def _write_lm_encodings(lm_data_shard_cfg, device):
data = get_data(
word_emb_data_path_prefix=GLOVE_PREPROC_WITH_UNKS_PATH_PREFIX,
tokenized_trn_json_path=TOKENIZED_TRN_JSON_PATH,
tokenized_dev_json_path=TOKENIZED_DEV_JSON_PATH,
max_ans_len=MAX_ANS_LEN,
max_ctx_len=MAX_CTX_LEN)
write_lm_data_shard(data, lm_data_shard_cfg, device)
def _main(config, config_idx):
base_filename = config.name + '_cfg' + str(config_idx)
logger = set_up_logger('logs/' + base_filename + '.log')
title = '{}: {} ({}) config index {}'.format(__file__, config.name,
config.desc, config_idx)
logger.info('START ' + title + '\n\n{}\n'.format(config))
data = get_data(config)
if config.device != 'cpu':
assert 'theano' not in sys.modules
import theano.sandbox.cuda
theano.sandbox.cuda.use(config.device)
from model import get_model
model = get_model(config, data)
if not config.is_train:
if config.tst_load_model_path and not model.load_if_exists(
config.tst_load_model_path):
raise AssertionError('Failed loading model weights from {}'.format(
config.tst_load_model_path))
ans_hats = _tst_epoch(config, model, data)
write_test_predictions(ans_hats, config.tst_prd_json_path)
logger.info('END ' + title)
return
# Training loop
epoch_results = []
max_em = -np.inf
max_f1 = -np.inf
np_rng = np.random.RandomState(config.seed // 2)
for epoch in range(1, config.max_num_epochs + 1):
trn_loss, trn_acc, trn_samples_per_sec = _trn_epoch(
config, model, data, epoch, np_rng)
dev_min_loss, dev_prx_loss, dev_max_acc, dev_prx_acc, dev_em, dev_f1 = _dev_epoch(
config, model, data)
if dev_em > max_em:
model.save('models/' + base_filename + '_best_em.pkl')
max_em = dev_em
if dev_f1 > max_f1:
model.save('models/' + base_filename + '_best_f1.pkl')
max_f1 = dev_f1
if epoch % 5 == 0:
model.save('models/' + base_filename +
'_e{:03d}.pkl'.format(epoch))
epoch_results.append(
EpochResult(trn_loss, trn_acc, dev_min_loss, dev_prx_loss,
dev_max_acc, dev_prx_acc, dev_em, dev_f1))
if config.plot:
plot_epoch_results(epoch_results, 'logs/' + base_filename + '.png')
logger.info(
'\n\nFinished epoch {} for: (config index {}) (samples/sec: {:<.1f})\n{}\n\nResults:\n{}\n\n'
.format(epoch, config_idx, trn_samples_per_sec,
config.format_compared(),
format_epoch_results(epoch_results)))
logger.info('END ' + title)
def __init__(self, args, logger, out_dir):
self.args = args
self.logger = logger
self.out_dir = out_dir
self.vocab, train_x, test_x, self.overall_maxlen = dataset.get_data(
self.args.domain,
vocab_size=self.args.vocab_size,
maxlen=self.args.maxlen)
self.train_x = sequence.pad_sequences(train_x,
maxlen=self.overall_maxlen)
self.test_x = sequence.pad_sequences(test_x,
maxlen=self.overall_maxlen)
self.vis_path = self.out_dir + "/visualization"
U.mkdir_p(self.vis_path)
def _main(config, config_idx, train):
base_filename = config.name + '_cfg' + str(config_idx)
logger = set_up_logger('logs/' + base_filename + '.log')
title = '{}: {} ({}) config index {}'.format(__file__, config.name, config.desc, config_idx)
logger.info('START ' + title + '\n\n{}\n'.format(config))
data = get_data(config, train)
if config.device != 'cpu':
assert 'theano' not in sys.modules
import theano.sandbox.cuda
theano.sandbox.cuda.use(config.device)
from model import get_model
model = get_model(config, data)
if not train:
assert config.tst_load_model_path
if not model.load(config.tst_load_model_path):
raise AssertionError('Failed loading model weights from {}'.format(config.tst_load_model_path))
ans_hats = _tst_epoch(config, model, data)
write_test_predictions(ans_hats, config.pred_json_path)
logger.info('END ' + title)
return
# Training loop
epoch_results = []
max_em = -np.inf
max_f1 = -np.inf
np_rng = np.random.RandomState(config.seed // 2)
for epoch in range(1, config.max_num_epochs+1):
trn_loss, trn_acc, trn_samples_per_sec = _trn_epoch(config, model, data, epoch, np_rng)
dev_loss, dev_acc, dev_em, dev_f1 = _dev_epoch(config, model, data)
if dev_em > max_em:
model.save('models/' + base_filename + '_best_em.pkl')
max_em = dev_em
if dev_f1 > max_f1:
model.save('models/' + base_filename + '_best_f1.pkl')
max_f1 = dev_f1
if config.save_freq and epoch % config.save_freq == 0:
model.save('models/' + base_filename + '_e{:03d}.pkl'.format(epoch))
epoch_results.append(
EpochResult(trn_loss, trn_acc, dev_loss, dev_acc, dev_em, dev_f1))
if config.plot:
plot_epoch_results(epoch_results, 'logs/' + base_filename + '.png')
logger.info('\n\nFinished epoch {} for: (config index {}) (samples/sec: {:<.1f})\n{}\n\nResults:\n{}\n\n'.format(
epoch, config_idx, trn_samples_per_sec, config.format_compared(), format_epoch_results(epoch_results)))
