def train(args):
if not exists(args.path):
os.makedirs(args.path)
# make net
agent, agent_vocab, abstractor, net_args = configure_net(
args.abs_dir, args.ext_dir, args.cuda)
# configure training setting
assert args.stop > 0
train_params = configure_training(
'adam', args.lr, args.clip, args.decay, args.batch,
args.gamma, args.reward, args.stop, 'rouge-1'
)
train_batcher, val_batcher = build_batchers(args.batch)
# TODO different reward
reward_fn = compute_rouge_l
stop_reward_fn = compute_rouge_n(n=1)
# save abstractor binary
if args.abs_dir is not None:
abs_ckpt = {}
abs_ckpt['state_dict'] = load_best_ckpt(args.abs_dir)
abs_vocab = pkl.load(open(join(args.abs_dir, 'vocab.pkl'), 'rb'))
abs_dir = join(args.path, 'abstractor')
os.makedirs(join(abs_dir, 'ckpt'))
with open(join(abs_dir, 'meta.json'), 'w') as f:
json.dump(net_args['abstractor'], f, indent=4)
torch.save(abs_ckpt, join(abs_dir, 'ckpt/ckpt-0-0'))
with open(join(abs_dir, 'vocab.pkl'), 'wb') as f:
pkl.dump(abs_vocab, f)
# save configuration
meta = {}
meta['net'] = 'rnn-ext_abs_rl'
meta['net_args'] = net_args
meta['train_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
with open(join(args.path, 'agent_vocab.pkl'), 'wb') as f:
pkl.dump(agent_vocab, f)
# prepare trainer
grad_fn = get_grad_fn(agent, args.clip)
optimizer = optim.Adam(agent.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer, 'max', verbose=True,
factor=args.decay, min_lr=0,
patience=args.lr_p)
pipeline = A2CPipeline(meta['net'], agent, abstractor,
train_batcher, val_batcher,
optimizer, grad_fn,
reward_fn, args.gamma,
stop_reward_fn, args.stop)
trainer = BasicTrainer(pipeline, args.path,
args.ckpt_freq, args.patience, scheduler,
val_mode='score')
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
assert args.net_type in ['ff', 'rnn']
# create data batcher, vocabulary
# batcher
with open(join(DATA_DIR, 'vocab_cnt.pkl'), 'rb') as f:
wc = pkl.load(f)
word2id = make_vocab(wc, args.vsize)
train_batcher, val_batcher = build_batchers(args.net_type, word2id,
args.cuda, args.debug)
# make net
net, net_args = configure_net(args.net_type, len(word2id), args.emb_dim,
args.conv_hidden, args.lstm_hidden,
args.lstm_layer, args.bi)
if args.w2v:
# NOTE: the pretrained embedding having the same dimension
# as args.emb_dim should already be trained
embedding, _ = make_embedding({i: w
for w, i in word2id.items()}, args.w2v)
net.set_embedding(embedding)
# configure training setting
criterion, train_params = configure_training(args.net_type, 'adam',
args.lr, args.clip,
args.decay, args.batch)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['net'] = 'ml_{}_extractor'.format(args.net_type)
meta['net_args'] = net_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
val_fn = basic_validate(net, criterion)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.Adam(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'min',
verbose=True,
factor=args.decay,
min_lr=0,
patience=args.lr_p)
if args.cuda:
net = net.cuda()
pipeline = BasicPipeline(meta['net'], net, train_batcher, val_batcher,
args.batch, val_fn, criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path, args.ckpt_freq, args.patience,
scheduler)
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
# create data batcher, vocabulary
# batcher
with open(join(DATA_DIR, 'vocab_cnt.pkl'), 'rb') as f:
wc = pkl.load(f)
word2id = make_vocab(wc, args.vsize)
train_batcher, val_batcher = build_batchers(word2id, args.cuda, args.debug)
# make net
print('vocab size:', len(word2id))
ids = [id for word, id in word2id.items()]
print(max(ids))
print(list(sorted(ids))[0])
net, net_args = configure_net(len(word2id), args.emb_dim, args.n_hidden,
args.bi, args.n_layer, args.load_from)
# configure training setting
criterion, train_params = configure_training('adam', args.lr, args.clip,
args.decay, args.batch)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['net'] = 'base_abstractor'
