def eval(args):
paddle.set_device(args.device)
if not args.init_from_ckpt:
raise ValueError('init_from_ckpt should be set when eval.')
vocab = load_vocab(args.vocab_file, args.max_characters_per_token)
elmo = ELMo(args.batch_size,
args.char_embed_dim,
args.projection_dim,
vocab.size,
dropout=args.dropout,
num_layers=args.num_layers,
num_highways=args.num_highways,
char_vocab_size=vocab.char_size)
elmo.eval()
elmo_loss = ELMoLoss()
# Loads pre-trained parameters.
weight_state_dict = paddle.load(args.init_from_ckpt + '.pdparams')
elmo.set_state_dict(weight_state_dict)
print("Loaded checkpoint from %s" % args.init_from_ckpt)
dev_dataset = OneBillionWordDataset(args.dev_data_path,
vocab,
args.batch_size,
args.unroll_steps,
mode='test',
shuffle=False,
seed=args.seed)
dev_dataloader = DataLoader(dev_dataset, return_list=True, batch_size=None)
total_step = total_loss = 0
total_time = 0.0
batch_start_time = time.time()
for step, inputs in enumerate(dev_dataloader, start=1):
ids, next_ids, ids_reverse, next_ids_reverse = inputs
outputs = elmo([ids, ids_reverse])
loss = elmo_loss(outputs, [next_ids, next_ids_reverse])
ppl = paddle.exp(loss)
total_loss += loss.numpy()[0]
total_step += 1
total_time += (time.time() - batch_start_time)
if step % args.log_freq == 0:
print("Eval step %d - loss: %.4f - Perplexity: %.4f - %.3fs/step" %
(step, loss.numpy()[0] * args.unroll_steps, ppl.numpy()[0],
total_time / args.log_freq))
total_time = 0.0
batch_start_time = time.time()
avg_loss = total_loss / total_step
avg_ppl = math.exp(avg_loss)
print("Eval - average loss: %.4f - average Perplexity: %.4f" %
(avg_loss * args.unroll_steps, avg_ppl))
def __init__(self, model, options, vocab, nnvecs=1):
self.word_counts, words, chars, pos, cpos, rels, treebanks, langs = vocab
self.model = model
self.nnvecs = nnvecs
# Load ELMo if the option is set
if options.elmo is not None:
from elmo import ELMo
self.elmo = ELMo(options.elmo, options.elmo_gamma,
options.elmo_learn_gamma)
self.elmo.init_weights(model)
else:
self.elmo = None
extra_words = 2 # MLP padding vector and OOV vector
self.words = {word: ind for ind, word in enumerate(words, extra_words)}
self.word_lookup = self.model.add_lookup_parameters(
(len(self.words) + extra_words, options.word_emb_size))
extra_pos = 2 # MLP padding vector and OOV vector
self.pos = {pos: ind for ind, pos in enumerate(cpos, extra_pos)}
self.pos_lookup = self.model.add_lookup_parameters(
(len(cpos) + extra_pos, options.pos_emb_size))
self.irels = rels
self.rels = {rel: ind for ind, rel in enumerate(rels)}
extra_chars = 1 # OOV vector
self.chars = {char: ind for ind, char in enumerate(chars, extra_chars)}
self.char_lookup = self.model.add_lookup_parameters(
(len(chars) + extra_chars, options.char_emb_size))
extra_treebanks = 1 # Padding vector
self.treebanks = {
treebank: ind
for ind, treebank in enumerate(treebanks, extra_treebanks)
}
self.treebank_lookup = self.model.add_lookup_parameters(
(len(treebanks) + extra_treebanks, options.tbank_emb_size))
# initialise word vectors with external embeddings where they exist
# This part got ugly - TODO: refactor
if not options.predict:
self.external_embedding = defaultdict(lambda: {})
if options.ext_word_emb_file and options.word_emb_size > 0:
# Load pre-trained word embeddings
for lang in langs:
embeddings = utils.get_external_embeddings(
options,
emb_file=options.ext_word_emb_file,
lang=lang,
words=self.words.viewkeys())
self.external_embedding["words"].update(embeddings)
if options.ext_char_emb_file and options.char_emb_size > 0:
# Load pre-trained character embeddings
for lang in langs:
embeddings = utils.get_external_embeddings(
options,
emb_file=options.ext_char_emb_file,
lang=lang,
words=self.chars,
chars=True)
self.external_embedding["chars"].update(embeddings)
if options.ext_emb_dir:
