# [29] START
model_path = "uni_joint_2layer_gelu_synth.pth.tar"
# [29] END
# [30] START
model_synth_swe = Trans_model(embedding_size, src_vocab_size, trg_vocab_size,
src_pad_idx, num_heads, num_encoder_layers,
num_decoder_layers, forward_expansion, dropout,
max_len, device, 'gelu')
# [30] END
# [31] START
sp.Load(swedish_model)
criterion = nn.CrossEntropyLoss(ignore_index=sp.pad_id())
optimizer = optim.Adam(model_synth_swe.parameters(), lr=learning_rate)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1.0, gamma=0.95)
# [31] END
# [32] START
if load_model == True:
checkpoint = torch.load(model_path, map_location='cpu')
model_synth_swe.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
model_synth_swe.to(device)
# [32] END
# [33] START
translate_sentence(model_synth_swe, sent1, device, sami_model, swedish_model)
# [33] END
src_vocab_size,
trg_vocab_size,
src_pad_idx,
num_heads,
num_encoder_layers,
num_decoder_layers,
forward_expansion,
dropout,
max_len,
device,
'gelu'
)
sp.Load(swedish_model)
criterion = nn.CrossEntropyLoss(ignore_index=sp.pad_id())
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1.0, gamma=0.95)
load_model = False
if load_model == True:
checkpoint = torch.load(model_path, map_location='cpu')
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
model.to(device)
sentence = "Sámediggi lea sámiid álbmotválljen orgána Norggas."
sentence2 = "Deaŧalaš lea gozihit álgoálbmotoli nationála ja riikkaidgaskasaš forain."
def main(DATA, MAX_LINES, MAX_PADDING, MIN_LEN_SENTENCE, SIZE_VOCAB, SHOW_SENTENCES, LR, EPOCHS):
if DATA == 'europarl':
URLS=["http://www.statmt.org/europarl/v10/training-monolingual/europarl-v10.es.tsv.gz"]
FILES = ["europarl-v10.es.tsv.gz"]
CORPORA = ["europarl-v10.es.tsv"]
elif DATA == 'newscarl':
URLS=[
"http://data.statmt.org/news-crawl/es/news.2007.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2008.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2009.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2010.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2011.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2012.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2013.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2014.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2015.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2016.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2017.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2018.es.shuffled.deduped.gz",
"http://data.statmt.org/news-crawl/es/news.2019.es.shuffled.deduped.gz"
]
FILES=[
"news.2007.es.shuffled.deduped.gz",
"news.2008.es.shuffled.deduped.gz",
"news.2009.es.shuffled.deduped.gz",
"news.2010.es.shuffled.deduped.gz",
"news.2011.es.shuffled.deduped.gz",
"news.2012.es.shuffled.deduped.gz",
"news.2013.es.shuffled.deduped.gz",
"news.2014.es.shuffled.deduped.gz",
"news.2015.es.shuffled.deduped.gz",
"news.2016.es.shuffled.deduped.gz",
"news.2017.es.shuffled.deduped.gz",
"news.2018.es.shuffled.deduped.gz",
"news.2019.es.shuffled.deduped.gz"
]
CORPORA=[
"news.2007.es.shuffled.deduped",
"news.2008.es.shuffled.deduped",
"news.2009.es.shuffled.deduped",
"news.2010.es.shuffled.deduped",
"news.2011.es.shuffled.deduped",
"news.2012.es.shuffled.deduped",
"news.2013.es.shuffled.deduped",
"news.2014.es.shuffled.deduped",
"news.2015.es.shuffled.deduped",
"news.2016.es.shuffled.deduped",
"news.2017.es.shuffled.deduped",
"news.2018.es.shuffled.deduped",
"news.2019.es.shuffled.deduped"
]
print('File download') #---------------------------------------------
for u, f in zip(URLS, FILES):
print(u)
sys.stdout.flush()
if path.exists(f):
print('File already downloaded'.format(f))
else:
wget.download(u, './'+f)
print('Unzipping {}'.format(f)) #----------------------------------------------
for f, c in zip(FILES, CORPORA):
print(f)
sys.stdout.flush()
if path.exists(f):
with gzip.open(f, 'rb') as f_in:
with open(c, 'wb') as f_out:
shutil.copyfileobj(f_in, f_out)
else:
print('File already unzipped')
print('Join all into one') #---------------------------------------------------
sys.stdout.flush()
with open('corpus.es', 'wb') as outfile:
for fname in CORPORA:
print('Joining file:', fname)
sys.stdout.flush()
with open(fname, 'rb') as infile:
for line in infile:
outfile.write(line)
print('Delete auxiliar files') #-----------------------------------------------
sys.stdout.flush()
for f in CORPORA:
os.remove(f)
print('Reduced File') #--------------------------------------------------------
sys.stdout.flush()
with open('corpus.es', 'rb') as cfile:
with open('corpus.reduced.es', 'wb') as rfile:
count = 0
while count < MAX_LINES:
line = cfile.readline()
rfile.write(line)
count += 1
print('Number of lines in the reduced file:', count)
print('Read Data') #----------------------------------------------------------
sys.stdout.flush()
FILE = 'corpus.reduced.es'
data = []
with open(FILE, 'rb') as corpus_file:
Lines = corpus_file.readlines()
for line in Lines:
data.append(line)
print('Preprocessing the Data') #----------------------------------------------
sys.stdout.flush()
print(data[50])
print(sentence_to_words(data[50]))
sys.stdout.flush()
preprocessed_data = []
perc = 0
for ind, s in enumerate(data):
words = sentence_to_words(s)
if len(words) >= MIN_LEN_SENTENCE and len(words) <= MAX_PADDING :
preprocessed_data.append(words)
if ind > perc:
print('{}/{} sentences preprocessed'.format(perc, len(data)))
sys.stdout.flush()
perc += SHOW_SENTENCES
print('Length data readed: ', len(data))
print('Length after preprocessing: ', len(preprocessed_data))
print(preprocessed_data[50])
sys.stdout.flush()
print('Build Dictionary') #---------------------------------------------------
sys.stdout.flush()
FED = 20 #nmber of elements we will show
word_dict, complete_dict = build_dict(preprocessed_data, SIZE_VOCAB)
list_dict = [key for key in word_dict.keys()]
print('First {} elements of the dictionary: {}'.format(FED, list_dict[:FED]))
sys.stdout.flush()
print('Selected {}/{} words'.format(SIZE_VOCAB, len(complete_dict)))
sys.stdout.flush()
# We will add the XXX as 0 in the dictionary, so when a custom sentence is inputted, the XXX will mark the word the model has to guess
word_dict['XXX'] = 0
word_dict['PAD'] = 1
word_dict['INFREQ'] = 2
print('Convert and Pad') #-----------------------------------------------------
sys.stdout.flush()
int_data, int_data_len = convert_and_pad_data(word_dict, preprocessed_data, MAX_PADDING)
print(int_data[50], int_data_len[50])
sys.stdout.flush()
print('Extract Word') #-------------------------------------------------------
sys.stdout.flush()
# Check there is no sentence with all 2's
int_data_pre = [d for d, lend in zip(int_data, int_data_len) if len(set(d[:lend])) > 1]
len_data_pre = [lend for d, lend in zip(int_data, int_data_len) if len(set(d[:lend])) > 1]
print('{} of the {} sentences were only 2\'s'.format(len(int_data)-len(int_data_pre), len(int_data)))
sys.stdout.flush()
masked_data = []
word_masked = []
masked_data = int_data_pre.copy()
for idx, (sentence, len_sentence) in enumerate(zip(int_data_pre, len_data_pre)):
acceptable_value = False
while acceptable_value == False:
idx_word = random.randint(0, len_sentence-1)
if int_data_pre[idx][idx_word] != 2:
acceptable_value = True
word_masked.append(int_data_pre[idx][idx_word]) #save the word extracted
masked_data[idx][idx_word] = 0 #put this word to 0
