def create_dataset(path_to_dataset,batch_size,split_ratio=0.7,min_vocab_freq=10,max_vocab_size=4000):
text_field = Field(tokenize="spacy",tokenizer_language="en",batch_first=True,init_token="<sos>",eos_token="<eos>",lower=True)
def transform(caption):
caption = caption.strip().lower().split()
return caption
dataset = CocoCaptions(annFile=os.path.join(path_to_dataset,"captions_train2014.json"),text_field=text_field,transform=transform)
train,val = dataset.split(split_ratio=split_ratio)
test = CocoCaptions(annFile=os.path.join(path_to_dataset,"captions_val2014.json"),text_field=text_field,transform=transform)
print("Dataset loaded")
print("Train set size:",len(train))
text_field.build_vocab(dataset.text,min_freq=min_vocab_freq,max_size=max_vocab_size)
SOS_TOKEN = text_field.vocab.stoi['<sos>']
EOS_TOKEN = text_field.vocab.stoi['<eos>']
UNK_TOKEN = text_field.vocab.stoi['<unk>']
PAD_TOKEN = text_field.vocab.stoi['<pad>']
print("Vocabuly build")
print("Vocabuly statistics")
print("\nMost common words in the vocabulary:\n",text_field.vocab.freqs.most_common(10))
print("Size of the vocabulary:",len(text_field.vocab))
print("Max sequence lenght",dataset.max_seq_len)
train_iter,val_iter = BucketIterator.splits((train,val),repeat=False,batch_size=batch_size)
test_iter = BucketIterator(test,batch_size=batch_size,repeat=False,train=False)
vocab_dict = text_field.vocab.stoi
return {"data_iters":(train_iter,val_iter,test_iter),"fields":text_field,
"word_to_num_vocab":vocab_dict,"num_to_word_vocab":{y:x for x,y in vocab_dict.items()},
"num_classes":len(text_field.vocab),"tokens":(SOS_TOKEN,EOS_TOKEN,UNK_TOKEN,PAD_TOKEN),"max_seq_len":dataset.max_seq_len}
def load_dataloaders(args):
logger.info("Preparing dataloaders...")
FR = torchtext.data.Field(tokenize=dum_tokenizer, lower=True, init_token="<sos>", eos_token="<eos>",\
batch_first=True)
EN = torchtext.data.Field(tokenize=dum_tokenizer, lower=True, batch_first=True)
train_path = os.path.join("./data/", "df.csv")
if not os.path.isfile(train_path):
tokenize_data(args)
train = torchtext.data.TabularDataset(train_path, format="csv", \
fields=[("EN", EN), ("FR", FR)])
FR.build_vocab(train)
EN.build_vocab(train)
train_iter = BucketIterator(train, batch_size=args.batch_size, repeat=False, sort_key=lambda x: (len(x["EN"]), len(x["FR"])),\
shuffle=True, train=True)
train_length = len(train)
logger.info("Loaded dataloaders.")
return train_iter, FR, EN, train_length
def eval(self, test_set):
print("========test only=======")
test_iter = BucketIterator(test_set,
batch_size=self.args.batch,
train=False,
shuffle=False,
device=self.model.device,
sort_key=lambda x: len(x.TOKENS))
with torch.no_grad():
self.run_a_epoch("final test",
test_iter,
need_backward=False,
epoch_num=0,
save_output=os.path.join(self.args.output_path,
"check"),
max_step=ceil(len(test_set) / self.args.batch))
print(self.model.argumentRoleClsLayer.arg_mask_init.tolist())
print(self.model.argumentRoleClsLayer.arg_mask.tolist())
def train(self,
epochs,
save_path,
load_previous=True,
clip=10,
batch_size=128):
save_dir = os.path.split(save_path)[0]
os.makedirs(save_dir, exist_ok=True)
if load_previous and os.path.exists(save_path):
self._logger.debug(f'Loading model state from {save_path}')
self.model.load_state_dict(torch.load(save_path))
train_iterator, test_iterator = BucketIterator.splits(
(self.train_data, self.test_data),
batch_size=batch_size,
device=self.device,
sort_key=lambda x: len(x.src))
optimizer = optim.Adam(self.model.parameters())
trg_pad_idx = self.trg_field.vocab.stoi[WordToPhonemeModel.PAD_TOKEN]
criterion = nn.CrossEntropyLoss(ignore_index=trg_pad_idx)
# Training loop
self._logger.debug(f'Beginning training for {epochs} epoch(s)')
for epoch in range(epochs):
train_loss = self._train_iter(train_iterator, optimizer, criterion,
clip)
test_loss = self._evaluate_iter(test_iterator, criterion)
if test_loss < self.best_test_loss:
# Save model if better
self.best_test_loss = test_loss
torch.save(self.model.state_dict(), save_path)
self._logger.debug(
f'| Epoch: {epoch+1:03} | Train Loss: {train_loss:.3f} | Train PPL: {math.exp(train_loss):7.3f} | Test Loss: {test_loss:.3f} | Test PPL: {math.exp(test_loss):7.3f} |'
)
# Save model
torch.save(self.model.state_dict(), save_path)