# For every language, load the data for the word and character
# embeddings from a directory.
for lang in langs:
if options.word_emb_size > 0:
embeddings = utils.get_external_embeddings(
options,
emb_dir=options.ext_emb_dir,
lang=lang,
words=self.words.viewkeys())
self.external_embedding["words"].update(embeddings)
if options.char_emb_size > 0:
embeddings = utils.get_external_embeddings(
options,
emb_dir=options.ext_emb_dir,
lang=lang,
words=self.chars,
chars=True)
self.external_embedding["chars"].update(embeddings)
self.init_lookups(options)
elmo_emb_size = self.elmo.emb_dim if self.elmo else 0
self.lstm_input_size = (
options.word_emb_size + elmo_emb_size + options.pos_emb_size +
options.tbank_emb_size + 2 *
(options.char_lstm_output_size if options.char_emb_size > 0 else 0)
)
print "Word-level LSTM input size: " + str(self.lstm_input_size)
self.bilstms = []
if options.no_bilstms > 0:
self.bilstms.append(
BiLSTM(self.lstm_input_size,
options.lstm_output_size,
self.model,
dropout_rate=0.33))
for i in range(1, options.no_bilstms):
self.bilstms.append(
BiLSTM(2 * options.lstm_output_size,
options.lstm_output_size,
self.model,
dropout_rate=0.33))
#used in the PaddingVec
self.word2lstm = self.model.add_parameters(
(options.lstm_output_size * 2, self.lstm_input_size))
self.word2lstmbias = self.model.add_parameters(
(options.lstm_output_size * 2))
else:
self.word2lstm = self.model.add_parameters(
(self.lstm_input_size, self.lstm_input_size))
self.word2lstmbias = self.model.add_parameters(
(self.lstm_input_size))
self.char_bilstm = BiLSTM(options.char_emb_size,
options.char_lstm_output_size,
self.model,
dropout_rate=0.33)
self.charPadding = self.model.add_parameters(
(options.char_lstm_output_size * 2))
def eval():
paddle.disable_static()
n_gpus = dist.get_world_size()
rank = dist.get_rank()
if n_gpus > 1:
dist.init_parallel_env()
args = parse_args()
if not args.init_from_ckpt:
raise ValueError('init_from_ckpt should be set when eval.')
vocab = load_vocab(args.vocab_file, args.max_characters_per_token)
elmo = ELMo(args.batch_size,
args.char_embed_dim,
args.projection_dim,
vocab.size,
dropout=args.dropout,
num_layers=args.num_layers,
num_highways=args.num_highways,
char_vocab_size=vocab.char_size)
if n_gpus > 1:
elmo = paddle.DataParallel(elmo)
elmo.eval()
elmo_loss = ELMoLoss()
# Loads pre-trained parameters.
weight_state_dict = paddle.load(args.init_from_ckpt + '.pdparams')
elmo.set_state_dict(weight_state_dict)
print("Loaded checkpoint from %s" % args.init_from_ckpt)
dev_dataset = OneBillionWordDataset(args.dev_data_path,
vocab,
args.batch_size,
args.unroll_steps,
n_gpus,
rank,
mode='test',
shuffle=False,
seed=args.random_seed)
# FIXME(xiemoyuan): When DataLoader support setting batch_size to None,
# setting batch_size to None.
dev_dataloader = DataLoader(dev_dataset, return_list=True, batch_size=1)
total_step = total_loss = 0
total_time = 0.0
batch_start_time = time.time()
for step, inputs in enumerate(dev_dataloader, start=1):