logger.info('END ' + title)
def prepare_data(datapaths, embedding_path=None, embedding='word2vec', embedd_dim=100, prompt_id=1, vocab_size=0, tokenize_text=True, \
to_lower=True, sort_by_len=False, vocab_path=None, score_index=3):
assert len(datapaths) == 3, "data paths should include train, dev and test path"
(train_x, train_y), (dev_x, dev_y), (test_x, test_y), vocab, overal_maxlen, overal_maxnum = \
reader.get_data(datapaths, prompt_id, vocab_size, tokenize_text=True, to_lower=True, sort_by_len=False, vocab_path=None, score_index=6)
X_train, y_train, mask_train = padding_sentence_sequences(train_x, train_y, overal_maxnum, overal_maxlen, post_padding=True)
X_dev, y_dev, mask_dev = padding_sentence_sequences(dev_x, dev_y, overal_maxnum, overal_maxlen, post_padding=True)
X_test, y_test, mask_test = padding_sentence_sequences(test_x, test_y, overal_maxnum, overal_maxlen, post_padding=True)
train_mean = y_train.mean(axis=0)
train_std = y_train.std(axis=0)
dev_mean = y_dev.mean(axis=0)
dev_std = y_dev.std(axis=0)
test_mean = y_test.mean(axis=0)
test_std = y_test.std(axis=0)
Y_train = reader.get_model_friendly_scores(y_train, prompt_id)
Y_dev = reader.get_model_friendly_scores(y_dev, prompt_id)
Y_test = reader.get_model_friendly_scores(y_test, prompt_id)
scaled_train_mean = reader.get_model_friendly_scores(train_mean, prompt_id)
logger.info('Statistics:')
logger.info(' train X shape: ' + str(X_train.shape))
logger.info(' dev X shape: ' + str(X_dev.shape))
logger.info(' test X shape: ' + str(X_test.shape))
logger.info(' train Y shape: ' + str(Y_train.shape))
logger.info(' dev Y shape: ' + str(Y_dev.shape))
logger.info(' test Y shape: ' + str(Y_test.shape))
logger.info(' train_y mean: %s, stdev: %s, train_y mean after scaling: %s' %
(str(train_mean), str(train_std), str(scaled_train_mean)))
if embedding_path:
embedd_dict, embedd_dim, _ = load_word_embedding_dict(embedding, embedding_path, vocab, logger, embedd_dim)
embedd_matrix = build_embedd_table(vocab, embedd_dict, embedd_dim, logger, caseless=True)
else:
embedd_matrix = None
return (X_train, Y_train, mask_train), (X_dev, Y_dev, mask_dev), (X_test, Y_test, mask_test), vocab, len(vocab), embedd_matrix, overal_maxlen, overal_maxnum, scaled_train_mean
def main():
# Get the arguments from the command-line except the filename
argv = sys.argv[1:]
params = get_params(argv)
params['data_params'], type_dict = reader.get_data_settings(params['config_file'])
header, raw_data = reader.get_data(params)
type_selector, converters = reader.validate_data_format(params['data_params'], header)
key_fails, failed_rows, surveys, out_keys = data_processor.process_rows(raw_data, header, params, converters)
if 'logfile' in params:
with open(params['logfile'], 'w') as logfile:
now = datetime.datetime.now()
logfile.writelines(['Completed: ' + now.strftime("%B %d, %Y") + '\n'])
logfile.writelines(['Total of %s rows were not completed' % len(failed_rows) + '\n'])
for key in key_fails.keys():
logfile.writelines([str(key_fails[key]) + ' ' + str(key) + ' key(s) could not be read' + '\n'])
logfile.writelines(['%s surveys read, %s surveys failed' % (len(raw_data), len(failed_rows)) + '\n'])
logfile.writelines(['%s percent success rate' % str(
round((len(raw_data) - len(failed_rows)) / len(raw_data) * 100, 2)) + '\n'])
out_header = header + out_keys
writer.write_data(surveys, out_header, params)
def prepare_sentence_data(datapath, vocab_path, embedding_path=None, embedding='glove', embedd_dim=100, prompt_id=1, vocab_size=0, tokenize_text=True, to_lower=True, sort_by_len=False):
assert len(datapath) == 1, "data paths should include train, dev and test path"
(train_x, train_y, train_prompts), vocab, overal_maxlen, overal_maxnum = reader.get_data(datapath, vocab_path, prompt_id, vocab_size, tokenize_text=True, to_lower=True, sort_by_len=False)
X_train, y_train, mask_train = utils.padding_sentence_sequences(train_x, train_y, overal_maxnum, overal_maxlen, post_padding=True)
train_mean = y_train.mean(axis=0)
train_std = y_train.std(axis=0)
# Convert scores to boundary of [0 1] for training and evaluation (loss calculation)
#######################################
#这里我改了 prompt_id ->train_prompts
#######################################
if prompt_id ==-1:
Y_train = reader.get_model_friendly_scores(y_train, train_prompts)
scaled_train_mean = Y_train.mean()
elif prompt_id!=-1:
Y_train = reader.get_model_friendly_scores(y_train, prompt_id)
scaled_train_mean = reader.get_model_friendly_scores(train_mean, prompt_id)