self._logger.info(save_path)
return self.best_test_loss
def load_naive_iterators(args, path, fields):
train, valid, test = NaiveDatasetClassification.splits(\
exts = (args.src_ext, args.trg_ext),
fields = fields,
root=path)
# Some stats
print("Stats for dataset in ", path)
# Train
count = Counter()
for e in train.examples:
count.update([e.label])
train_d = dict(count)
train_d['name'] = "train"
# Val
count = Counter()
for e in valid.examples:
count.update([e.label])
val_d = dict(count)
val_d['name'] = 'val'
# Test
count = Counter()
for e in test.examples:
count.update([e.label])
test_d = dict(count)
test_d['name'] = 'test'
d_list = [train_d, val_d, test_d]
df = data_stats.dicts_to_pandas(d_list)
md = tabulate(df, headers='keys', tablefmt='pipe')
print(md)
fields[1].build_vocab(train)
# Data iterators
train_iterator, valid_iterator, test_iterator = BucketIterator.splits(
(train, valid, test),
batch_size=args.batch_size,
sort_within_batch=True,
sort_key=lambda x: len(x.text),
device=args.device)
return train_iterator, valid_iterator, test_iterator, fields[1]
def create(cls, config):
src_field = Field(init_token='<sos>',
eos_token='<eos>',
pad_token='<pad>',
include_lengths=True)
trg_field = Field(init_token='<sos>',
eos_token='<eos>',
pad_token='<pad>',
lower=True,
include_lengths=True)
train = TranslationDataset(path=config.train_prefix,
exts=config.exts,
fields=(src_field, trg_field))
valid = TranslationDataset(path=config.valid_prefix,
exts=config.exts,
fields=(src_field, trg_field))
test = TranslationDataset(path=config.test_prefix,
exts=config.exts,
fields=(src_field, trg_field))
train_it, valid_it, test_it = BucketIterator.splits(
[train, valid, test],
batch_sizes=config.batch_sizes,
sort_key=TranslationDataset.sort_key,
device=-1)
src_field.build_vocab(train, min_freq=10)
trg_field.build_vocab(train, min_freq=10)
src_voc = src_field.vocab
trg_voc = trg_field.vocab
model = Seq2Seq.create(src_voc, trg_voc, config)
if config.use_cuda:
model = model.cuda()
return Trainer(model, train_it, valid_it, test_it, config.valid_step,
config.checkpoint_path, config.pool_size)
def load_dataset(data_dir, word_dir, batch_size, device):
TEXT = data.Field(batch_first=True, include_lengths=True, lower=True)
LABEL = data.LabelField(batch_first=True)
fields = {'sentence1': ('premise', TEXT),
'sentence2': ('hypothesis', TEXT),
'gold_label': ('label', LABEL)}
trainDataset, valDataset, testDataset = data.TabularDataset.splits(
path=data_dir,
format='json',
train='snli_1.0_train.jsonl',
validation='snli_1.0_dev.jsonl',
test='snli_1.0_test.jsonl',
fields=fields,
filter_pred=lambda x: x.label != '-'
)
vectors = Vectors('glove.6B.200d.txt', word_dir)
TEXT.build_vocab(trainDataset, vectors=vectors, unk_init=nn.init.xavier_uniform)
LABEL.build_vocab(trainDataset)
train_iter, val_iter = BucketIterator.splits(
datasets=(trainDataset, valDataset),
batch_sizes=(batch_size, batch_size),
device=device,
sort_key=lambda x: len(x.premise) + len(x.hypothesis),
sort_within_batch=True,
repeat=False,
shuffle=True
)
test_iter = Iterator(
dataset=testDataset,
batch_size=batch_size,
device=device,
sort=False,
repeat=False,
shuffle=False
)
return TEXT, LABEL, train_iter, val_iter, test_iter
def score(self, sentences: List[str]) -> np.ndarray:
results = []
fields = [('data', self.field)]
examples = [Example.fromlist([s], fields) for s in sentences]
dataset = Dataset(examples, fields)
dataloader = BucketIterator(dataset,
self.config.metrics.classifier.batch_size,
repeat=False,
shuffle=False,
device=self.config.device_name)
for batch in dataloader:
scores = self.model(batch.data)
scores = torch.sigmoid(scores)
scores = scores.detach().cpu().numpy().tolist()
results.extend(scores)
return np.mean(results)
def splits(cls, train, valid, batch_size=2, device="cuda"):
train_dataset = train.get("Dataset")
valid_dataset = valid.get("Dataset")
field_names = [field[0] for field in train.get("Field")]
device = device
train_iter, valid_iter = BucketIterator.splits(
(train_dataset, valid_dataset),
batch_size=batch_size,
device=device,
sort_key=lambda x: len(vars(x)[field_names[0]]),
sort_within_batch=True)
train_dataloader = MiniBatchWrapper(train_iter, field_names[0],
field_names[1])
valid_dataloader = MiniBatchWrapper(valid_iter, field_names[0],
