def __init__(self, DATA_DIR, dset, is_train, verbose, reward='RECALL'):
# this method returns simulator, state/action vocabularies, and the maximum number of actions
n_words = 100 # 374000 # words for the vocabulary
vocab_path = os.path.join(
DATA_DIR, 'data/D_cbow_pdw.pkl'
) # Path to the python dictionary containing the vocabulary.
wordemb_path = os.path.join(
DATA_DIR, 'data/D_cbow_pdw.pkl'
) # Path to the python dictionary containing the word embeddings.
dataset_path = os.path.join(
DATA_DIR, 'data/msa_dataset.hdf5'
) # path to load the hdf5 dataset containing queries and ground-truth documents.
docs_path = os.path.join(
DATA_DIR,
'data/msa_corpus.hdf5') # Path to load the articles and links.
docs_path_term = os.path.join(
DATA_DIR,
'data/msa_corpus.hdf5') # Path to load the articles and links.
############################
# Search Engine Parameters #
############################
n_threads = 1 # 20 # number of parallel process that will execute the queries on the search engine.
index_name = 'index' # index name for the search engine. Used when engine is 'lucene'.
index_name_term = 'index_terms' # index name for the search engine. Used when engine is 'lucene'.
use_cache = False # If True, cache (query-retrieved docs) pairs. Watch for memory usage.
max_terms_per_doc = 15 # Maximum number of candidate terms from each feedback doc. Must be always less than max_words_input .
self.actions_space = dict()
self.states_space = dict()
self.current_queries = []
self.D_gt_id = []
# self.word2vec = Word2Vec.load_word2vec_format(GLOVE_FILE, binary=binary)
self.word_index = None
# self.embedding_dim = self.word2vec.syn0.shape[1]
self.vocab = utils.load_vocab(vocab_path, n_words)
vocabinv = {}
for k, v in self.vocab.items():
vocabinv[v] = k
self.reward = reward
self.is_train = is_train
self.search = Search(engine=lucene_search.LuceneSearch(
DATA_DIR, self.vocab, n_threads, max_terms_per_doc, index_name,
index_name_term, docs_path, docs_path_term, use_cache))
t0 = time()
dh5 = dataset_hdf5.DatasetHDF5(dataset_path)
self.qi = dh5.get_queries(dset)
self.dt = dh5.get_doc_ids(dset)
# print("Loading queries and docs {}".format(time() - t0))
self.reset()
def __init__(self, cfg, dset, is_train, verbose, rewardtype='RECALL'):
# this method returns simulator, state/action vocabularies, and the maximum number of actions
n_words = 374000 # 100 words for the vocabulary
DATA_DIR = cfg['data']['base_path']
vocab_path = os.path.join(
DATA_DIR, 'data/D_cbow_pdw_8B.pkl'
) # Path to the python dictionary containing the vocabulary.
wordemb_path = os.path.join(
DATA_DIR, 'data/D_cbow_pdw_8B.pkl'
) # Path to the python dictionary containing the word embeddings.
dataset_path = os.path.join(
DATA_DIR, 'data/msa_dataset.hdf5'
) # path to load the hdf5 dataset containing queries and ground-truth documents.
docs_path = os.path.join(
DATA_DIR,
'data/msa_corpus.hdf5') # Path to load the articles and links.
docs_path_term = os.path.join(
DATA_DIR,
'data/msa_corpus.hdf5') # Path to load the articles and links.
############################
# Search Engine Parameters #
############################
n_threads = 1 # 20 # number of parallel process that will execute the queries on the search engine.
index_name = 'index' # index name for the search engine. Used when engine is 'lucene'.
index_name_term = 'index_terms' # index name for the search engine. Used when engine is 'lucene'.
use_cache = False # If True, cache (query-retrieved docs) pairs. Watch for memory usage.
max_terms_per_doc = 15 # Maximum number of candidate terms from each feedback doc. Must be always less than max_words_input .
#self.batch_size_train=2 # The batch size during training.
self.cfg = cfg
t0 = time()
self.vocab = utils.load_vocab(vocab_path, n_words)
vocabinv = {}
for k, v in self.vocab.items():
vocabinv[v] = k
self.vocabinv = vocabinv
self.rewardtype = rewardtype
self.is_train = is_train
self.search = Search(engine=lucene_search.LuceneSearch(
DATA_DIR, self.vocab, n_threads, max_terms_per_doc, index_name,
index_name_term, docs_path, docs_path_term, use_cache))
self.batch_size = cfg['agent']['batch_size']
t0 = time()
dh5 = dataset_hdf5.DatasetHDF5(dataset_path)
self.qi = dh5.get_queries(dset)
cfg['data']['querydb_size'] = len(self.qi)
self.dt = dh5.get_doc_ids(dset)
print("Loading queries and docs {}".format(time() - t0))
self.counsteps = 0
def train():
if prm.optimizer.lower() == 'adam':
optimizer=adam
elif prm.optimizer.lower() == 'sgd':
optimizer=sgd
elif prm.optimizer.lower() == 'rmsprop':
optimizer=rmsprop
elif prm.optimizer.lower() == 'adadelta':
optimizer=adadelta
options = locals().copy()
print 'parameters:', str(options)
prm_k = vars(prm).keys()
prm_d = vars(prm)
prm_k.sort()
for x in prm_k:
if not x.startswith('__'):
print x,'=', prm_d[x]
print 'loading Vocabulary...'
