cand_gen = SimpleQCandidateGenerator(fb_subset=args.fb_subset,
lex_name=lex_name,
wd_emb_util=wd_emb_util,
vb=args.verbose)
for q_idx, qa in enumerate(qa_list):
LogInfo.begin_track('Entering Q %d / %d [%s]:',
q_idx, len(qa_list), qa['utterance'].encode('utf-8'))
sub_idx = q_idx / 1000 * 1000
sub_dir = '%s/data/%d-%d' % (args.output_dir, sub_idx, sub_idx + 999)
if not os.path.exists(sub_dir):
os.makedirs(sub_dir)
opt_sc_fp = '%s/%d_schema' % (sub_dir, q_idx)
link_fp = '%s/%d_links' % (sub_dir, q_idx)
if os.path.isfile(opt_sc_fp):
LogInfo.end_track('Skip this question, already saved.')
continue
Tt.start('single_q')
cand_gen.single_question_candgen(q_idx=q_idx, qa=qa,
link_fp=link_fp, opt_sc_fp=opt_sc_fp)
Tt.record('single_q')
LogInfo.end_track() # End of Q
if __name__ == '__main__':
LogInfo.begin_track('[kangqi.task.compQA.candgen_acl18.simpq_candgen] ... ')
_args = parser.parse_args()
main(_args)
LogInfo.end_track('All Done.')
Tt.display()
def main(args):
LogInfo.begin_track('Learning starts ... ')
# ==== Loading Necessary Util ==== #
LogInfo.begin_track('Loading Utils ... ')
wd_emb_util = WordEmbeddingUtil(wd_emb=args.word_emb, dim_emb=args.dim_emb)
LogInfo.end_track()
# ==== Loading Dataset ==== #
data_config = literal_eval(args.data_config) # including data_name, dir, max_length and others
data_config['wd_emb_util'] = wd_emb_util
# data_config['kb_emb_util'] = kb_emb_util
data_config['verbose'] = args.verbose
dataset = QScDataset(**data_config)
dataset.load_size() # load size info
# ==== Build Model First ==== #
LogInfo.begin_track('Building Model and Session ... ')
gpu_options = tf.GPUOptions(allow_growth=True,
per_process_gpu_memory_fraction=args.gpu_fraction)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
intra_op_parallelism_threads=8))
rm_config = literal_eval(args.rm_config) # Relation Matching
rm_name = rm_config['name']
del rm_config['name']
assert rm_name in ('Compact', 'Separated')
rm_config['n_words'] = dataset.word_size
rm_config['n_mids'] = dataset.mid_size
rm_config['dim_emb'] = args.dim_emb
rm_config['q_max_len'] = dataset.q_max_len
rm_config['sc_max_len'] = dataset.sc_max_len
rm_config['path_max_len'] = dataset.path_max_len
rm_config['pword_max_len'] = dataset.path_max_len * dataset.item_max_len
rm_config['verbose'] = args.verbose
if rm_name == 'Compact':
LogInfo.logs('RelationMatchingKernel: Compact')
rm_kernel = CompactRelationMatchingKernel(**rm_config)
else:
LogInfo.logs('RelationMatchingKernel: Separated')
rm_kernel = SeparatedRelationMatchingKernel(**rm_config)
el_kernel = EntityLinkingKernel(
e_max_size=dataset.e_max_size, e_feat_len=dataset.e_feat_len, verbose=args.verbose)
model_config = literal_eval(args.model_config)
model_config['sess'] = sess
model_config['objective'] = args.eval_mode # relation_only / normal
model_config['relation_kernel'] = rm_kernel
model_config['entity_kernel'] = el_kernel
model_config['extra_len'] = dataset.extra_len
model_config['verbose'] = args.verbose
compq_model = CompqModel(**model_config)
LogInfo.begin_track('Showing final parameters: ')
for var in tf.global_variables():
LogInfo.logs('%s: %s', var.name, var.get_shape().as_list())
LogInfo.end_track()
saver = tf.train.Saver()
LogInfo.begin_track('Parameter initializing ... ')
start_epoch = 0
best_valid_f1 = 0.
resume_flag = False
model_dir = None
if args.resume_model_name not in ('', 'None'):
model_dir = '%s/%s' % (args.output_dir, args.resume_model_name)
if os.path.exists(model_dir):
resume_flag = True
if resume_flag:
start_epoch, best_valid_f1 = load_model(saver=saver, sess=sess, model_dir=model_dir)
else:
dataset.load_init_emb() # loading parameters for embedding initialize
LogInfo.logs('Running global_variables_initializer ...')
sess.run(tf.global_variables_initializer(),
feed_dict={rm_kernel.w_embedding_init: dataset.word_init_emb,
rm_kernel.m_embedding_init: dataset.mid_init_emb})
LogInfo.end_track('Start Epoch = %d', start_epoch)
LogInfo.end_track('Model Built.')