meta['net_args'] = net_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
if args.cuda:
net = net.cuda()
val_fn = basic_validate(net, criterion)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.AdamW(net.parameters(), **train_params['optimizer'][1])
#optimizer = optim.Adagrad(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'min',
verbose=True,
factor=args.decay,
min_lr=0,
patience=args.lr_p)
pipeline = BasicPipeline(meta['net'], net, train_batcher, val_batcher,
args.batch, val_fn, criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path, args.ckpt_freq, args.patience,
scheduler)
print('start training with the following hyper-parameters:')
trainer.train()
def main(args):
assert args.net_type in ['ff', 'rnn']
# create data batcher, vocabulary
# batcher
with open(join(DATA_DIR, 'vocab_cnt.pkl'), 'rb') as f:
wc = pkl.load(f)
word2id = make_vocab(wc, args.vsize)
train_batcher, val_batcher = build_batchers(args.net_type, word2id,
args.cuda, args.debug)
# make net
net, net_args = configure_net(args.net_type,
len(word2id), args.emb_dim, args.conv_hidden,
args.lstm_hidden, args.lstm_layer, args.bi)
if args.w2v:
# NOTE: the pretrained embedding having the same dimension
# as args.emb_dim should already be trained
embedding, _ = make_embedding(
{i: w for w, i in word2id.items()}, args.w2v)
net.set_embedding(embedding)
# configure training setting
criterion, train_params = configure_training(
args.net_type, 'adam', args.lr, args.clip, args.decay, args.batch
)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['net'] = 'ml_{}_extractor'.format(args.net_type)
meta['net_args'] = net_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
val_fn = basic_validate(net, criterion)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.Adam(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer, 'min', verbose=True,
factor=args.decay, min_lr=0,
patience=args.lr_p)
if args.cuda:
net = net.cuda()
pipeline = BasicPipeline(meta['net'], net,
train_batcher, val_batcher, args.batch, val_fn,
criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path,
args.ckpt_freq, args.patience, scheduler)
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def train(args):
if not exists(args.path):
os.makedirs(args.path)
# make net
if args.docgraph or args.paragraph:
agent, agent_vocab, abstractor, net_args = configure_net_graph(
args.abs_dir, args.ext_dir, args.cuda, args.docgraph,
args.paragraph)
else:
agent, agent_vocab, abstractor, net_args = configure_net(
args.abs_dir, args.ext_dir, args.cuda, True, False, args.rl_dir)
if args.bert_stride > 0:
assert args.bert_stride == agent._bert_stride
# configure training setting
assert args.stop > 0
train_params = configure_training('adam', args.lr, args.clip, args.decay,
args.batch, args.gamma, args.reward,
args.stop, 'rouge-1')
if args.docgraph or args.paragraph:
if args.bert:
train_batcher, val_batcher = build_batchers_graph_bert(
args.batch, args.key, args.adj_type, args.max_bert_word,
args.docgraph, args.paragraph)
else:
train_batcher, val_batcher = build_batchers_graph(
args.batch, args.key, args.adj_type, args.gold_key,
args.docgraph, args.paragraph)
elif args.bert:
train_batcher, val_batcher = build_batchers_bert(
args.batch, args.bert_sent, args.bert_stride, args.max_bert_word)
else:
train_batcher, val_batcher = build_batchers(args.batch)
# TODO different reward
if args.reward == 'rouge-l':
reward_fn = compute_rouge_l
elif args.reward == 'rouge-1':
reward_fn = compute_rouge_n(n=1)
elif args.reward == 'rouge-2':
reward_fn = compute_rouge_n(n=2)
elif args.reward == 'rouge-l-s':
reward_fn = compute_rouge_l_summ
else:
raise Exception('Not prepared reward')
stop_reward_fn = compute_rouge_n(n=1)
# save abstractor binary
if args.abs_dir is not None:
abs_ckpt = {}
abs_ckpt['state_dict'] = load_best_ckpt(args.abs_dir, reverse=True)
abs_vocab = pkl.load(open(join(args.abs_dir, 'vocab.pkl'), 'rb'))
abs_dir = join(args.path, 'abstractor')
os.makedirs(join(abs_dir, 'ckpt'))