# FIXME(xiemoyuan): When DataLoader support setting batch_size to None,
# deleting the operation of squeeze.
for j in range(len(inputs)):
inputs[j] = paddle.squeeze(inputs[j], axis=0)
ids, next_ids, ids_reverse, next_ids_reverse = inputs
outputs = elmo([ids, ids_reverse])
loss = elmo_loss(outputs, [next_ids, next_ids_reverse])
ppl = paddle.exp(loss)
total_loss += loss.numpy()[0]
total_step += 1
total_time += (time.time() - batch_start_time)
if rank == 0:
if step % args.log_freq == 0:
print(
"Eval step %d - loss: %.4f - Perplexity: %.4f - %.3fs/step"
% (step, loss.numpy()[0] * args.unroll_steps,
ppl.numpy()[0], total_time / args.log_freq))
total_time = 0.0
batch_start_time = time.time()
avg_loss = total_loss / total_step
avg_ppl = math.exp(avg_loss)
if rank == 0:
print("Eval - average loss: %.4f - average Perplexity: %.4f" %
(avg_loss * args.unroll_steps, avg_ppl))
from tensorflow.contrib import seq2seq
from elmo import ELMo
from data import NERData
import os
total_epoch = 5000
hidden_size = 200
vocab_size = 5000
max_length = 128
entity_class = 8
lr = 1e-4
batch_size = 256
ner = NERData(batch_size, max_length)
elmo = ELMo(batch_size, hidden_size, vocab_size)
def network(X):
w = tf.get_variable("fcn_w", [1, hidden_size, entity_class])
b = tf.get_variable("fcn_b", [entity_class])
w_tile = tf.tile(w, [batch_size, 1, 1])
logists = tf.nn.softmax(tf.nn.xw_plus_b(X, w_tile, b), name="logists")
return logists
def train():
X = tf.placeholder(shape=[batch_size, max_length],
dtype=tf.int32,
name="X")
for _ in range(len(SENT2VEC))
]
for i in range(len(SENT2VEC)):
s2vsingle[i].load_state(SENT2VEC[i])
s2vsingle[i].set_w2v_path(PATH_TO_W2V)
s2vsingle[i] = s2vsingle[i].cuda()
sent2vec = Sent2Vec(s2vsingle, 'concat')
params_model = {'bsize': 64, 'pool_type': 'mean',
'which_layer': 'all',
'optfile': ELMO_OPTIONS,
'wgtfile': ELMO_WEIGHT}
elmo = ELMo(params_model)
elmo = elmo.cuda()
gensen_1 = GenSenSingle(
model_folder=FOLDER_PATH,
filename_prefix=PREFIX1,
pretrained_emb=PRETRAIN_EMB,
cuda=True
)
gensen_2 = GenSenSingle(
model_folder=FOLDER_PATH,
filename_prefix=PREFIX2,
pretrained_emb=PRETRAIN_EMB,
cuda=True
)
gensen = GenSen(gensen_1, gensen_2)
# The list contains train, valid and test dataset. Each tuple in the list is the dataset of word tokens and the dataset of character tokens
datasets, field_word, field_char = gen_language_model_corpus(WikiText2)
train_data, valid_data, test_data = datasets
VOCAB_DIM = len(field_char.vocab)
OUTPUT_DIM = len(field_word.vocab)
# OTHER HYPER-PARAMETERS
BATCH_SIZE = 32
N_EPOCHS = 100
CLIP = 1
best_valid_loss = float('inf')
# PAD_IDX = field_word.vocab.stoi["<pad>"] # PAD token for word, NOT CHAR
model = ELMo(VOCAB_DIM, OUTPUT_DIM, CHAR_EMB_DIM, HID_DIM, PRJ_DIM, FILTERS,
CHAR_LEN, N_LAYERS).to(DEVICE)
# Initialize
model.init_weights()
print(f'The model has {count_parameters(model):,} trainable parameters')
import time
# criterion = cal_loss
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters())
train_losses = []
test_losses = []
for epoch in range(1, N_EPOCHS + 1):
def train(args):
paddle.set_device(args.device)
n_procs = dist.get_world_size()
rank = dist.get_rank()
if n_procs > 1:
dist.init_parallel_env()
vocab = load_vocab(args.vocab_file, args.max_characters_per_token)
elmo = ELMo(args.batch_size,
args.char_embed_dim,
args.projection_dim,
vocab.size,
dropout=args.dropout,
num_layers=args.num_layers,
num_highways=args.num_highways,
char_vocab_size=vocab.char_size)
if n_procs > 1:
elmo = paddle.DataParallel(elmo)
elmo.train()
gloabl_norm_clip = nn.ClipGradByGlobalNorm(args.max_grad_norm)
optimizer = paddle.optimizer.Adagrad(learning_rate=args.lr,
parameters=elmo.parameters(),
initial_accumulator_value=1.0,
grad_clip=gloabl_norm_clip)
elmo_loss = ELMoLoss()
