def evaluate(self, ptb_path='PTB'):
raw_data = load_dataset(ptb_path)
train_data, dev_data, test_data, word_to_id, id_to_word = raw_data
print(datetime_str(), 'Training model ...')
model = train(train_data, word_to_id, id_to_word)
print(datetime_str(), 'Training model finished.')
unigram_probs = self.train_unigram_model(train_data, word_to_id)
allpreds = [['prev', 'pred1', 'pred2', 'pred3',
'true_prob', 'true_rank', 'kl_uniform', 'kl_unigram']]
print(datetime_str(), 'Testing model ...')
allpreds.extend(self.predict_probs(model, id_to_word, train_data, unigram_probs, 'train'))
allpreds.extend(self.predict_probs(model, id_to_word, dev_data, unigram_probs, 'valid'))
allpreds.extend(self.predict_probs(model, id_to_word, test_data, unigram_probs, 'test'))
print(datetime_str(), 'Testing model finished.')
save_preds(allpreds, preds_fname=PREDS_FNAME)
scores = score_preds(PREDS_FNAME, ptb_path)
for method_name, method_scores in scores.items():
for metric, value in method_scores.items():
print('{} {}: {}'.format(method_name.capitalize(), metric, value))
print()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--ptb_path', default='PTB', help='Path to PTB data')
args = parser.parse_args()
raw_data = load_dataset(args.ptb_path)
train_data, dev_data, test_data, word_to_id, id_to_word = raw_data
model = train(train_data, word_to_id, id_to_word)
unigram_probs = train_unigram_model(train_data, word_to_id)
allpreds = [[
'prev', 'pred1', 'pred2', 'pred3', 'true_prob', 'true_rank',
'kl_uniform', 'kl_unigram'
]]
allpreds.extend(
predict_probs(model, id_to_word, train_data, unigram_probs, 'train'))
allpreds.extend(
predict_probs(model, id_to_word, dev_data, unigram_probs, 'valid'))
allpreds.extend(
predict_probs(model, id_to_word, test_data, unigram_probs, 'test'))
save_preds(allpreds, preds_fname=PREDS_FNAME)
scores = score_preds(PREDS_FNAME, args.ptb_path)
for method_name, method_scores in scores.items():
for metric, value in method_scores.items():
print('{} {}: {}'.format(method_name.capitalize(), metric, value))
print()
def sampling(self, size,
ptb_path='PTB',
start_text=None,
batch_size=20,
temperature=1.0,
pretrained_model=None):
"""
Parameters:
size (int): - number of generated tokens
start_text (str): - list of tokens separated by a space
pretrained_model (str): - path to pretrained model
Returns:
generated text (str)
"""
raw_data = load_dataset(ptb_path)
train_data, dev_data, test_data, word_to_id, id_to_word = raw_data
if pretrained_model is None:
print(datetime_str(), 'Training model ...')
model = train(train_data, word_to_id, id_to_word)
print(datetime_str(), 'Training model finished.')