# print Y_train.shape
logger.info('Statistics:')
logger.info(' train X shape: ' + str(X_train.shape))
logger.info(' train Y shape: ' + str(Y_train.shape))
logger.info(' train_y mean: %s, stdev: %s, train_y mean after scaling: %s' %
(str(train_mean), str(train_std), str(scaled_train_mean)))
if embedding_path:
embedd_dict, embedd_dim, _ = utils.load_word_embedding_dict(embedding, embedding_path, vocab, logger, embedd_dim)
embedd_matrix = utils.build_embedd_table(vocab, embedd_dict, embedd_dim, logger, caseless=True)
else:
embedd_matrix = None
return (X_train, Y_train, mask_train), vocab, len(vocab), embedd_matrix, overal_maxlen, overal_maxnum, scaled_train_mean
def main():
# Get the arguments from the command-line except the filename
argv = sys.argv[1:]
params = get_params(argv)
print('Program started, input file: %s' % params['filepath'])
if 'debug' in params.keys():
if params['debug'] == '1':
print('Setting debug level: Debug')
logging.getLogger().setLevel(logging.DEBUG)
else:
print('Setting debug level: Warnings')
logging.getLogger().setLevel(logging.WARNING)
else:
logging.getLogger().setLevel(logging.INFO)
params['data_params'], type_dict = reader.get_data_settings(params['config_file'])
header, raw_data = reader.get_data(params)
type_selector, converters = reader.validate_data_format(params['data_params'], header)
key_fails, failed_rows, surveys,quality_assessment, out_keys,meta, raw_surveys = data_processor.process_rows(raw_data, header, params, converters)
if 'logfile' in params:
if params['logfile'].endswith('txt'):
writer.write_log(logfiler.write_txt_log(params, key_fails, raw_data, failed_rows, meta), params['logfile'], 'Log:')
elif params['logfile'].endswith('csv'):
writer.write_log(logfiler.create_pbi_log(surveys,quality_assessment, meta['attribute_quality'], meta, failed_rows,params), params['logfile'], ['Key', 'Value'])
else:
logging.CRITICAL('Logfile not written, wrong file extension provided')
out_header = surveys[0].keys()
if params['outfile'].endswith('.csv'):
writer.write_data(surveys, out_header, params,rounding=9,raw_surveys=raw_surveys)
if len(quality_assessment) > 0:
writer.write_data(quality_assessment, quality_assessment[0].keys(), params, sub_file='group_qual')
if len(meta['attribute_quality']) > 0 and 'logfile' in params:
writer.write_data(meta['attribute_quality'], meta['attribute_quality'][0].keys(), params,
sub_file='attr_qual')
else:
logging.critical("Your output filename: %s must end with .csv" % params['outfile'])
"--emb_technique",
dest="emb_technique",
type=str,
metavar='<str>',
default='w2v',
help="embedding technique (w2v or fasttext)")
parser.add_argument("-o",
"--output",
dest="out_file",
type=str,
metavar='<str>',
help="output file name")
args = parser.parse_args()
vocab, train_x, overall_maxlen = dataset.get_data(vocab_size=args.vocab_size,
maxlen=args.maxlen,
lang=args.lang)
train_x = sequence.pad_sequences(train_x, maxlen=overall_maxlen)
print('Number of training examples: ', len(train_x))
print('Length of vocab: ', len(vocab))
optimizer = get_optimizer(args.algorithm)
model = create_model(args, overall_maxlen, vocab)
model_param = config.model_param_file[args.lang].format(
args.emb_technique, config.word_emb_training_type, args.epochs,
config.filter_word_on)
model.load_weights(model_param)
model.compile(optimizer=optimizer,
loss=ut.max_margin_loss,
metrics=[ut.max_margin_loss])
from reader import get_data
from basic_prepro import process, do_tokeinze
text, title = get_data('001.pdf')
processed = process(text)
print(len(do_tokeinze(processed)))
embedding = np.zeros(300,dtype=np.float32)
sen_embedding = sen_embedding + embedding
return sen_embedding
elif embedding_type == 'WORD2VEC':
sen_embedding = np.zeros(300,dtype=np.float32)
for word in sentence:
try:
embedding = model_word2vec[word]
except:
embedding = np.zeros(300,dtype=np.float32)
sen_embedding = sen_embedding + embedding
return sen_embedding
return None
x, y , ac_num, ac_words = reader.get_data()
x_embedding = []
for node in x:
node_embedding = []
for sent in node:
temp_sent = sent.split(" ")
sent_embedding = sentence_embedding([temp_value.strip() for temp_value in temp_sent],"WORD2VEC")
node_embedding.append(sent_embedding)
x_embedding.append(node_embedding)
x_embedding = np.asarray(x_embedding)
#done with the data preperation step
x_embedding,y = shuffle_copy(x_embedding,y)
evals = []
# 0.63
}
#assert args.domain in {'restaurant', 'beer'}
if args.seed > 0:
np.random.seed(args.seed)