field_names[1])
return train_dataloader, valid_dataloader
def prepare_data(self):
self.text_field = Field(sequential=True,
fix_length=200,
include_lengths=True)
self.label_field = LabelField()
train_val, test = IMDB.splits(self.text_field, self.label_field)
random.seed(42)
train, val = train_val.split(random_state=random.getstate())
self.text_field.build_vocab(
train, vectors=GloVe()) #vectors=FastText('simple'))
self.label_field.build_vocab(train)
self.train_iter, self.test_iter, self.val_iter = BucketIterator.splits(
(train, test, val), batch_size=self.batch_size)
self.train_iter.sort_within_batch = True
self.val_iter.sort_within_batch = True
def load_or_generate_dataset(batch_size=128):
download_spacy_models() # TODO check if need to download then do it
spacy_en, spacy_de = load_tokenize_models()
SRC = Field(tokenize=partial(tokenize_de, spacy_de=spacy_de),
init_token=START_TOKEN,
eos_token=END_TOKEN,
lower=True)
TRG = Field(tokenize=partial(tokenize_en, spacy_en=spacy_en),
init_token=START_TOKEN,
eos_token=END_TOKEN,
lower=True)
train_data, valid_data, test_data = Multi30k.splits(exts=('.de', '.en'),
fields=(SRC, TRG))
logging.debug(f"Number of training examples: {len(train_data.examples)}")
logging.debug(f"Number of validation examples: {len(valid_data.examples)}")
logging.debug(f"Number of testing examples: {len(test_data.examples)}")
SRC.build_vocab(train_data, min_freq=2)
TRG.build_vocab(train_data, min_freq=2)
logging.debug(f"Unique tokens in source (de) vocabulary: {len(SRC.vocab)}")
logging.debug(f"Unique tokens in target (en) vocabulary: {len(TRG.vocab)}")
train_iterator, valid_iterator, test_iterator = BucketIterator.splits(
(train_data, valid_data, test_data),
batch_size=batch_size,
device=get_available_device())
# TODO class
return {
'train_data': train_iterator,
'valid_data': valid_iterator,
'test_data': test_iterator,
'n_src_words': len(SRC.vocab),
'n_trg_words': len(TRG.vocab),
'trg_pad_idx': TRG.vocab.stoi[TRG.pad_token],
'src_vocab': SRC.vocab,
'trg_vocab': TRG.vocab,
'src_field': SRC,
'trg_field': TRG,
}
def load_iters(batch_size=1, device="cpu", data_path='data'):
TEXT = data.Field(batch_first=True, include_lengths=True, lower=True)
LABEL = data.LabelField(batch_first=True)
fields = {
'sentence1': ('premise', TEXT),
'sentence2': ('hypothesis', TEXT),
'gold_label': ('label', LABEL)
}
train_data, dev_data, test_data = data.TabularDataset.splits(
path=data_path,
train='snli_1.0_train.jsonl',
validation='snli_1.0_dev.jsonl',
test='snli_1.0_test.jsonl',
format='json',
fields=fields,
filter_pred=lambda ex: ex.label !=
'-' # filter the example which label is '-'(means unlabeled)
)
TEXT.build_vocab(train_data, vectors=GloVe(name='6B', dim=300))
LABEL.build_vocab(train_data)
# TEXT.build_vocab(train_data, vectors=vectors, unk_init=torch.Tensor.normal_)
train_iter, dev_iter = BucketIterator.splits(
(train_data, dev_data),
batch_sizes=(batch_size, batch_size),
device=device,
sort_key=lambda x: len(x.premise) + len(x.hypothesis), #按 合并文本长度 排序，
sort_within_batch=True, #方便后面pytorch lstm进行pack和pad
repeat=False,
shuffle=True)
test_iter = Iterator(test_data,
batch_size=batch_size,
device=device,
sort=False,
sort_within_batch=False,
repeat=False,
shuffle=False)
return train_iter, dev_iter, test_iter, TEXT, LABEL
def init(model_config, device='cpu'):
logging.critical("[CRITICAL] %s device is selected" % device)
logging.info('[INFO] Using directory %s for the translation pair with filename %s' % (os.path.abspath(model_config['global']['dataset_path']), model_config['global']['translate_pair']))
#initialize the field for src language
src_field = Field(tokenize = english_tokenizer,
init_token = '<sos>',
eos_token = '<eos>',
lower = True)
#initialize the field for trg language
trg_field = Field(tokenize = hindi_tokenizer,
init_token = '<sos>',
eos_token = '<eos>',
lower = True)
train_data, valid_data, test_data = load_datasets(model_config['global']['dataset_path'], model_config['global']['dataset_file_names'], model_config['global']['translate_pair'], model_config['global']['lang_extensions'], [src_field, trg_field])
#initialize the vocabulary
src_field.build_vocab(train_data, min_freq = 1)
trg_field.build_vocab(train_data, min_freq = 1)
#display dataset stats