print(masked_data[50])
print(word_masked[50])
sys.stdout.flush()
print('Split Train, Valid and Test') #-----------------------------------------
sys.stdout.flush()
train_x, valid_x, train_y, valid_y, train_len, valid_len = train_test_split(masked_data, word_masked, len_data_pre, test_size=0.25, random_state=42)
valid_x, test_x, valid_y, test_y, valid_len, test_len = train_test_split(valid_x, valid_y, valid_len, test_size=0.4, random_state=42)
print('train: ', len(train_x), len(train_y), len(train_len))
print('valid: ', len(valid_x), len(valid_y), len(valid_len))
print('test: ', len(test_x), len(test_y), len(test_len))
sys.stdout.flush()
print('Cleaning Variables') #--------------------------------------------------
sys.stdout.flush()
preprocessed_data = None
list_dict = None
int_data = None
int_data_len = None
int_data_pre = None
len_data_pre = None
masked_data = None
word_masked = None
data = None
Lines = None
complete_dict = None
# Training ------------------------------------------------------------------
BATCH_SIZE = 128
d_model = 256
heads = 8
N = 6
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
sys.stdout.flush()
print('Preparing Input Masks')
sys.stdout.flush()
msk_input_loader = create_mask(train_x, BATCH_SIZE)
valid_msk_input_loader = create_mask(valid_x, BATCH_SIZE)
print('Preparing Train Data Loaders')
sys.stdout.flush()
train_torch_x = torch.tensor(train_x).clone()
train_torch_y = torch.tensor(train_y).clone()
train_sample_ds = torch.utils.data.TensorDataset(train_torch_x, train_torch_y)
train_loader = torch.utils.data.DataLoader(train_sample_ds, batch_size=BATCH_SIZE)
print('Preparing Validation Data Loaders')
sys.stdout.flush()
valid_torch_x = torch.tensor(valid_x).clone()
valid_torch_y = torch.tensor(valid_y).clone()
valid_sample_ds = torch.utils.data.TensorDataset(valid_torch_x, valid_torch_y)
valid_loader = torch.utils.data.DataLoader(valid_sample_ds, batch_size=BATCH_SIZE)
print('Initialize Model')
sys.stdout.flush()
model = Transformer(SIZE_VOCAB, d_model, N, heads).to(device)
for p in model.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
print('Define Loss and Optimizer')
sys.stdout.flush()
loss_function = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=LR, betas=(0.9, 0.98), eps=1e-9)
mask_sample_ds = None
train_torch_x = None
train_torch_y = None
train_sample_ds = None
print('Going to train')
sys.stdout.flush()
losses, valid_losses = train(model, train_loader, msk_input_loader, valid_loader, valid_msk_input_loader, EPOCHS, optimizer, loss_function, device)
# Testing ------------------------------------------------------------------
print('Preparing Input Masks')
msk_test_input_loader = create_mask(test_x, BATCH_SIZE)
print('Preparing Data Loaders')
test_torch_x = torch.tensor(test_x).clone()
test_torch_y = torch.tensor(test_y).clone()
test_sample_ds = torch.utils.data.TensorDataset(test_torch_x, test_torch_y)
test_loader = torch.utils.data.DataLoader(test_sample_ds, batch_size=BATCH_SIZE)
test_torch_x = None
test_torch_y = None
test_sample_ds = None
print('Going to test')
test_loss = evaluate(model, test_loader, msk_test_input_loader, loss_function, BATCH_SIZE, device)
print(test_loss)
# Custom Sentence ------------------------------------------------------------
test_sentences = ["Ha habido una XXX en Colombia durant la presentación del presidente",
"Todas las tropas han sido XXX a America",
"Estaba pensando que quizas XXX deberías hacerlo",
"Todo lo que llevo esta dentro de mí XXX"]
for custom_sentence in test_sentences:
resulting_word = guess_word(custom_sentence)
print('Initial Sentence: \t {}'.format(custom_sentence))
print('Word Guessed: \t\t {}'.format(resulting_word))