# ###############################################################################################################################
# ## Prepare data
# #
from keras.preprocessing import sequence
import reader as dataset
#vocab, train_x, test_x, overall_maxlen = dataset.get_data(args.domain, vocab_size=args.vocab_size, maxlen=args.maxlen)
vocab, train_x, overall_maxlen = dataset.get_data(args.input_path,
args.out_dir_path,
vocab_size=args.vocab_size,
maxlen=args.maxlen)
train_x = sequence.pad_sequences(train_x, maxlen=overall_maxlen)
#test_x = sequence.pad_sequences(test_x, maxlen=overall_maxlen)
print('Number of training examples: ', len(train_x))
print('Length of vocab: ', len(vocab))
def sentence_batch_generator(data, batch_size):
n_batch = len(data) // batch_size
batch_count = 0
np.random.shuffle(data)
while True:
def prepare_sentence_data(datapaths, embedding_path=None, embedding='word2vec', embedd_dim=100, prompt_id=1, vocab_size=0, tokenize_text=True, \
to_lower=True, sort_by_len=False, vocab_path=None, score_index=6):
assert len(
datapaths) == 4, "data paths should include train, dev and test path"
(train_x, train_y, train_prompts, train_ids), (dev_x, dev_y, dev_prompts, dev_ids), (test_x, test_y, test_prompts, test_ids), vocab, overal_maxlen, overal_maxnum = \
reader.get_data(datapaths, prompt_id, vocab_size, tokenize_text=True, to_lower=True, sort_by_len=False, vocab_path=None, score_index=6)
train_d, max_sentnum = reader.read_description(datapaths[3],
vocab,
len(train_x),
tokenize_text=True,
to_lower=True)
dev_d, max_sentnum = reader.read_description(datapaths[3],
vocab,
len(dev_x),
tokenize_text=True,
to_lower=True)
test_d, max_sentnum = reader.read_description(datapaths[3],
vocab,
len(test_x),
tokenize_text=True,
to_lower=True)
X_train, y_train, mask_train = utils.padding_sentence_sequences(
train_x, train_y, overal_maxnum, overal_maxlen, post_padding=True)
X_dev, y_dev, mask_dev = utils.padding_sentence_sequences(
dev_x, dev_y, overal_maxnum, overal_maxlen, post_padding=True)
X_test, y_test, mask_test = utils.padding_sentence_sequences(
test_x, test_y, overal_maxnum, overal_maxlen, post_padding=True)
D_train, mask_d_train = utils.padding_des_sequences(train_d,
max_sentnum,
overal_maxlen,
post_padding=True)
D_dev, mask_d_dev = utils.padding_des_sequences(dev_d,
max_sentnum,
overal_maxlen,
post_padding=True)
D_test, mask_d_test = utils.padding_des_sequences(test_d,
max_sentnum,
overal_maxlen,
post_padding=True)
if prompt_id:
train_pmt = np.array(train_prompts, dtype='int32')
dev_pmt = np.array(dev_prompts, dtype='int32')
test_pmt = np.array(test_prompts, dtype='int32')
train_mean = y_train.mean(axis=0)
train_std = y_train.std(axis=0)
dev_mean = y_dev.mean(axis=0)
dev_std = y_dev.std(axis=0)
test_mean = y_test.mean(axis=0)
test_std = y_test.std(axis=0)
# We need the dev and test sets in the original scale for evaluation
# dev_y_org = y_dev.astype(reader.get_ref_dtype())
# test_y_org = y_test.astype(reader.get_ref_dtype())
# Convert scores to boundary of [0 1] for training and evaluation (loss calculation)
Y_train = reader.get_model_friendly_scores(y_train, prompt_id)
Y_dev = reader.get_model_friendly_scores(y_dev, prompt_id)
Y_test = reader.get_model_friendly_scores(y_test, prompt_id)
scaled_train_mean = reader.get_model_friendly_scores(train_mean, prompt_id)
# print Y_train.shape
logger.info('Statistics:')
logger.info(' train X shape: ' + str(X_train.shape))
logger.info(' dev X shape: ' + str(X_dev.shape))
logger.info(' test X shape: ' + str(X_test.shape))
logger.info(' train Y shape: ' + str(Y_train.shape))
logger.info(' dev Y shape: ' + str(Y_dev.shape))
logger.info(' test Y shape: ' + str(Y_test.shape))
logger.info(
' train_y mean: %s, stdev: %s, train_y mean after scaling: %s' %
(str(train_mean), str(train_std), str(scaled_train_mean)))
if embedding_path:
embedd_dict, embedd_dim, _ = utils.load_word_embedding_dict(
embedding, embedding_path, vocab, logger, embedd_dim)
embedd_matrix = utils.build_embedd_table(vocab,
embedd_dict,
embedd_dim,
logger,
caseless=True)
else:
embedd_matrix = None
return (X_train, Y_train, D_train, mask_train, train_ids), (X_dev, Y_dev, D_dev, mask_dev, dev_ids), (X_test, Y_test, D_test ,mask_test, test_ids), \
vocab, len(vocab), embedd_matrix, overal_maxlen, overal_maxnum, max_sentnum, scaled_train_mean
import calculator.flow
import viewer.team_metrics
import viewer.dash
import logging
import logging.config
import config
with open("log_config.yaml", "r") as f:
log_config = yaml.safe_load(f.read())
logging.config.dictConfig(log_config)
logger = logging.getLogger(__name__)
logger.info("Starting seshat. Let's do team magic!")