print_dataset_statistics(train_data, valid_data, test_data, model_config['global']['lang_extensions'], [src_field, trg_field])
model = create_seq2seq_model(model_config, len(src_field.vocab), len(trg_field.vocab), device)
optimizer = optim.Adam(model.parameters())
#defining the loss function
loss_function = nn.CrossEntropyLoss(ignore_index = trg_field.vocab.stoi[trg_field.pad_token])
logging.info(model.apply(init_weights))
logging.info('[INFO] Model has %s trainable parameters' % (count_parameters(model)))
logging.info('[INFO] About to start the primary training loop')
train_iterator, valid_iterator, test_iterator = BucketIterator.splits(
(train_data, valid_data, test_data),
batch_size = model_config['global']['batch_size'],
device = device)
cache_file_name = "%s-%s-%s-epoch-%s.pt" % (model_config['global']['name'], model_config['global']['lang_extensions'][0], model_config['global']['lang_extensions'][1], model_config['global']['epochs'])
cache_file_path = os.path.join(model_config['global']['cache_path'], cache_file_name)
stats = execute_training_loop(model, train_iterator, valid_iterator, loss_function, optimizer, model_config['global']['clip_value'], src_field, trg_field, epochs=model_config['global']['epochs'], model_cache_path=os.path.abspath(cache_file_path))
stats_file_name = "%s-%s-%s-epoch-%s-stats.pickle" % (model_config['global']['name'], model_config['global']['lang_extensions'][0], model_config['global']['lang_extensions'][1], model_config['global']['epochs'])
store_object(stats, os.path.join(model_config['global']['cache_path'], stats_file_name))
logging.info("[INFO] loading the model %s" % (cache_file_name))
model.load_state_dict(torch.load(os.path.abspath(cache_file_path)))
test_loss, test_bleu = evaluate_model(model, test_iterator, loss_function, src_field, trg_field)
logging.info(f'[INFO] | Test Loss: {test_loss:.3f} Test Bleu: {test_bleu:.3f} | Test PPL: {math.exp(test_loss):7.3f} |')
def main():
train, test, field = dataset_reader(train=True, stop=900000)
evl, _ = dataset_reader(train=False, fields=field)
field.build_vocab(train, evl)
_, evl_iter = BucketIterator.splits((train, evl),
batch_sizes=(1024, 1024),
device=device,
sort_within_batch=False,
repeat=False,
sort=False)
model = RNNCNN(num_embeddings=len(field.vocab),
embedding_dim=256).to(device)
model.load_state_dict(torch.load('model/rnn_cnn_model.pkl'))
with open('data/rnn_cnn_result.txt', 'w+') as f:
f.write('')
model.eval()
with torch.no_grad():
for i, data in tqdm.tqdm(enumerate(evl_iter),
total=evl_iter.__len__()):
inputs = torch.cat((
data.plat_form,
data.biz_type,
data.payed_day,
data.payed_hour,
data.cate1_id,
data.cate2_id,
data.cate3_id,
data.preselling_shipped_day,
data.seller_uid_field,
data.company_name,
data.rvcr_prov_name,
data.rvcr_city_name,
),
dim=1)
outputs = model(inputs, 'test', field)
day = outputs * 3 + 3
with open('data/rnn_cnn_result.txt', 'a+') as f:
for b in range(day.size(0)):
sign_day = str(float(day[b]))
f.write(sign_day + '\n')
def main():
setup_seed(2020)
POST = Field(tokenize=tokenize_post, init_token='<sos>', eos_token='<eos>')
QUERY = Field(tokenize=tokenize_query,
init_token='<sos>',
eos_token='<eos>')
RESP = Field(tokenize=tokenize_resp, init_token='<sos>', eos_token='<eos>')
# first 'post' is the key in loaded json, second 'post' is the key in batch
fields = {
'post': ('post', POST),
'query': ('query', QUERY),
'resp': ('resp', RESP)
}
train_data, valid_data, test_data = TabularDataset.splits(
path='./data',
train='train.json',
validation='valid.json',
test='test.json',
format='json',
fields=fields)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
POST.build_vocab(train_data, min_freq=1)
QUERY.build_vocab(train_data, min_freq=1)
RESP.build_vocab(train_data, min_freq=1)
# print(POST.vocab.__dict__)
batch_size = 10
train_iter, val_iter, test_iter = BucketIterator.splits(
(train_data, valid_data, test_data),
batch_sizes=(batch_size, batch_size, batch_size),
device=device,
sort_key=lambda x: len(x.post),
sort_within_batch=True,
# sort according to the len, for padding in LSTM
repeat=False)
cnt = 0
for i, batch in enumerate(train_iter):
if cnt == 0:
post = batch.post
print(post.size())
print(batch.post)
cnt += 1
def eval_dataset(model,
dataset,
batch_size,
loss_fn,
device,
text_embeds,
optimizer,