tf.get_default_graph().finalize()
# ==== Constructing Data_Loader ==== #
LogInfo.begin_track('Creating DataLoader ... ')
dataset.load_cands() # first loading all the candidates
if args.eval_mode == 'relation_only':
ro_change = 0
for cand_list in dataset.q_cand_dict.values():
ro_change += add_relation_only_metric(cand_list) # for "RelationOnly" evaluation
LogInfo.logs('RelationOnly F1 change: %d schemas affected.', ro_change)
optm_dl_config = {'dataset': dataset, 'mode': 'train',
'batch_size': args.optm_batch_size, 'proc_ob_num': 5000, 'verbose': args.verbose}
eval_dl_config = dict(optm_dl_config)
spt = args.dl_neg_mode.split('-') # Neg-${POOR_CONTRIB}-${POOR_MAX_SAMPLE}
optm_dl_config['poor_contribution'] = int(spt[1])
optm_dl_config['poor_max_sample'] = int(spt[2])
optm_dl_config['shuffle'] = False
optm_train_data = CompqPairDataLoader(**optm_dl_config)
eval_dl_config['batch_size'] = args.eval_batch_size
eval_data_group = []
for mode in ('train', 'valid', 'test'):
eval_dl_config['mode'] = mode
eval_data = CompqSingleDataLoader(**eval_dl_config)
eval_data.renew_data_list()
eval_data_group.append(eval_data)
(eval_train_data, eval_valid_data, eval_test_data) = eval_data_group
LogInfo.end_track() # End of loading data & dataset
# ==== Free memories ==== #
for item in (wd_emb_util, dataset.wd_emb_util, data_config):
del item
# ==== Ready for learning ==== #
LogInfo.begin_track('Learning start ... ')
output_dir = args.output_dir
if not os.path.exists(output_dir + '/detail'):
os.makedirs(output_dir + '/detail')
if not os.path.exists(output_dir + '/result'):
os.makedirs(output_dir + '/result')
if os.path.isdir(output_dir + '/TB'):
shutil.rmtree(output_dir + '/TB')
tf.summary.FileWriter(output_dir + '/TB/optm', sess.graph) # saving model graph information
# optm_summary_writer = tf.summary.FileWriter(output_dir + '/TB/optm', sess.graph)
# eval_train_summary_writer = tf.summary.FileWriter(output_dir + '/TB/eval_train', sess.graph)
# eval_valid_summary_writer = tf.summary.FileWriter(output_dir + '/TB/eval_valid', sess.graph)
# eval_test_summary_writer = tf.summary.FileWriter(output_dir + '/TB/eval_test', sess.graph)
# LogInfo.logs('TensorBoard writer defined.')
# TensorBoard imformation
status_fp = output_dir + '/status.csv'
with open(status_fp, 'a') as bw:
bw.write('%s\t%s\t%s\t%s\t%s\t%s\t%s\n' % (
'Epoch', 'T_loss', 'T_F1', 'v_F1', 'Status', 't_F1',
datetime.now().strftime('%Y-%m-%d_%H:%M:%S')
))
patience = args.max_patience
for epoch in range(start_epoch + 1, args.max_epoch + 1):
if patience == 0:
LogInfo.logs('Early stopping at epoch = %d.', epoch)
break
LogInfo.begin_track('Epoch %d / %d', epoch, args.max_epoch)
update_flag = False
LogInfo.begin_track('Optimizing ...')
train_loss = compq_model.optimize(optm_train_data, epoch, ob_batch_num=1)
LogInfo.end_track('T_loss = %.6f', train_loss)
LogInfo.begin_track('Eval-Training ...')
train_f1 = compq_model.evaluate(eval_train_data, epoch, ob_batch_num=50,
detail_fp=output_dir + '/detail/train_%03d.txt' % epoch)
LogInfo.end_track('T_F1 = %.6f', train_f1)
LogInfo.begin_track('Eval-Validating ...')
valid_f1 = compq_model.evaluate(eval_valid_data, epoch, ob_batch_num=50,
detail_fp=output_dir + '/detail/valid_%03d.txt' % epoch)
LogInfo.logs('v_F1 = %.6f', valid_f1)
if valid_f1 > best_valid_f1:
best_valid_f1 = valid_f1
update_flag = True
patience = args.max_patience
else:
patience -= 1
LogInfo.logs('Model %s, best v_F1 = %.6f [patience = %d]',
'updated' if update_flag else 'stayed',
valid_f1,
patience)
LogInfo.end_track()
LogInfo.begin_track('Eval-Testing ... ')
test_f1 = compq_model.evaluate(eval_test_data, epoch, ob_batch_num=20,
detail_fp=output_dir + '/detail/test_%03d.txt' % epoch,
result_fp=output_dir + '/result/test_schema_%03d.txt' % epoch)
LogInfo.end_track('t_F1 = %.6f', test_f1)
with open(status_fp, 'a') as bw:
bw.write('%d\t%8.6f\t%8.6f\t%8.6f\t%s\t%8.6f\t%s\n' % (
epoch, train_loss, train_f1, valid_f1,
'UPDATE' if update_flag else str(patience), test_f1,
datetime.now().strftime('%Y-%m-%d_%H:%M:%S')
))
save_epoch_dir = '%s/model_epoch_%d' % (output_dir, epoch)
save_best_dir = '%s/model_best' % output_dir
if args.save_epoch:
delete_dir(save_epoch_dir)
save_model(saver=saver, sess=sess, model_dir=save_epoch_dir, epoch=epoch, valid_metric=valid_f1)
if update_flag and args.save_best: # just create a symbolic link
delete_dir(save_best_dir)
os.symlink(save_epoch_dir, save_best_dir) # symlink at directory level
elif update_flag and args.save_best:
delete_dir(save_best_dir)
save_model(saver=saver, sess=sess, model_dir=save_best_dir, epoch=epoch, valid_metric=valid_f1)
LogInfo.end_track() # End of epoch
LogInfo.end_track() # End of learning
Tt.display()
LogInfo.end_track('All Done.')