config = config.get()
projects = []
for source_config in config["input"]:
logging.info(f"Reading data for {source_config['name']}")
data = reader.get_data(source_config)
cycle_data = calculator.flow.cycle_data(data, source_config)
team_metrics = viewer.team_metrics.Team_Metrics(cycle_data, source_config)
projects.append(team_metrics)
dash = viewer.dash.Dash(projects, config)
server = dash.server
if __name__ == "__main__":
dash.run()
def _main(config):
base_filename = config.name
logger_filename = 'logs/' + base_filename + '.log'
logger = set_up_logger(logger_filename)
title = '{}: {} ({})'.format(__file__, config.name, config.desc)
logger.info('START ' + title + '\n\n{}\n'.format(config))
data = get_data(
word_emb_data_path_prefix=GLOVE_PREPROC_WITH_UNKS_PATH_PREFIX,
tokenized_trn_json_path=TOKENIZED_TRN_JSON_PATH,
tokenized_dev_json_path=TOKENIZED_DEV_JSON_PATH,
max_ans_len=MAX_ANS_LEN,
max_ctx_len=MAX_CTX_LEN)
if config.device != 'cpu':
assert 'theano' not in sys.modules
import theano.sandbox.cuda
theano.sandbox.cuda.use(config.device)
from model import get_model
model = get_model(config, data)
lm_data = get_lm_data(config.lm_layer) if config.mode == 'LM' else None
# Training loop
epoch_results = []
max_em = -np.inf
max_f1 = -np.inf
np_rng = np.random.RandomState(config.seed // 2)
for epoch in range(1, config.max_num_epochs + 1):
trn_loss, trn_acc, trn_samples_per_sec, trn_num_all_samples, trn_num_valid_samples, \
trn_mean_grad_norm, trn_max_grad_norm, trn_min_grad_norm, trn_num_unsafe_samples = \
_trn_epoch(config, model, data, lm_data, epoch, np_rng)
dev_loss, dev_acc, dev_em, dev_f1, dev_num_all_samples, dev_num_valid_samples = \
_dev_epoch(config, model, data, lm_data)
best_filename = base_filename
if dev_em > max_em:
model.save('models/' + best_filename + '_best_em.pkl')
max_em = dev_em
if dev_f1 > max_f1:
model.save('models/' + best_filename + '_best_f1.pkl')
max_f1 = dev_f1
if config.save_freq and epoch % config.save_freq == 0:
model.save('models/' + base_filename +
'_e{:03d}.pkl'.format(epoch))
epoch_results.append(
EpochResult(trn_loss, trn_acc, dev_loss, dev_acc, dev_em, dev_f1))
if config.plot:
plot_epoch_results(epoch_results, 'logs/' + base_filename + '.png')
logger.info(
('\n\nEpc {} {}: (smp/sec: {:<.1f})'
' (trn: {}/{}) (dev: {}/{})'
' (grad: avg:{} max:{} min:{}) (low probability predictions:{})'
'\n{}\n\nResults:\n{}\n\n').format(
epoch, config.name, trn_samples_per_sec,
trn_num_valid_samples, trn_num_all_samples,
dev_num_valid_samples, dev_num_all_samples,
trn_mean_grad_norm, trn_max_grad_norm, trn_min_grad_norm,
trn_num_unsafe_samples, config.format_compared(),
format_epoch_results(epoch_results)))
logger.info('END ' + title)
# ###############################################################################################################################
# ## Prepare data
# #
from keras.preprocessing import sequence
import reader as dataset # 有3个方法
# create_vocab
# read_dataset
# get_data
# args = parser.parse_args() 这里保存了所有的参数
# 这一步就会创建vocab
vocab, train_x, test_x, overall_maxlen = dataset.get_data(
args.domain, vocab_size=args.vocab_size,
maxlen=args.maxlen) # 读取已经经过preprocessing的数据.