stage,
csv_file,
update_grad=False):
cols = ['loss', 'acc']
(iterator, ) = BucketIterator.splits(datasets=(dataset, ),
batch_sizes=[batch_size],
device=device,
shuffle=True)
metrics = []
for batch in iterator:
(prem_embeds, prem_lens, hyp_embeds, hyp_lens,
labels) = batch_cols(batch, text_embeds)
predictions = model.forward(prem_embeds, prem_lens, hyp_embeds,
hyp_lens)
loss = loss_fn(predictions, labels)
acc = accuracy(predictions, labels)
vals = [loss, acc]
stats = get_stats(cols, vals)
metrics.append(stats)
print(
yaml.dump({
stage: {
k: round(i, 3) if isinstance(i, float) else i
for k, i in stats.items()
}
}))
if update_grad:
optimizer.zero_grad()
loss.backward()
optimizer.step()
df = pd.DataFrame(metrics, columns=cols)
df.to_csv(csv_file)
(loss, acc) = list(df.mean())
return (loss, acc, df)
def _set_data(self, dataset: str, batch_size: int):
train_ds, valid_ds, test_ds, TEXT = load_real_dataset(dataset)
self._reverse = lambda x: TEXT.reverse(x)
self.first_token = TEXT.vocab.stoi["<sos>"]
self.vocab_size = len(TEXT.vocab)
self.length = TEXT.max_length
self.train_batchmanager, self.val_batchmanager, self.test_batchmanager = BucketIterator.splits(
(train_ds, valid_ds, test_ds),
batch_sizes=(batch_size, 2 * batch_size, 2 * batch_size),
device="cpu",
sort_key=lambda x: len(x.text),
sort_within_batch=False,
repeat=False)
if self.evaluate_test:
self.eval_test = MetricsEval(test_ds, TEXT, "Test")
if self.evaluate_valid:
self.eval_valid = MetricsEval(valid_ds, TEXT, "Valid")
def __init__(self, data_path, batch_size, device, embedding=None):
self.TEXT = Field(sequential=True, tokenize=self.tokenize, lower=True, include_lengths = True)
self.LABEL = Field(sequential=False, use_vocab=False)
self.datafield = [("id", None), ("sentence", self.TEXT), ("label", self.LABEL)]
train_data, dev_data_a, test_data_a = TabularDataset.splits(path=data_path, train='train.csv',
validation="dev_a.csv", test="test_a.csv",
format='csv', skip_header=True, fields=self.datafield)
dev_data_b, test_data_b = TabularDataset.splits(path=data_path, validation="dev_b.csv", test="test_b.csv",
format='csv', skip_header=True, fields=self.datafield)
self.TEXT.build_vocab(train_data)
if embedding:
self.TEXT.vocab.load_vectors(embedding)
self.embedding = self.TEXT.vocab.vectors.to(device)
else:
self.embedding = None
# self.train_iter, self.val_iter = BucketIterator(train_data, batch_size=batch_size, device=device,
# sort_key=lambda x: len(x.sentence), sort_within_batch=True)
self.train_iter, self.dev_a_iter, self.test_a_iter, self.dev_b_iter, self.test_b_iter = \
BucketIterator.splits((train_data, dev_data_a, test_data_a, dev_data_b, test_data_b), batch_size=batch_size,
sort_within_batch=True, sort_key=lambda x: len(x.sentence), device=device)
def predict(self, input_text):
input_fields = [('id', self.ID), ('text', self.TEXT),
('label', self.LABEL)]
input_example = Example()
input_example = input_example.fromlist([1, input_text, 1],
input_fields)
input_dataset = Dataset([input_example], input_fields)
input_iter = BucketIterator(input_dataset,
batch_size=1,
device=self.device,
repeat=False)
pred, attn, prob = test_model(self.encoder, self.classifier, 1,
input_iter)
pred = pred[0][1]
prob = prob[0][1]
attn = attn[0].tolist()
attn = [i[0] for i in attn]
return pred, prob, attn
def get_iterator(datafields, data_dir, bs):
trn, vld = TabularDataset.splits(
path=data_dir,
train='train.csv', validation="val.csv",
format='csv',
skip_header=True,
fields=datafields)
_, label_field = datafields[0]
_, text_field = datafields[1]
text_field.build_vocab(trn, vectors="glove.6B.100d")
label_field.build_vocab(trn)
train_iterator, valid_iterator = BucketIterator.splits(
(trn, vld),
batch_size=bs,
sort_key=lambda x: len(x.text),
device=device)
return train_iterator, valid_iterator
def load_iters(batch_size=32, device="cpu", data_path='data', vectors=None):
TEXT = data.Field(lower=True, batch_first=True, include_lengths=True)
LABEL = data.LabelField(batch_first=True)
train_fields = [(None, None), (None, None), ('text', TEXT),
('label', LABEL)]
test_fields = [(None, None), (None, None), ('text', TEXT)]
train_data = data.TabularDataset.splits(
path=data_path,
train='train.tsv',
format='tsv',
fields=train_fields,
skip_header=True)[0] # return is a tuple.