with open(join(abs_dir, 'meta.json'), 'w') as f:
json.dump(net_args['abstractor'], f, indent=4)
torch.save(abs_ckpt, join(abs_dir, 'ckpt/ckpt-0-0'))
with open(join(abs_dir, 'vocab.pkl'), 'wb') as f:
pkl.dump(abs_vocab, f)
# save configuration
meta = {}
meta['net'] = 'rnn-ext_abs_rl'
meta['net_args'] = net_args
meta['train_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
with open(join(args.path, 'agent_vocab.pkl'), 'wb') as f:
pkl.dump(agent_vocab, f)
# prepare trainer
grad_fn = get_grad_fn(agent, args.clip)
optimizer = optim.Adam(agent.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'max',
verbose=True,
factor=args.decay,
min_lr=1e-5,
patience=args.lr_p)
if args.docgraph or args.paragraph:
entity = True
else:
entity = False
pipeline = SCPipeline(meta['net'], agent, abstractor, train_batcher,
val_batcher, optimizer, grad_fn, reward_fn, entity,
args.bert)
trainer = BasicTrainer(pipeline,
args.path,
args.ckpt_freq,
args.patience,
scheduler,
val_mode='score')
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
# create data batcher, vocabulary
# batcher
word2id = pkl.load(open(join(args.abs_dir, 'vocab.pkl'), 'rb'))
# reward func
reward_func = None
reward_weight = 0.
# make net
net, net_args = configure_net(args.abs_dir)
bert = net._bert
train_batcher, val_batcher = build_batchers(word2id, args.cuda, args.debug)
# configure training setting
criterion, train_params = configure_training('adam', args.lr, args.clip,
args.decay, args.batch)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['net'] = 'base_abstractor'
meta['net_args'] = net_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
if args.cuda:
net = net.cuda()
val_fn = rl_validate(net,
reward_func=reward_func,
reward_coef=reward_weight,
_bleu=args.bleu,
f1=args.f1)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.AdamW(net.parameters(), **train_params['optimizer'][1])
#optimizer = optim.Adagrad(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'max',
verbose=True,
factor=args.decay,
min_lr=0,
patience=args.lr_p)
pipeline = AbsSelfCriticalPipeline(meta['net'],
net,
train_batcher,
val_batcher,
args.batch,
val_fn,
optimizer,
grad_fn,
weights=[args.r1, args.r2, args.rl],
_bleu=args.bleu,
f1=args.f1)
trainer = BasicTrainer(pipeline,
args.path,
args.ckpt_freq,
args.patience,
scheduler,
val_mode='score')
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
# create data batcher, vocabulary
# batcher
with open(join(args.data_path, 'vocab_cnt.pkl'), 'rb') as f:
wc = pkl.load(f)
word2id = make_vocab(wc, args.vsize, args.max_target_sent) #一个word的词典
train_batcher, val_batcher = build_batchers(word2id, args.cuda, args.debug)
# make net
if args.w2v:
# NOTE: the pretrained embedding having the same dimension
# as args.emb_dim should already be trained
embedding, _ = make_embedding({i: w
for w, i in word2id.items()},
args.w2v) #提供一个embedding矩阵
net, net_args = configure_net(len(word2id), args.emb_dim,
args.n_hidden, args.bi, args.n_layer,
args.sampling_teaching_force,
args.self_attn, args.hi_encoder,
embedding)
else:
print("please provide pretrain_w2v")
return
# configure training setting
criterion, train_params = configure_training('adam', args.lr, args.clip,
args.decay, args.batch)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
net_args_backup = net_args.copy()
del net_args_backup["embedding"]
meta = {}
meta['net'] = 'base_abstractor'
meta['net_args'] = net_args_backup
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
val_fn = basic_validate(net, criterion)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.Adam(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'min',
verbose=True,
factor=args.decay,
min_lr=0,
patience=args.lr_p)
if args.cuda:
net = net.cuda()
pipeline = BasicPipeline(meta['net'], net, train_batcher, val_batcher,
args.batch, val_fn, criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path, args.ckpt_freq, args.patience,
scheduler)