else:
model = train(train_data, word_to_id, id_to_word, pretrained_model=pretrained_model)
unigram_probs = self.train_unigram_model(train_data, word_to_id)
generated_texts = ancestral_sampling_batch(model,
unigram_probs,
word_to_id,
id_to_word,
size=size,
batch_size=batch_size,
start_text=start_text,
temperature=temperature)
print('\n'.join(generated_texts))
def main(job_id, params):
print params
trainerr, validerr, testerr = train(saveto=params['model'][0],
reload_=params['reload'][0],
dim_word=params['dim_word'][0],
dim=params['dim'][0],
n_words=params['n-words'][0],
decay_c=params['decay-c'][0],
lrate=params['learning-rate'][0],
optimizer=params['optimizer'][0],
maxlen=30,
batch_size=32,
valid_batch_size=16,
validFreq=5000,
dispFreq=10,
saveFreq=1000,
sampleFreq=10,
use_dropout=params['use-dropout'][0])
return validerr
def main(job_id, params):
print params
validerr = train(
saveto=params['model'][0],
reload_=params['reload'][0],
dim_word=params['dim_word'][0],
dim=params['dim'][0],
n_words=params['n-words'][0],
decay_c=params['decay-c'][0],
lrate=params['learning-rate'][0],
optimizer=params['optimizer'][0],
maxlen=30,
batch_size=32,
valid_batch_size=16,
validFreq=5000,
dispFreq=10,
saveFreq=1000,
sampleFreq=1000,
dataset='staff.data.train.tok',
valid_dataset='staff.data.valid.tok',
dictionary='staff.data.train.pkl',
use_dropout=params['use-dropout'][0])
return validerr
def main(job_id, params):
print params
validerr = train(
saveto=params['model'][0],
reload_=params['reload'][0],
dim_word=params['dim_word'][0],
dim=params['dim'][0],
n_words=params['n-words'][0],
decay_c=params['decay-c'][0],
lrate=params['learning-rate'][0],
optimizer=params['optimizer'][0],
maxlen=30,
batch_size=32,
valid_batch_size=16,
validFreq=5000,
dispFreq=10,
saveFreq=1000,
sampleFreq=5,
dataset='/data/lisatmp4/anirudhg/ptb/train.txt',
valid_dataset='/data/lisatmp4/anirudhg/ptb/valid.txt',
dictionary='/data/lisatmp4/anirudhg/ptb/dictionary_pbt.pkl',
use_dropout=params['use-dropout'][0])
return validerr
def main(job_id, params):
print params
validerr = train(
saveto=params['model'][0],
reload_=params['reload'][0],
dim_word=params['dim_word'][0],
dim=params['dim'][0],
n_words=params['n-words'][0],
decay_c=params['decay-c'][0],
lrate=params['learning-rate'][0],
optimizer=params['optimizer'][0],
maxlen=30,
batch_size=32,
valid_batch_size=16,
validFreq=5000,
dispFreq=10,
saveFreq=1000,
sampleFreq=1000,
dataset='/ichec/work/dl4mt_data/nec_files/wiki.tok.txt.gz',
valid_dataset='/ichec/work/dl4mt_data/nec_files/newstest2011.en.tok',
dictionary='/ichec/work/dl4mt_data/nec_files/wiki.tok.txt.gz.pkl',
use_dropout=params['use-dropout'][0])
return validerr
import argparse
import logging
import pprint
import config_lm
from lm import train
logger = logging.getLogger(__name__)
if __name__ == '__main__':
# Get the arguments
parser = argparse.ArgumentParser()
parser.add_argument("--proto", help="Prototype config to use for model configuration")
parser.add_argument("--platoon", action='store_true', help="Prototype config to use for model configuration")
args = parser.parse_args()
config = getattr(config_lm, args.proto)()
config['platoon'] = args.platoon
config['name'] = args.proto
logger.info("Model options:\n{}".format(pprint.pformat(config)))
train(**config)
import numpy
from lm import train
trainerr, validerr, testerr = train(
train_path='../data/morph/finnish.clean.train10k',
validation_path='../data/morph/finnish.clean.test',
dictionary_path='../data/morph/morph.vocab',
model_path='/l/senarvi/theano-rnnlm/model-train10k-lstm.npz',
reload_state=True,
dim_word=256,
dim=1024,
n_words=30000,
decay_c=0.,
lrate=0.0001,
optimizer='adam',
maxlen=30,
batch_size=32,
valid_batch_size=16,
validFreq=5000,
dispFreq=20,
saveFreq=40,
sampleFreq=20)
print(validerr)
import numpy
from lm import train
trainerr, validerr, testerr = train(train_path='../data/morph/finnish.clean.train10k',
validation_path='../data/morph/finnish.clean.test',
dictionary_path='../data/morph/morph.vocab',
model_path='/l/senarvi/theano-rnnlm/model-train10k-lstm.npz',
reload_state=True,
dim_word=256,
dim=1024,
n_words=30000,
decay_c=0.,
lrate=0.0001,
optimizer='adam',
maxlen=30,
batch_size=32,
valid_batch_size=16,
validFreq=5000,
dispFreq=20,
saveFreq=40,
sampleFreq=20)
print(validerr)