test_data = data.TabularDataset.splits(path='data',
train='test.tsv',
format='tsv',
fields=test_fields,
skip_header=True)[0]
TEXT.build_vocab(train_data.text, vectors=vectors)
LABEL.build_vocab(train_data.label)
train_data, dev_data = train_data.split([0.8, 0.2])
train_iter, dev_iter = BucketIterator.splits(
(train_data, dev_data),
batch_sizes=(batch_size, batch_size),
device=device,
sort_key=lambda x: len(x.text),
sort_within_batch=True,
repeat=False,
shuffle=True)
test_iter = Iterator(test_data,
batch_size=batch_size,
device=device,
sort=False,
sort_within_batch=False,
repeat=False,
shuffle=False)
return train_iter, dev_iter, test_iter, TEXT, LABEL
def getdata():
europarl_en = open('europarl-v7.cs-en.en',
encoding='utf-8').read().split('\n')
europarl_cs = open('europarl-v7.cs-en.cs',
encoding='utf-8').read().split('\n')
EN_TEXT = Field(tokenize=tokenize_en)
CS_TEXT = Field(tokenize=tokenize_cs,
init_token="<sos>",
eos_token="<eos>")
raw_data = {
'English': [line for line in europarl_en],
'Czech': [line for line in europarl_cs]
}
df = pd.DataFrame(raw_data, columns=["English", "Czech"])
# remove very long sentences and sentences where translations are
df['en_len'] = df['English'].str.count(' ')
df['cs_len'] = df['Czech'].str.count(' ')
df = df.query('cs_len < 80 & en_len < 80')
df = df.query('cs_len < en_len * 1.5 & cs_len * 1.5 > en_len')
train, val = train_test_split(df, test_size=0.1)
train.to_csv("train.csv", index=False)
val.to_csv("val.csv", index=False)
data_fields = [('English', EN_TEXT), ('Czech', CS_TEXT)]
train, val = TabularDataset.splits(path='./',
train='train.csv',
validation='val.csv',
format='csv',
fields=data_fields)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
EN_TEXT.build_vocab(train, min_freq=2)
CS_TEXT.build_vocab(train, min_freq=2)
BATCH_SIZE = 16
INPUT_DIM = len(EN_TEXT.vocab)
OUTPUT_DIM = len(CS_TEXT.vocab)
PAD_IDX = EN_TEXT.vocab.stoi['<pad>']
train_iterator, valid_iterator = BucketIterator.splits(
(train, val), batch_size=BATCH_SIZE, device=device)
return train_iterator, valid_iterator, INPUT_DIM, OUTPUT_DIM, PAD_IDX
def load_dataset(db_name, batch_size):
"""
Load the csv datasets into torchtext files
Inputs:
db_name (string)
The name of the dataset. This name must correspond to the folder name.
batch_size
The batch size
"""
print "Loading " + db_name + "..."
i = 1
print('num', i)
tokenize = lambda x: x.split()
TEXT = Field(sequential=True, tokenize=tokenize, lower=True)
LABEL = Field(sequential=False, use_vocab=False)
tv_datafields = [("sentence", TEXT), ("label", LABEL)]
trn, vld = TabularDataset.splits(
path=DATA_ROOT + db_name, # the root directory where the data lies
train='train.csv',
validation="test.csv",
format='csv',
skip_header=False,
fields=tv_datafields)
TEXT.build_vocab(trn)
print "vocab size: %i" % len(TEXT.vocab)
train_iter, val_iter = BucketIterator.splits(
(trn, vld),
batch_sizes=(batch_size, batch_size),
device=-1, # specify dont use gpu
sort_key=lambda x: len(x.sentence), # sort the sentences by length
sort_within_batch=False,
repeat=False)
return train_iter, val_iter, len(TEXT.vocab)
def get_iterators(opt, device=None):
"""
Get dataset iterator and necessary fields information
:param opt: opt from argparser.
:param device: device to create the data.
