.embed_dim**-0.5
}],
betas=(0.9, 0.98),
eps=1e-9)
lr_schedule = get_inverse_sqrt_schedule_with_warmup(
optimizer, args.warmup_steps, args.total_train_steps)
train_data = DataLoader(vocabs,
args.train_data,
args.per_gpu_train_batch_size,
for_train=True,
rank=local_rank,
num_replica=args.world_size)
model.eval()
#dev_data = DataLoader(vocabs, cur_dev_data, args.dev_batch_size, for_train=False)
#bleu = validate(device, model, dev_data, beam_size=5, alpha=0.6, max_time_step=10)
step, epoch = 0, 0
tr_stat = Statistics()
logger.info("start training")
model.train()
best_dev_bleu = 0.
while global_step <= args.total_train_steps:
for batch in train_data:
#step_start = time.time()
batch = move_to_device(batch, device)
if args.arch == 'rg':
loss, acc = model(
batch,
update_mem_bias=(global_step >
args.update_retriever_after))
else:
loss, acc = model(batch)
tr_stat.update({
'loss': loss.item() * batch['tgt_num_tokens'],
'tokens': batch['tgt_num_tokens'],
'acc': acc
})
tr_stat.step()
loss.backward()
#step_cost = time.time() - step_start
#print ('step_cost', step_cost)
step += 1
if not (step % args.gradient_accumulation_steps
== -1 % args.gradient_accumulation_steps):
continue
if args.world_size > 1:
average_gradients(model)
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
lr_schedule.step()
optimizer.zero_grad()
global_step += 1
if args.world_size == 1 or (dist.get_rank() == 0):
if global_step % args.print_every == -1 % args.print_every:
logger.info("epoch %d, step %d, loss %.3f, acc %.3f",
epoch, global_step,
tr_stat['loss'] / tr_stat['tokens'],
tr_stat['acc'] / tr_stat['tokens'])
tr_stat = Statistics()
if global_step % args.eval_every == -1 % args.eval_every:
model.eval()
max_time_step = 256 if global_step > 2 * args.warmup_steps else 5
bleus = []
for cur_dev_data in args.dev_data:
dev_data = DataLoader(vocabs,
cur_dev_data,
args.dev_batch_size,
for_train=False)
bleu = validate(device,
model,
dev_data,
beam_size=5,
alpha=0.6,
max_time_step=max_time_step)
bleus.append(bleu)
bleu = sum(bleus) / len(bleus)
logger.info("epoch %d, step %d, dev bleu %.2f", epoch,
global_step, bleu)
if bleu > best_dev_bleu:
testbleus = []
for cur_test_data in args.test_data:
test_data = DataLoader(vocabs,
cur_test_data,
args.dev_batch_size,
for_train=False)
testbleu = validate(device,
model,
test_data,
beam_size=5,
alpha=0.6,
max_time_step=max_time_step)
testbleus.append(testbleu)
testbleu = sum(testbleus) / len(testbleus)
logger.info("epoch %d, step %d, test bleu %.2f", epoch,
global_step, testbleu)
torch.save({
'args': args,
'model': model.state_dict()
}, '%s/best.pt' % (args.ckpt, ))
if not args.only_save_best:
torch.save(
{
'args': args,
'model': model.state_dict()
},
'%s/epoch%d_batch%d_devbleu%.2f_testbleu%.2f' %
(args.ckpt, epoch, global_step, bleu,
testbleu))
best_dev_bleu = bleu
model.train()
if args.rebuild_every > 0 and (global_step % args.rebuild_every
== -1 % args.rebuild_every):
model.retriever.drop_index()
torch.cuda.empty_cache()
next_index_dir = '%s/batch%d' % (args.ckpt, global_step)
if args.world_size == 1 or (dist.get_rank() == 0):
model.retriever.rebuild_index(next_index_dir)
dist.barrier()
else:
dist.barrier()
model.retriever.update_index(next_index_dir, args.nprobe)
if global_step > args.total_train_steps:
break
epoch += 1
logger.info('rank %d, finish training after %d steps', local_rank,
global_step)
L = 32
adam = Adam(bnn.params.parameters(), 0.001)
T = 2500
x1, x2 = -6, 6
y1, y2 = -100, 100
for i in range(T):
adam.zero_grad()
bnn.params.sample()
loss = lossf(X_,Y_)
loss.backward()
adam.step()
if i % 100 == 0:
print i, loss.data.numpy()[0]
N = 500
xx = varify(np.linspace(x1,x2,N).astype('float32').reshape(N,1))
yys = list()
for i in range(32):
bnn.params.sample()
yys.append(model(xx).data.numpy())