train_x = sequence.pad_sequences(train_x, maxlen=overall_maxlen)
test_x = sequence.pad_sequences(test_x, maxlen=overall_maxlen)
train_x = train_x[0:3000] # 限制长度进行test
print('Number of training examples: ', len(train_x))
print('Length of vocab: ', len(vocab))
def sentence_batch_generator(data, batch_size):
n_batch = len(data) // batch_size
batch_count = 0
np.random.shuffle(data)
while True:
def set_rss():
for key in rss.keys():
if key not in rss_contents.keys():
rss_data = reader.get_data(rss[key])
rss_contents[key] = rss_data
def _main(config, config_idx, train):
base_filename = config.name + '_cfg' + str(config_idx)
logger = set_up_logger('logs/' + base_filename + '.log')
title = '{}: {} ({}) config index {}'.format(__file__, config.name,
config.desc, config_idx)
logger.info('START ' + title + '\n\n{}\n'.format(config))
data = get_data(config, train)
if config.device != 'cpu':
assert 'theano' not in sys.modules
import theano.sandbox.cuda
theano.sandbox.cuda.use(config.device)
from model import get_model
model = get_model(config, data)
if not train:
assert config.tst_load_model_path
if not model.load(config.tst_load_model_path):
raise AssertionError('Failed loading model weights from {}'.format(
config.tst_load_model_path))
ans_hats = _tst_epoch(config, model, data)
write_test_predictions(ans_hats, config.pred_json_path)
logger.info('END ' + title)
return
# Training loop
epoch_results = []
max_em = -np.inf
max_f1 = -np.inf
epochs_with_no_improvement = 0
np_rng = np.random.RandomState(config.seed // 2)
for epoch in range(1, config.max_num_epochs + 1):
trn_loss, trn_acc, trn_samples_per_sec = _trn_epoch(
config, model, data, epoch, np_rng)
dev_loss, dev_acc, dev_em, dev_f1 = _dev_epoch(config, model, data)
if dev_em > max_em:
model.save('models/' + base_filename + '_best_em.pkl')
max_em = dev_em
# Best EM so far, reset epochs_with_no_improvement
epochs_with_no_improvement = 0
if dev_f1 > max_f1:
model.save('models/' + base_filename + '_best_f1.pkl')
max_f1 = dev_f1
# Best F1 so far, reset epochs_with_no_improvement
epochs_with_no_improvement = 0
if dev_em <= max_em and dev_f1 <= max_f1:
# Neither dev_em nor dev_f1 are better than max, increment epochs
# with no improvement.
epochs_with_no_improvement += 1
if config.save_freq and epoch % config.save_freq == 0:
model.save('models/' + base_filename +
'_e{:03d}.pkl'.format(epoch))
epoch_results.append(
EpochResult(trn_loss, trn_acc, dev_loss, dev_acc, dev_em, dev_f1))
if config.plot:
plot_epoch_results(epoch_results, 'logs/' + base_filename + '.png')
logger.info(
'\n\nFinished epoch {} for: (config index {}) (samples/sec: {:<.1f})\n{}\n\nResults:\n{}\n\n'
.format(epoch, config_idx, trn_samples_per_sec,
config.format_compared(),
format_epoch_results(epoch_results)))
# Check if we have to do early stopping.
if epochs_with_no_improvement > config.patience:
logger.info("Patience exceeded.")
break
logger.info('END ' + title)
'rmsprop', 'sgd', 'adagrad', 'adadelta', 'adam', 'adamax'
}
# assert args.domain in {'restaurant', 'beer'}
if args.seed > 0:
np.random.seed(args.seed)
# ###############################################################################################################################
# ## Prepare data
# #
from keras.preprocessing import sequence
import reader as dataset
vocab, train_x, overall_maxlen = dataset.get_data(args.domain,
vocab_size=args.vocab_size,
maxlen=args.maxlen,
require_test=False)
train_x = sequence.pad_sequences(train_x, maxlen=overall_maxlen)
# train_x = train_x[0:30000]
print('Number of training examples: ', len(train_x))
print('Length of vocab: ', len(vocab))
def sentence_batch_generator(data, batch_size):
n_batch = len(data) // batch_size
batch_count = 0
np.random.shuffle(data)
while True:
if batch_count == n_batch:
anss_val = np.array([_np_ans_word_idxs_to_ans_idx(ans_stt, ans_end, max_ans_len) for \
ans_stt, ans_end in anss], dtype=np.int32)
ans_stts_val = anss[:, 0]
ans_ends_val = anss[:, 1]
gpu_anss = torch.from_numpy(anss_val)
gpu_ans_stts = torch.from_numpy(ans_stts_val)
gpu_ans_ends = torch.from_numpy(ans_ends_val)
return gpu_anss, gpu_ans_stts, gpu_ans_ends
config = Config()
base_filename = config.name + '_cfg' + str(0)
logger = set_up_logger('logs/' + base_filename + '.log')
title = '{}: {}'.format(__file__, config.name)
logger.info('START ' + title + '\n\n{}\n'.format(config))
data = get_data(config, train=True)
emb_val = data.word_emb_data.word_emb # (voc size, emb_dim)
first_known_word = data.word_emb_data.first_known_word
assert config.emb_dim == emb_val.shape[1]
assert first_known_word > 0
emb_val[:first_known_word] = 0
emb = torch.from_numpy(emb_val)
#load all the data, train and dev data
trn_ctxs, trn_ctx_masks, trn_ctx_lens, trn_qtns, trn_qtn_masks, trn_qtn_lens, trn_qtn_ctx_idxs, trn_anss, trn_ans_stts, trn_ans_ends = _gpu_dataset(