:return: train_iter, test_iter, dataset.fields
"""
import random
random.seed(42)
dataset = load_dataset(opt.use_cws)
dataset.fields["text"].build_vocab(dataset)
dataset.fields["author"].build_vocab(dataset)
dataset.fields["book"].build_vocab(dataset)
train, test = dataset.split(split_ratio=0.7)
train_iter, test_iter = BucketIterator.splits(
(train, test), # first one is default to train (shuffle each epoch)
batch_sizes=(opt.train_batch_size, opt.eval_batch_size),
device=device,
sort_key=lambda x: len(x.text),
sort_within_batch=True)
return train_iter, test_iter, dataset.fields
def load_dataset(batch_size):
spacy_de = spacy.load('de')
spacy_en = spacy.load('en')
url = re.compile('(<url>.*</url>)')
def tokenize_de(text):
return [tok.text for tok in spacy_de.tokenizer(url.sub('@URL@', text))]
def tokenize_en(text):
return [tok.text for tok in spacy_en.tokenizer(url.sub('@URL@', text))]
DE = Field(tokenize=tokenize_de, include_lengths=True,
init_token='<sos>', eos_token='<eos>')
EN = Field(tokenize=tokenize_en, include_lengths=True,
init_token='<sos>', eos_token='<eos>')
train, val, test = Multi30k.splits(exts=('.de', '.en'), fields=(DE, EN))
DE.build_vocab(train.src, min_freq=2)
EN.build_vocab(train.trg, max_size=10000)
train_iter, val_iter, test_iter = BucketIterator.splits(
(train, val, test), batch_size=batch_size, repeat=False)
return train_iter, val_iter, test_iter, DE, EN
def create_iterator(
corpus: Corpus,
batch_size: int = 1,
device: str = "cpu",
sort: Optional[bool] = None,
) -> BucketIterator:
if sort is not None:
sort_key: Optional[Callable] = lambda e: len(e.sen)
else:
sort_key = None
iterator = BucketIterator(
dataset=corpus,
batch_size=batch_size,
device=device,
shuffle=False,
sort=sort,
sort_key=sort_key,
)
return iterator
def data_iter(train_data, valid_data, TEXT, LABEL):
train = MyDataset(train_data,
text_field=TEXT,
label_field=LABEL,
test=False)
valid = MyDataset(valid_data,
text_field=TEXT,
label_field=LABEL,
test=False)
TEXT.build_vocab(train)
train_iter, val_iter = BucketIterator.splits(
(train, valid), # 构建数据集所需的数据集
batch_sizes=(100, 100),
# 如果使用gpu，此处将-1更换为GPU的编号
device=device,
# the BucketIterator needs to be told what function it should use to group the data.
sort_key=lambda x: len(x.text),
sort_within_batch=False,
repeat=False)
return train_iter, val_iter
def load_dataset(batch_size, device):
"""
Load the dataset from the files into iterator and initialize the vocabulary
:param batch_size
:param device
:return: source and data iterators
"""
source = Field(tokenize=tokenize_en,
init_token='<sos>',
eos_token='<eos>',
lower=True)
train_data, valid_data, test_data = TranslationDataset.splits(
path=DATA_FOLDER,
exts=(POSITIVE_FILE_EXTENSION, NEGATIVE_FILE_EXTENSION),
fields=(source, source))
source.build_vocab(train_data, min_freq=5)
return source, BucketIterator.splits((train_data, valid_data, test_data),
shuffle=True,
batch_size=batch_size,
device=device)
def load_dataset(batch_size):
spacy_de = spacy.load('de')
spacy_en = spacy.load('en')
url = re.compile('(<url>.*</url>)')
def tokenize_de(text):
return [tok.text for tok in spacy_de.tokenizer(url.sub('@URL@', text))]
def tokenize_en(text):
return [tok.text for tok in spacy_en.tokenizer(url.sub('@URL@', text))]
DE = Field(tokenize=tokenize_de, include_lengths=True,
init_token='<sos>', eos_token='<eos>')
EN = Field(tokenize=tokenize_en, include_lengths=True,
init_token='<sos>', eos_token='<eos>')
train, val, test = Multi30k.splits(exts=('.de', '.en'), fields=(DE, EN))
DE.build_vocab(train.src, min_freq=2)
EN.build_vocab(train.trg, max_size=10000)
train_iter, val_iter, test_iter = BucketIterator.splits(
(train, val, test), batch_size=batch_size, repeat=False)
return train_iter, val_iter, test_iter, DE, EN
def get_data_iterator(batch_size, device):
SRC = Field(tokenize = tokenize_de,
init_token = '<sos>',
eos_token = '<eos>',
lower = True)
TRG = Field(tokenize = tokenize_en,
init_token = '<sos>',
eos_token = '<eos',