vocab = utils.load_vocab(prm.vocab_path, prm.n_words)
options['vocab'] = vocab
options['vocabinv'] = {}
for k,v in vocab.items():
options['vocabinv'][v] = k
print 'Loading Environment...'
if prm.engine.lower() == 'lucene':
import lucene_search
options['engine'] = lucene_search.LuceneSearch()
elif prm.engine.lower() == 'elastic':
import elastic_search
options['engine'] = elastic_search.ElasticSearch()
print 'Loading Dataset...'
dh5 = dataset_hdf5.DatasetHDF5(prm.dataset_path)
qi_train = dh5.get_queries(dset='train')
dt_train = dh5.get_doc_ids(dset='train')
qi_valid = dh5.get_queries(dset='valid')
dt_valid = dh5.get_doc_ids(dset='valid')
qi_test = dh5.get_queries(dset='test')
dt_test = dh5.get_doc_ids(dset='test')
if prm.train_size == -1:
train_size = len(qi_train)
else:
train_size = min(prm.train_size, len(qi_train))
if prm.valid_size == -1:
valid_size = len(qi_valid)
else:
valid_size = min(prm.valid_size, len(qi_valid))
if prm.test_size == -1:
test_size = len(qi_test)
else:
test_size = min(prm.test_size, len(qi_test))
print '%d train examples' % len(qi_train)
print '%d valid examples' % len(qi_valid)
print '%d test examples' % len(qi_test)
# This create the initial parameters as np ndarrays.
# Dict name (string) -> np ndarray
params, exclude_params = init_params(options)
if prm.wordemb_path:
print 'loading pre-trained word embeddings'
params = load_wemb(params, vocab)
options['W'] = params['W']
if prm.reload_model:
load_params(prm.reload_model, params)
print 'Building model'
# This create Theano Shared Variable from the parameters.
# Dict name (string) -> Theano Tensor Shared Variable
# params and tparams have different copy of the weights.
tparams = init_tparams(params)
for kk, value in tparams.iteritems():
tparams[kk] = theano.shared(value, name=kk)
iin, out, updates, f_pred, consider_constant \
= build_model(tparams, options)
#get only parameters that are not in the exclude_params list
tparams_ = OrderedDict([(kk, vv) for kk, vv in tparams.iteritems() if kk not in exclude_params])
grads = tensor.grad(out[0], wrt=itemlist(tparams_), consider_constant=consider_constant)
lr = tensor.scalar(name='lr')
f_grad_shared, f_update = optimizer(lr, tparams_, grads, iin, out, updates)
history_errs = []
best_p = None
if prm.validFreq == -1:
validFreq = len(qi_train) / prm.batch_size_train
else:
validFreq = prm.validFreq
if prm.saveFreq == -1:
saveFreq = len(qi_train) / prm.batch_size_train
else:
saveFreq = prm.saveFreq
uidx = 0 # the number of update done
estop = False # early stop
start_time = time.time()
print 'Optimization'
try:
for eidx in xrange(prm.max_epochs):
n_samples = 0
# Get new shuffled index for the training set.
kf = get_minibatches_idx(len(qi_train), prm.batch_size_train, shuffle=True)
for _, train_index in kf:
st = time.time()
uidx += 1
qi, qi_i, qi_lst, D_gt_id, D_gt_url = get_samples(qi_train, dt_train, train_index, options)
# share the current queries with the search engine.
options['current_queries'] = qi_lst
n_samples += len(qi)
is_train = 1.
out = f_grad_shared(qi_i, D_gt_id, is_train)
cost = out.pop(0)
cost_ent = out.pop(0)
lr_t = f_update(prm.lrate)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
return 1., 1., 1.