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
abstractor = get_abstractor(args.abs_dir, args.beam_search, args.cuda)
for split in ('train', 'val'):
decode(args, split)
embedding = abstractor._net._decoder._embedding
word2id = abstractor._word2id
id2words = {i: w for w, i in word2id.items()}
elmo = None
if args.elmo:
elmo = get_elmo(dropout=args.elmo_dropout,
vocab_to_cache=[id2words[i] for i in range(len(id2words))],
cuda=args.cuda,
projection_dim=args.elmo_projection)
args.emb_dim = elmo.get_output_dim()
meta = {
'net': '{}_discriminator'.format('cnn'),
'net_args': {
'vocab_size': len(abstractor._word2id),
'emb_dim': embedding.embedding_dim,
'kernel_num': args.kernel_num,
'kernel_sizes': args.kernel_sizes,
'class_num': 2,
'dropout': args.dropout,
'max_norm': args.max_norm,
'static': args.static,
},
'training_params': {
'optimizer': ('adam', {'lr': args.lr}),
'batch_size': args.batch,
'clip_grad_norm': args.clip,
'lr_decay': args.decay,
}
}
net = ConvNet(**meta['net_args'])
if elmo:
meta['net_args']['elmo'] = {
'dropout': args.elmo_dropout,
'projection': args.elmo_projection,
}
net.set_elmo_embedding(elmo)
else:
net.set_embedding(embedding.weight)
train_batcher, val_batcher = build_batchers(args, word2id)
def criterion(logit, target):
return F.cross_entropy(logit, target, reduce=False)
val_fn = basic_validate(net, criterion)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, net.parameters()), lr=args.lr)
scheduler = ReduceLROnPlateau(optimizer, 'min', verbose=True,
factor=args.decay, min_lr=0,
patience=args.lr_p)
if args.cuda:
net = net.cuda()
pipeline = BasicPipeline('discriminator', net,
train_batcher, val_batcher, args.batch, val_fn,
criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path,
args.ckpt_freq, args.patience, scheduler)
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def train(args):
assert args.encoder in ['BiLSTM', 'DeepLSTM', 'Transformer']
assert args.decoder in ['SL', 'PN']
assert args.emb_type in ['W2V', 'BERT']
# create data batcher, vocabulary
# batcher
with open(join(DATA_DIR, 'vocab_cnt.pkl'), 'rb') as f:
wc = pkl.load(f)
word2id = make_vocab(wc, args.vsize)
train_batcher, val_batcher = build_batchers(args.decoder, args.emb_type,
word2id, args.cuda, args.debug)
# make model
model, model_args = configure_net(args.encoder, args.decoder, args.emb_type, len(word2id),
args.emb_dim, args.conv_hidden, args.encoder_hidden,
args.encoder_layer)
if args.emb_type == 'W2V':
# NOTE: the pretrained embedding having the same dimension
# as args.emb_dim should already be trained
w2v_path='./CNNDM/word2vec/word2vec.128d.226k.bin'
embedding, _ = make_embedding(
{i: w for w, i in word2id.items()}, w2v_path)
model.set_embedding(embedding)
# configure training setting
criterion, train_params = configure_training(
args.decoder, 'adam', args.lr, args.clip, args.decay, args.batch
)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['model_args'] = model_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
val_fn = basic_validate(model, criterion, args.decoder)
grad_fn = get_basic_grad_fn(model, args.clip)
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer, 'min', verbose=True,
factor=args.decay, min_lr=2e-5,
patience=args.lr_p)
if args.cuda:
model = model.cuda()
pipeline = BasicPipeline(model, args.decoder,
train_batcher, val_batcher, args.batch, val_fn,
criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path,
args.ckpt_freq, args.patience, scheduler)
# for name, para in net.named_parameters():
# if para.requires_grad:
# print(name)
print('Start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
# create data batcher, vocabulary
# batcher
word2id = pkl.load(open(join(args.abs_dir, 'vocab.pkl'), 'rb'))
# reward func
if args.reward_model_dir is not None:
assert args.reward_data_dir is not None
reward_func = cloze_reward(args.reward_model_dir, args.cloze_device)
reward_weight = args.reward_weight
else:
reward_func = None
reward_weight = 0.