lower = True)
train_data, valid_data, test_data = Multi30k.splits(exts = ('.de', '.en'), fields = (SRC, TRG))
SRC.build_vocab(train_data, min_freq = 2)
TRG.build_vocab(train_data, min_freq = 2)
train_iterator, valid_iterator, test_iterator = BucketIterator.splits(
(train_data, valid_data, test_data),
batch_size = batch_size,
device = device)
return train_iterator, valid_iterator, test_iterator, SRC, TRG
def load_dataset(batch_size, device=0):
spacy_de = spacy.load('de')
spacy_en = spacy.load('en')
def tokenize_de(text):
return [tok.text for tok in spacy_de.tokenizer(text)]
def tokenize_en(text):
return [tok.text for tok in spacy_en.tokenizer(text)]
DE = Field(tokenize=tokenize_de, init_token='<sos>', eos_token='<eos>')
EN = Field(tokenize=tokenize_en, init_token='<sos>', eos_token='<eos>')
train, val, test = Multi30k.splits(exts=('.de', '.en'), fields=(DE, EN))
DE.build_vocab(train.src)
EN.build_vocab(train.trg)
train_iter, val_iter, test_iter = BucketIterator.splits(
(train, val, test), batch_size=batch_size, device=device, repeat=False)
return train_iter, val_iter, test_iter, DE, EN
def load_dataset(batch_size, debug=True, shuffle_dataset=True):
spacy_en = spacy.load('en')
def tokenize_en(line):
return [token.text for token in spacy_en.tokenizer(line)]
def tokenize_zh(line):
return [token for token in jieba.cut(line)]
EN = Field(tokenize=tokenize_en,
include_lengths=True,
init_token='<sos>',
eos_token='<eos>')
ZH = Field(tokenize=tokenize_zh,
include_lengths=True,
init_token='<sos>',
eos_token='<eos>')
exts = ['.en', '.zh']
fields = [('src', EN), ('trg', ZH)]
train_dataset = TranslationDataset(train_sentence_path,
exts=exts,
fields=fields)
val_dataset = TranslationDataset(val_sentence_path,
exts=exts,
fields=fields)
print('Datasets Built!')
EN.build_vocab(train_dataset.src, min_freq=2)
ZH.build_vocab(train_dataset.trg, max_size=100000)
print('Vocabularies Built!')
val_iter, *_ = BucketIterator.splits(
(val_dataset, ),
shuffle=shuffle_dataset,
batch_size=batch_size,
repeat=False,
sort_key=lambda x: interleave_keys(len(x.src), len(x.trg)))
print('Training Iterators Built!')
return val_iter, val_dataset, ZH, EN
# multi30k dataloader
train,val,test = datasets.Multi30k.splits(exts=(".en",".de"),fields=(EN,DE),root=data_path)
# wmt14 dataloader (better than using datasets.WMT14.splits since it's slow)
#train,val,test = datasets.TranslationDataset.splits(exts=(".en",".de"),fields=[("src",EN),("trg",DE)],path=os.path.join(data_path,"wmt14"),
# train="train.tok.clean.bpe.32000",validation="newstest2013.tok.bpe.32000",test="newstest2014.tok.bpe.32000")
print("Dataset loaded")
EN.build_vocab(train.src,min_freq=3)
DE.build_vocab(train.trg,max_size=50000)
print("Vocabularies build")
train_iter,val_iter = BucketIterator.splits((train, val),batch_size=3)
test_iter = BucketIterator(test,batch_size=3)
print("Start iterating through data")
for i,batch in enumerate(train_iter):
print(batch.src) # the source language
print(batch.trg) # the target language
break
for i,batch in enumerate(val_iter):
print(batch.src) # the source language
print(batch.trg) # the target language
break
for i,batch in enumerate(test_iter):
skip_header=True,
fields=[('id',None),('text',TEXT),('label',LABEL)],
filter_pred = lambda x: True if len(x.text) > 1 else False)
# 토큰 레벨 문장의 길이가 1 이상인 경우만 허용
TEXT.build_vocab(train_data,min_freq=2)
LABEL.build_vocab(train_data)
# print (TEXT.vocab)
# print (len(TEXT.vocab),len(LABEL.vocab))
# print (TEXT.vocab.itos[:5])
# print (LABEL.vocab.itos)
train_loader, test_loader = BucketIterator.splits((train_data,test_data),sort_key=lambda x:len(x.text), sort_within_batch=True,
repeat=False,shuffle=True,
batch_size=32,device=DEVICE)
for batch in train_loader:
break
EPOCH = 5
BATCH_SIZE = 32
EMBED = 300
KERNEL_SIZES = [3,4,5]
KERNEL_DIM = 100
LR = 0.001
# model = CNNClassifier(len(TEXT.vocab), EMBED, 1, KERNEL_DIM, KERNEL_SIZES)
model = RNN(len(TEXT.vocab), EMBED, KERNEL_DIM, 1, bidirec=False)