'trn', data.trn, config)
dev_ctxs, dev_ctx_masks, dev_ctx_lens, dev_qtns, dev_qtn_masks, dev_qtn_lens, dev_qtn_ctx_idxs, dev_anss, dev_ans_stts, dev_ans_ends = _gpu_dataset(
'dev', data.dev, config)
def print_param(mdoel):
for name, param in model.state_dict().items():
def prepare_sentence_data(
datapaths,
embedding_path=None,
embedding='word2vec',
emb_dim=100,
prompt_id=1,
vocab_size=0,
tokenize_text=True,
to_lower=True,
vocab_path=None,
score_index=6,
need_context=True
):
assert len(datapaths) == 3, "data paths should include train, dev and test path"
(train_x, train_y, train_prompts, train_text), \
(dev_x, dev_y, dev_prompts, dev_text), \
(test_x, test_y, test_prompts, test_text), \
vocab, overall_maxlen, overall_maxnum = \
reader.get_data(
datapaths,
prompt_id,
vocab_size,
tokenize_text,
to_lower,
vocab_path,
score_index)
X_train, y_train, mask_train = utils.padding_sentence_sequences(train_x, train_y, overall_maxnum, overall_maxlen,
post_padding=True)
X_dev, y_dev, mask_dev = utils.padding_sentence_sequences(dev_x, dev_y, overall_maxnum, overall_maxlen,
post_padding=True)
X_test, y_test, mask_test = utils.padding_sentence_sequences(test_x, test_y, overall_maxnum, overall_maxlen,
post_padding=True)
if need_context:
context, context_len, context_num = reader.get_context(prompt_id, vocab, to_lower)
else:
# Dummy context
context = [[0]]
context_len = 1
context_num = 1
train_context = [context] * len(train_x)
dev_context = [context] * len(dev_x)
test_context = [context] * len(test_x)
train_context, _, _ = utils.padding_sentence_sequences(train_context, train_y, context_num, context_len, post_padding=True)
dev_context, _, _ = utils.padding_sentence_sequences(dev_context, dev_y, context_num, context_len, post_padding=True)
test_context, _, _ = utils.padding_sentence_sequences(test_context, test_y, context_num, context_len, post_padding=True)
train_mean = y_train.mean(axis=0)
train_std = y_train.std(axis=0)
dev_mean = y_dev.mean(axis=0)
dev_std = y_dev.std(axis=0)
test_mean = y_test.mean(axis=0)
test_std = y_test.std(axis=0)
# We need the dev and test sets in the original scale for evaluation
# dev_y_org = y_dev.astype(reader.get_ref_dtype())
# test_y_org = y_test.astype(reader.get_ref_dtype())
# Convert scores to boundary of [0 1] for training and evaluation (loss calculation)
Y_train = utils.get_model_friendly_scores(y_train, prompt_id)
Y_dev = utils.get_model_friendly_scores(y_dev, prompt_id)
Y_test = utils.get_model_friendly_scores(y_test, prompt_id)
scaled_train_mean = utils.get_model_friendly_scores(train_mean, prompt_id)
scaled_dev_mean = utils.get_model_friendly_scores(dev_mean, prompt_id)
scaled_test_mean = utils.get_model_friendly_scores(test_mean, prompt_id)
# print Y_train.shape
logger.info('Statistics:')
logger.info(' train X shape: ' + str(X_train.shape))
logger.info(' dev X shape: ' + str(X_dev.shape))
logger.info(' test X shape: ' + str(X_test.shape))
if need_context:
logger.info(' train context shape: ' + str(train_context.shape))
logger.info(' dev context shape: ' + str(dev_context.shape))
logger.info(' test context shape: ' + str(test_context.shape))
logger.info(' train Y shape: ' + str(Y_train.shape))
logger.info(' dev Y shape: ' + str(Y_dev.shape))
logger.info(' test Y shape: ' + str(Y_test.shape))
logger.info(' train_y mean: %s, stdev: %s, train_y mean after scaling: %s' %
(str(train_mean), str(train_std), str(scaled_train_mean)))
logger.info(' dev_y mean: %s, stdev: %s, dev_y mean after scaling: %s' %
(str(dev_mean), str(dev_std), str(scaled_dev_mean)))
logger.info(' test_y mean: %s, stdev: %s, test_y mean after scaling: %s' %
(str(test_mean), str(test_std), str(scaled_test_mean)))
if embedding_path:
emb_dict, emb_dim, _ = utils.load_word_embedding_dict(embedding, embedding_path, vocab, logger, emb_dim)
emb_matrix = utils.build_embedding_table(vocab, emb_dict, emb_dim, logger, caseless=True)
else:
emb_matrix = None
return (X_train, Y_train, mask_train, train_context, train_text), \
(X_dev, Y_dev, mask_dev, dev_context, dev_text), \
(X_test, Y_test, mask_test, test_context, test_text), \
vocab, len(vocab), emb_matrix, overall_maxlen, overall_maxnum, scaled_train_mean, context_len, context_num
embedding = np.zeros(300,dtype=np.float32)
sen_embedding = sen_embedding + embedding
return sen_embedding
elif embedding_type == 'WORD2VEC':
sen_embedding = np.zeros(300,dtype=np.float32)
for word in sentence:
try:
embedding = model_word2vec[word]
except:
embedding = np.zeros(300,dtype=np.float32)
sen_embedding = sen_embedding + embedding
return sen_embedding
return None
x, y , ac_num, ac_words = reader.get_data()
x_embedding = []
for node in x:
node_embedding = []
for sent in node:
temp_sent = sent.split(" ")