# make net
if args.docgraph or args.paragraph:
net, net_args = configure_net_graph(args.abs_dir, args.docgraph,
args.paragraph)
else:
net, net_args = configure_net(args.abs_dir)
bert = net._bert
if bert:
print('model use bert')
import logging
print('disable')
logging.getLogger('transformers.tokenization_utils').setLevel(
logging.ERROR)
logging.getLogger('transformers.tokenization_utils').disabled = True
if args.docgraph or args.paragraph:
if bert:
tokenizer = net._bert_model._tokenizer
train_batcher, val_batcher = build_batchers_graph_bert(
tokenizer, args.cuda, args.debug, args.key, net._adj_type,
args.docgraph, args.reward_data_dir)
else:
train_batcher, val_batcher = build_batchers_graph(
word2id, args.cuda, args.debug, args.key, net._adj_type,
args.docgraph, args.reward_data_dir)
else:
if bert:
tokenizer = net._bert_model._tokenizer
train_batcher, val_batcher = build_batchers_bert(
tokenizer, args.cuda, args.debug)
else:
train_batcher, val_batcher = build_batchers(
word2id, args.cuda, args.debug)
# configure training setting
criterion, train_params = configure_training('adam', args.lr, args.clip,
args.decay, args.batch)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['net'] = 'base_abstractor'
meta['net_args'] = net_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
local_coh_fun = None
# prepare trainer
if args.cuda:
net = net.cuda()
multigpu = False
val_fn = rl_validate(net,
reward_func=reward_func,
reward_coef=reward_weight,
bert=bert)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.AdamW(net.parameters(), **train_params['optimizer'][1])
#optimizer = optim.Adagrad(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'max',
verbose=True,
factor=args.decay,
min_lr=0,
patience=args.lr_p)
print('rouge weights:', [args.r1, args.r2, args.rl])
pipeline = AbsSelfCriticalPipeline(
meta['net'],
net,
train_batcher,
val_batcher,
args.batch,
val_fn,
optimizer,
grad_fn,
reward_func,
reward_weight,
local_coh_fun,
0.,
accumulate_g_step=args.accumulate_g_step,
weights=[args.r1, args.r2, args.rl],
bert=bert,
multigpu=multigpu,
ml_loss=args.ml_loss)
trainer = BasicTrainer(pipeline,
args.path,
args.ckpt_freq,
args.patience,
scheduler,
val_mode='score')
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
# create data batcher, vocabulary
# batcher
with open(join(DATA_DIR, 'vocab_cnt.pkl'), 'rb') as f:
wc = pkl.load(f)
word2id = make_vocab(wc, args.vsize)
train_batcher, val_batcher = build_batchers(word2id, args.cuda, args.debug)
# make net
net, net_args = configure_net(len(word2id), args.emb_dim, args.n_hidden,
args.bi, args.n_layer)
if args.w2v:
# NOTE: the pretrained embedding having the same dimension
# as args.emb_dim should already be trained
embedding, oov = make_embedding({i: w
for w, i in word2id.items()},
args.w2v)
net.set_embedding(embedding)
# configure training setting
criterion, train_params = configure_training('adam', args.lr, args.clip,
args.decay, args.batch)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['net'] = 'base_abstractor'
meta['net_args'] = net_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
val_fn = basic_validate(net, criterion)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.Adam(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'min',
verbose=True,
factor=args.decay,
min_lr=0,
patience=args.lr_p)
if args.cuda:
net = net.cuda()
pipeline = BasicPipeline(meta['net'], net, train_batcher, val_batcher,
args.batch, val_fn, criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path, args.ckpt_freq, args.patience,
scheduler)
print('start training with the following hyper-parameters:')
print(meta)
# # Print model's state_dict
# print("Model's state_dict:")
# for param_tensor in net.state_dict():
# print(param_tensor, "\t", net.state_dict()[param_tensor].size())
#
# # Print optimizer's state_dict
# print("Optimizer's state_dict:")
# for var_name in optimizer.state_dict():
# print(var_name, "\t", optimizer.state_dict()[var_name])
# # IMPORT PRETRAINED MODEL PARAMETERS
# net.load_state_dict(torch.load(
# 'pretrained_eng_model/abstractor/ckpt/ckpt-0-0')['state_dict'])
# net.eval() # do I need that or not?
# copy net
# from copy import deepcopy
# net_copy = deepcopy(net)
# net_copy.load_state_dict(torch.load('pretrained_eng_model/abstractor/ckpt/ckpt-0-0', map_location='cpu')['state_dict'])
# for key in net_copy.state_dict():
# print('key: ', key)
# param = net_copy.state_dict()[key]
# print('param.shape: ', param.shape)
# print('param.requires_grad: ', param.requires_grad)
# print('param.shape, param.requires_grad: ', param.shape, param.requires_grad)
# print('isinstance(param, nn.Module) ', isinstance(param, nn.Module))
# print('isinstance(param, nn.Parameter) ', isinstance(param, nn.Parameter))
# print('isinstance(param, torch.Tensor): ', isinstance(param, torch.Tensor))
# print('=====')
# save current state dict
model_dict = net.state_dict()
# save some parameters for testing purposes if the dict was loaded successfully
p1 = net._embedding.weight[0][0].detach().cpu().numpy()
p2 = net._enc_lstm.weight_hh_l0[0][0].detach().cpu().numpy()
p3 = net._attn_wm.data[0][0].detach().cpu().numpy()
# print(p1)
# print(p2)
# print(p3)
# load dict from pretrained net
ABS_DIR = os.environ['ABS']
print(ABS_DIR)
# uncomment for gpu
# pretrained_dict = torch.load(ABS_DIR)['state_dict']
pretrained_dict = torch.load(ABS_DIR)['state_dict']
# skip embedding weights
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k != '_embedding.weight'
}
# overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
print('Model will be trained on device:')
print(model_dict['_embedding.weight'].device)
# load the new state dict
net.load_state_dict(model_dict)
# check if the update was correct
pn1 = net._embedding.weight[0][0].detach().cpu().numpy()
pn2 = net._enc_lstm.weight_hh_l0[0][0].detach().cpu().numpy()
pn3 = net._attn_wm.data[0][0].detach().cpu().numpy()
# print(pn1)
# print(pn2)
# print(pn3)
assert p1 == pn1 # embedding layer has to be the same
assert p2 != pn2
assert p3 != pn3
print('Embedding layer has not been overwritten')
# set updating of the parameters
for name, param in net.named_parameters():
#param.requires_grad = True
print(name, param.requires_grad)
trainer.train()
def main(args):
assert args.net_type in ['ff', 'rnn']
# create data batcher, vocabulary
# batcher
with open(join(DATA_DIR, 'vocab_cnt.pkl'), 'rb') as f:
wc = pkl.load(f)
word2id = make_vocab(wc, args.vsize)
id2words = {i: w for w, i in word2id.items()}
elmo = None
if args.elmo:
elmo = get_elmo(
dropout=args.elmo_dropout,
vocab_to_cache=[id2words[i] for i in range(len(id2words))],
cuda=args.cuda,
projection_dim=args.elmo_projection)
args.emb_dim = elmo.get_output_dim()
train_batcher, val_batcher = build_batchers(args.net_type, word2id,
args.cuda, args.debug)
# make net
net, net_args = configure_net(args.net_type, len(word2id), args.emb_dim,
args.conv_hidden, args.lstm_hidden,
args.lstm_layer, args.bi)
if elmo:
net_args['elmo'] = {