if np.mod(uidx, prm.dispFreq) == 0:
print '\n================================================================================'
print 'Epoch', eidx, 'Update', uidx, 'Cost', cost, 'LR_t', lr_t
print 'Time Minibatch Update: ' + str(time.time() - st)
print 'Input Query: ', qi[0].replace('\n','\\n')
print
print 'Target Docs: ', str(D_gt_url[0])
print
print 'Input Query Vocab: ', utils.idx2text(qi_i[0], options['vocabinv'])
for ii in range(prm.n_iterations):
prob = out.pop(0)
ans = out.pop(0)
metrics = out.pop(0)
bl = out.pop(0)
cost_bl = out.pop(0)
D_id = out.pop(0)
print
print 'Iteration', ii
print 'Baseline Value', bl.mean(), 'Cost', cost_bl
print ' '.join(prm.metrics_map.keys())
print metrics.mean(0)
print
print 'Retrieved Docs: ', str([options['engine'].id_title_map[d_id] for d_id in D_id[0]])
print
print 'Reformulated Query:', options['reformulated_queries'][ii][0]
print
print 'Query ANS: ',
for kk, word in enumerate(options['current_queries'][0][:ans.shape[1]]):
if word not in options['vocab'] and word != '':
word += '<unk>'
if ans[0,kk] == 1:
word = word.upper()
print str(word),
print
print
print 'prob[:,:,0].max(1).mean(), prob[:,:,0].mean(), prob[:,:,0].min(1).mean()', prob[:,:,0].max(1).mean(), prob[:,:,0].mean(), prob[:,:,0].min(1).mean()
print 'prob[:,:,1].max(1).mean(), prob[:,:,1].mean(), prob[:,:,1].min(1).mean()', prob[:,:,1].max(1).mean(), prob[:,:,1].mean(), prob[:,:,1].min(1).mean()
print '==================================================================================\n'
if np.mod(uidx, validFreq) == 0 or uidx == 1:
kf_train = get_minibatches_idx(len(qi_train), prm.batch_size_pred, shuffle=True, max_samples=train_size)
kf_valid = get_minibatches_idx(len(qi_valid), prm.batch_size_pred, shuffle=True, max_samples=valid_size)
kf_test = get_minibatches_idx(len(qi_test), prm.batch_size_pred, shuffle=True, max_samples=test_size)
print '\nEvaluating - Training Set'
train_metrics = pred_error(f_pred, qi_train, dt_train, options, kf_train)
print '\nEvaluating - Validation Set'
valid_metrics = pred_error(f_pred, qi_valid, dt_valid, options, kf_valid)
print '\nEvaluating - Test Set'
test_metrics = pred_error(f_pred, qi_test, dt_test, options, kf_test)
his = [train_metrics, valid_metrics, test_metrics]
history_errs.append(his)
metric_idx = prm.metrics_map[prm.reward.upper()]
if (uidx == 0 or
valid_metrics[-1, metric_idx] >= np.array(history_errs)[:,1,-1,metric_idx].max()):
best_p = unzip(tparams)
bad_counter = 0
print '====================================================================================================='
print ' '.join(prm.metrics_map.keys())
print
print 'Train:'
print train_metrics
print
print 'Valid:'
print valid_metrics
print
print 'Test:'
print test_metrics
print
print '====================================================================================================='
if (len(history_errs) > prm.patience and
valid_metrics[-1, metric_idx] <= np.array(history_errs)[:-prm.patience,
1,-1,metric_idx].max()):
bad_counter += 1
if bad_counter > prm.patience:
print 'Early Stop!'
estop = True
break
if prm.saveto and np.mod(uidx, saveFreq) == 0:
print 'Saving...',
if best_p is not None:
params = best_p
else:
params = unzip(tparams)
np.savez(prm.saveto, history_errs=history_errs, **params)
print 'Done'
print 'Seen %d samples' % n_samples
if estop:
break
except KeyboardInterrupt:
print "Training interupted"
return
from tensorflow.contrib.learn.python.learn.preprocessing import categorical_vocabulary
from sklearn.decomposition import PCA
from QueryReformulatorEnv import QueryReformulatorEnv
import unicodedata
# Parameters
# ==================================================
with open('config.yml', 'r') as ymlfile:
cfg = yaml.load(ymlfile)
# Data Preparation
# ==================================================
# Load data
print 'Loading queries Dataset...'
t0 = time()
dh5 = dataset_hdf5.DatasetHDF5(cfg['data']['query_dataset_path'])
qi_train = dh5.get_queries(dset='train')
dt_train = dh5.get_doc_ids(dset='train')
qi_valid = dh5.get_queries(dset='valid')
dt_valid = dh5.get_doc_ids(dset='valid')
qi_test = dh5.get_queries(dset='test')
dt_test = dh5.get_doc_ids(dset='test')
print("Loading queries and docs {}".format(time() - t0))
print '%d train examples' % len(qi_train)
print '%d valid examples' % len(qi_valid)
print '%d test examples' % len(qi_test)
#print 'qi_train',qi_train
#print 'dt_train',dt_train
# Build vocabulary
t0 = time()