# Loads pre-trained parameters.
if args.init_from_ckpt:
weight_state_dict = paddle.load(args.init_from_ckpt + '.pdparams')
opt_state_dict = paddle.load(args.init_from_ckpt + '.pdopt')
elmo.set_state_dict(weight_state_dict)
optimizer.set_state_dict(opt_state_dict)
print("Loaded checkpoint from %s" % args.init_from_ckpt)
train_dataset = OneBillionWordDataset(args.train_data_path,
vocab,
args.batch_size,
args.unroll_steps,
n_procs=n_procs,
rank=rank,
mode='train',
shuffle=True,
seed=args.seed)
train_dataloader = DataLoader(train_dataset,
return_list=True,
batch_size=None)
n_tokens_per_batch = args.batch_size * args.unroll_steps * n_procs
n_steps_per_epoch = int(train_dataset.number_of_tokens /
n_tokens_per_batch)
n_steps_total = args.epochs * n_steps_per_epoch
print("Training for %s epochs and %s steps" % (args.epochs, n_steps_total))
total_time = 0.0
batch_start_time = time.time()
for step, inputs in enumerate(train_dataloader, start=1):
ids, next_ids, ids_reverse, next_ids_reverse = inputs
outputs = elmo([ids, ids_reverse])
loss = elmo_loss(outputs, [next_ids, next_ids_reverse])
ppl = paddle.exp(loss)
loss *= args.unroll_steps
loss.backward()
optimizer.step()
optimizer.clear_grad()
total_time += (time.time() - batch_start_time)
if step % args.log_freq == 0:
print("step %d/%d - loss: %.4f - Perplexity: %.4f - %.3fs/step" %
(step, n_steps_total, loss.numpy()[0], ppl.numpy()[0],
total_time / args.log_freq))
total_time = 0.0
if rank == 0 and step % args.save_freq == 0:
save_params(elmo, optimizer, args.save_dir, step)
if step == n_steps_total:
# training done
if rank == 0:
save_params(elmo, optimizer, args.save_dir, 'final')
break
batch_start_time = time.time()
'epoch_size': 4
}
# Set up logger
logging.basicConfig(format='%(asctime)s : %(message)s', level=logging.DEBUG)
if __name__ == "__main__":
# Load InferSent model
params_model = {
'bsize': 64,
'pool_type': 'mean',
'which_layer': 'all',
'optfile': OPT_PATH,
'wgtfile': MODEL_PATH
}
model = ELMo(params_model)
params_senteval['elmo'] = model.cuda()
se = senteval.engine.SE(params_senteval, batcher, prepare)
transfer_tasks = [
'STS12', 'STS13', 'STS14', 'STS15', 'STS16', 'MR', 'CR', 'MPQA',
'SUBJ', 'SST2', 'SST5', 'TREC', 'MRPC', 'SICKEntailment',
'SICKRelatedness', 'STSBenchmark', 'Length', 'WordContent', 'Depth',
'TopConstituents', 'BigramShift', 'Tense', 'SubjNumber', 'ObjNumber',
'OddManOut', 'CoordinationInversion'
]
results_transfer = se.eval(transfer_tasks)
print('--------------------------------------------')
print('MR [Dev:%.1f/Test:%.1f]' %
(results_transfer['MR']['devacc'], results_transfer['MR']['acc']))