'dropout': args.elmo_dropout,
'projection': args.elmo_projection,
}
net.set_elmo_embedding(elmo)
elif args.w2v:
# NOTE: the pretrained embedding having the same dimension
# as args.emb_dim should already be trained
embedding, _ = make_embedding(id2words, args.w2v)
net.set_embedding(embedding)
# configure training setting
criterion, train_params = configure_training(args.net_type, 'adam',
args.lr, args.clip,
args.decay, args.batch)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {
'net': 'ml_{}_extractor'.format(args.net_type),
'net_args': net_args,
'traing_params': train_params
}
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
# prepare trainer
val_fn = basic_validate(net, criterion)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.Adam(filter(lambda p: p.requires_grad, net.parameters()),
**train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer,
'min',
verbose=True,
factor=args.decay,
min_lr=0,
patience=args.lr_p)
if args.cuda:
net = net.cuda()
pipeline = BasicPipeline(meta['net'], net, train_batcher, val_batcher,
args.batch, val_fn, criterion, optimizer, grad_fn)
trainer = BasicTrainer(pipeline, args.path, args.ckpt_freq, args.patience,
scheduler)
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()
def main(args):
# create data batcher, vocabulary
# batcher
word2id = pkl.load(open(join(args.abs_dir, 'vocab.pkl'), 'rb'))
train_batcher, val_batcher = build_batchers(word2id,
args.cuda, args.debug)
# make net
net, net_args = configure_net(args.abs_dir)
# configure training setting
criterion, train_params = configure_training(
'adam', args.lr, args.clip, args.decay, args.batch
)
# save experiment setting
if not exists(args.path):
os.makedirs(args.path)
with open(join(args.path, 'vocab.pkl'), 'wb') as f:
pkl.dump(word2id, f, pkl.HIGHEST_PROTOCOL)
meta = {}
meta['net'] = 'base_abstractor'
meta['net_args'] = net_args
meta['traing_params'] = train_params
with open(join(args.path, 'meta.json'), 'w') as f:
json.dump(meta, f, indent=4)
if args.coherence or args.all_local:
print('use coref score')
word_dict, model, sess = graph_init(model="cnndm")
coh_func = (model, sess, word_dict)
print('finish load coref model')
else:
coh_func = None
assert not (args.anaphora and args.apposition)
if args.anaphora:
print('use anaphora')
local_coh_fun = getAnaphoraReward
elif args.apposition:
print('use apposition')
local_coh_fun = getAppositionReward
elif args.all_local:
local_coh_fun = getLocalReward
else:
local_coh_fun = None
# prepare trainer
if args.cuda:
net = net.cuda()
val_fn = rl_validate(net, coherence_func=coh_func, coh_coef=args.coh_coef, local_coh_func=local_coh_fun, local_coh_coef=args.local_coh_coef)
grad_fn = get_basic_grad_fn(net, args.clip)
optimizer = optim.Adam(net.parameters(), **train_params['optimizer'][1])
#optimizer = optim.Adagrad(net.parameters(), **train_params['optimizer'][1])
scheduler = ReduceLROnPlateau(optimizer, 'max', verbose=True,
factor=args.decay, min_lr=0,
patience=args.lr_p)
pipeline = AbsSelfCriticalPipeline(meta['net'], net,
train_batcher, val_batcher, args.batch, val_fn, optimizer, grad_fn, coh_func, args.coh_coef, local_coh_fun, args.local_coh_coef)
trainer = BasicTrainer(pipeline, args.path,
args.ckpt_freq, args.patience, scheduler, val_mode='score')
print('start training with the following hyper-parameters:')
print(meta)
trainer.train()