sent_embedding = sentence_embedding([temp_value.strip() for temp_value in temp_sent],"WORD2VEC")
node_embedding.append(sent_embedding)
x_embedding.append(node_embedding)
x_embedding = np.asarray(x_embedding)
#done with the data preperation step
x_embedding,y = shuffle_copy(x_embedding,y)
evals = []
# 0.63
def train_model_each_cluster(args,cluster_size,embtype):
logger.info("Cluster Size: {}".format(cluster_size))
args.aspect_size = cluster_size
if args.seed > 0:
np.random.seed(args.seed)
aspect_file_name = config.aspect_file_name[args.lang].format(args.emb_technique, config.word_emb_training_type, args.epochs, config.filter_word_on, embtype, cluster_size)
model_path = config.model_param_file[args.lang].format(args.emb_technique, config.word_emb_training_type, args.epochs, config.filter_word_on)
util.createPath(aspect_file_name)
vocab, train_x, overall_maxlen = dataset.get_data(vocab_size=args.vocab_size,
maxlen=args.maxlen, lang=args.lang)
train_x = sequence.pad_sequences(train_x, maxlen=overall_maxlen)
print('Number of training examples: ', len(train_x))
print('Length of vocab: ', len(vocab))
optimizer = get_optimizer(args.algorithm)
logger.info('Building {} based model for {}'.format(args.emb_technique, args.lang))
model = create_model(args, overall_maxlen, vocab)
# freeze the word embedding layer
model.get_layer('word_emb').trainable = False
model.compile(optimizer=optimizer, loss=util.max_margin_loss, metrics=[util.max_margin_loss])
logger.info("-" * 80)
vocab_inv = {}
for w, ind in vocab.items():
vocab_inv[ind] = w
sen_gen = sentence_batch_generator(train_x, args.batch_size)
neg_gen = negative_batch_generator(train_x, args.batch_size, args.neg_size)
batches_per_epoch = len(train_x) // args.batch_size
min_loss = float('inf')
for ii in range(args.epochs):
t0 = time()
loss, max_margin_loss = 0., 0.
for b in tqdm(range(batches_per_epoch)):
sen_input = next(sen_gen)
neg_input = next(neg_gen)
batch_loss, batch_max_margin_loss = model.train_on_batch([sen_input, neg_input],
np.ones((args.batch_size, 1)))
loss += batch_loss / batches_per_epoch
max_margin_loss += batch_max_margin_loss / batches_per_epoch
tr_time = time() - t0
if loss < min_loss:
min_loss = loss
word_emb = K.get_value(model.get_layer('word_emb').embeddings)
aspect_emb = K.get_value(model.get_layer('aspect_emb').W)
word_emb = word_emb / np.linalg.norm(word_emb, axis=-1, keepdims=True)
aspect_emb = aspect_emb / np.linalg.norm(aspect_emb, axis=-1, keepdims=True)
aspect_file = open(aspect_file_name, 'wt', encoding='utf-8')
model.save(model_path)
for ind in range(len(aspect_emb)):
desc = aspect_emb[ind]
sims = word_emb.dot(desc.T)
ordered_words = np.argsort(sims)[::-1]
desc_list = [vocab_inv[w] + "|" + str(sims[w]) for w in ordered_words[:50]]
# print('Aspect %d:' % ind)
# print(desc_list)
aspect_file.write('Aspect %d:\n' % ind)
aspect_file.write(' '.join(desc_list) + '\n\n')
per_cluster_train_loss = loss
logger.info('Epoch %d, train: %is' % (ii, tr_time))
logger.info(
'Total loss: %.4f, max_margin_loss: %.4f, ortho_reg: %.4f' % (
loss, max_margin_loss, loss - max_margin_loss))
return per_cluster_train_loss
assert args.algorithm in {
'rmsprop', 'sgd', 'adagrad', 'adadelta', 'adam', 'adamax'
}
# assert args.domain in {'restaurant', 'beer'}
if args.seed > 0:
np.random.seed(args.seed)
# ###############################################################################################################################
# ## Prepare data
# #
from keras.preprocessing import sequence
import reader as dataset
vocab, train_x, test_x, overall_maxlen = dataset.get_data(
args.domain, vocab_size=args.vocab_size, maxlen=args.maxlen)
train_x = sequence.pad_sequences(train_x, maxlen=overall_maxlen)
test_x = sequence.pad_sequences(test_x, maxlen=overall_maxlen)
# train_x = train_x[0:30000]
print('Number of training examples: ', len(train_x))
print('Length of vocab: ', len(vocab))
def sentence_batch_generator(data, batch_size):
n_batch = len(data) // batch_size
batch_count = 0
np.random.shuffle(data)
while True:
if batch_count == n_batch:
if args.train:
print("Training...")
df_train = pd.read_json(args.data_dir + "csqa.train.json")
df_dev = pd.read_json(args.data_dir + "csqa.dev.json")
if args.random_emb:
train_embs = np.random.uniform(-1, 1, (406213, 768))
dev_embs = np.random.uniform(-1, 1, (45076, 768))
else:
train_embs = reader.load_embs(args.data_dir +
"csqa.train.embeddings.bin")
dev_embs = reader.load_embs(args.data_dir +
"csqa.dev.embeddings.bin")
train_data = reader.get_data(
df_train,
train_embs,
rel2idx,
subset=int(config["parameters"]["subset_train"]),
shuffle=False,
random=args.random_emb,
)
dev_data = reader.get_data(
df_dev,
dev_embs,
rel2idx,
subset=int(config["parameters"]["subset_dev"]),
shuffle=False,
random=args.random_emb,
)
classifier.train(train_data, dev_data)
if args.test:
