def main():
parse = argparse.ArgumentParser()
# ---------- environment setting: which gpu -------
parse.add_argument('-gpu', '--gpu', type=str, default='0', help='which gpu to use: 0 or 1')
# ---------- foler path of train/test data -------
parse.add_argument('-folder', '--folder_path', type=str,
default='datasets/Amazon-670K/data/deeplearning_data/adjacent_labels/',
help='path to train/test data')
# ---------- vocab and word embeddings --------
#parse.add_argument('-vocab', '--vocab_path', type=str, default='vocab.6B.300d.pkl', help='path to the vocab')
parse.add_argument('-word_embeddings', '--word_embedding_path', type=str,
default='word_emb.6B.300d.npy',
help='path to the word embeddings')
# ---------- model ----------
parse.add_argument('-max_seq_len', '--max_seq_len', type=int, default=500,
help='maximum sequence length')
parse.add_argument('-model', '--model', type=str, default='CNN', help='model: LSTM, biLSTM, CNN')
parse.add_argument('-pretrained_model', '--pretrained_model_path', type=str,
default=None, help='path to the pretrained model')
parse.add_argument('-cal_metrics', '--cal_metrics', type=int, default=1,
help='if calculate wts_p and wts_ndcg for baseline results')
parse.add_argument('-alpha', '--alpha', type=float, default=0.2,
help='trade off parameter between baseline score and refinement score')
# ---------- params for CNN ------------
parse.add_argument('-num_filters', '--num_filters', type=int,
default=32, help='number of filters in CNN')
parse.add_argument('-pooling_units', '--pooling_units', type=int,
default=64, help='number of pooling units')
parse.add_argument('-dropout_keep_prob', '--dropout_keep_prob', type=float,
default=0.5, help='keep probability in dropout layer')
filter_sizes = [2, 4, 8]
# ---------- training parameters --------
#parse.add_argument('-if_use_all_true', '--if_use_all_true', type=int, default=0, help='if use all true labels for training')
parse.add_argument('-if_output_all_labels', '--if_output_all_labels', type=int, default=0, help='if output all labels')
parse.add_argument('-n_epochs', '--n_epochs', type=int, default=10, help='number of epochs')
parse.add_argument('-batch_size', '--batch_size', type=int, default=256, help='batch size for training')
parse.add_argument('-batch_pid_size', '--batch_pid_size', type=int, default=5, help='batch pid size for testing')
#parse.add_argument('-num_candidate', '--num_candidate', type=int, default=20, help='number of candidate labels')
#parse.add_argument('-topk', '--topk', type=int, default=6, help='k in competitive layer')
parse.add_argument('-show_batches', '--show_batches', type=int,
default=1000, help='show how many batches have been processed.')
parse.add_argument('-lr', '--learning_rate', type=float, default=0.0001, help='learning rate')
parse.add_argument('-update_rule', '--update_rule', type=str, default='adam', help='update rule')
# ------ train or predict -------
parse.add_argument('-train', '--train', type=int, default=1, help='if training')
parse.add_argument('-test', '--test', type=int, default=0, help='if testing')
parse.add_argument('-predict', '--predict', type=int, default=0, help='if predicting')
args = parse.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
print('-------------- load vocab and word embeddings ---------------')
word_embeddings = np.load('datasets/material/' + args.word_embedding_path)
# word_embedding_dim
word_embedding_dim = word_embeddings.shape[-1]
print('word_embeddings shape: ' + str(word_embeddings.shape))
# add '<PAD>' embedding
word_embeddings = np.concatenate((np.zeros((1, word_embedding_dim)), word_embeddings), axis=0)
print('after add PAD embedding, word_embeddings shape:' + str(word_embeddings.shape))
print('-------------- load label embeddings ------------------------')
all_labels, label_embeddings = generate_label_embedding_from_file_2(args.folder_path + 'label.embeddings')
label_embeddings = np.array(label_embeddings)
label_dict = dict(zip(all_labels, range(len(all_labels))))
print('number of labels: ' + str(len(all_labels)))
# label_embedding_dim
label_embedding_dim = len(label_embeddings[all_labels[0]])
print('-------------- load label propensity ------------------------')
label_prop = load_pickle(args.folder_path + 'inv_prop_dict.pkl')
print('-------------- load train/test data -------------------------')
train_doc = load_pickle(args.folder_path + 'train_doc_wordID.pkl')
test_doc = load_pickle(args.folder_path + 'test_doc_wordID.pkl')
train_label = load_pickle(args.folder_path + 'train_label.pkl')
test_label = load_pickle(args.folder_path + 'test_label.pkl')
print('-------------- load candidate labels ------------------------')
if 'sleec' in args.model:
candidate_type = 'sleec'
elif 'pfastrexml' in args.model:
candidate_type = 'pfastrexml'
elif 'pfastxml' in args.model:
candidate_type = 'pfastxml'
elif 'fastxml' in args.model:
candidate_type = 'fastxml'
print('candidate from: ' + candidate_type)
candidate_folder_path = args.folder_path + candidate_type + '_candidate/'
train_candidate_label = load_pickle(candidate_folder_path + 'train_candidate_label.pkl')
test_candidate_label = load_pickle(candidate_folder_path + 'test_candidate_label.pkl')
print('============== create train/test data loader ...')
if 'XML' not in args.model:
train_loader = TrainDataLoader2(train_doc, train_label, train_candidate_label, label_dict, label_prop,
10, 10, max_seq_len=args.max_seq_len)
max_seq_len = train_loader.max_seq_len
#max_seq_len = args.max_seq_len
#train_loader = {}
print('max_seq_len: ' + str(max_seq_len))
test_loader = TestDataLoader2(test_doc, test_label, test_candidate_label, label_dict, label_prop,
max_seq_len=max_seq_len, if_cal_metrics=args.cal_metrics)
# test_loader = DataLoader3(test_doc, test_label, test_candidate_label, label_dict, args.batch_size,
# given_seq_len=True, max_seq_len=max_seq_len)
# ----------------------- train ------------------------
print('============== build model ...')
if 'biLSTM' in args.model:
print('build biLSTM model ...')
# biLSTM: max_seq_len, word_embedding_dim, hidden_dim, label_embedding_dim, num_classify_hidden, args.batch_size
model = biLSTM(max_seq_len, word_embedding_dim, 64, label_embedding_dim, 32, args)
args.if_use_seq_len = 1
elif 'LSTM' in args.model:
print('build LSTM model ...')
# LSTM: max_seq_len, word_embedding_dim, hidden_dim, label_embedding_dim, num_classify_hidden, args.batch_size
model = LSTM(max_seq_len, word_embedding_dim, 64, label_embedding_dim, 32, args)
args.if_use_seq_len = 1
elif 'CNN' in args.model:
print('build CNN model ...')
# CNN: sequence_length, word_embeddings, filter_sizes, label_embeddings, num_classify_hidden, args
# args.num_filters, args.pooling_units, args.batch_size, args.dropout_keep_prob
model = CNN(max_seq_len, word_embeddings, filter_sizes, label_embeddings, 32, args)
args.if_use_seq_len = 0
elif 'XML' in args.model:
print('build XML-CNN model ...')
train_loader = DataLoader5(train_doc, train_label, all_labels,
args.batch_size,
given_seq_len=False, max_seq_len=args.max_seq_len)
max_seq_len = train_loader.max_seq_len
test_loader = DataLoader5(test_doc, test_label, all_labels,
args.batch_size,
given_seq_len=True, max_seq_len=max_seq_len)
# max_seq_len, word_embedding, filter_sizes, label_output_dim, hidden_dim, args
# args.num_filters, args.pooling_units, args.batch_size, args.dropout_keep_prob
model = XML_CNN(max_seq_len, word_embeddings, filter_sizes, len(all_labels), 128, args)
args.if_output_all_labels = 1
args.if_use_seq_len = 0
print('================= model solver ...')
# solver: __init__(self, model, train_data, test_data, **kwargs):
solver = ModelSolver(model, train_loader, test_loader,
if_use_seq_len=args.if_use_seq_len,
if_output_all_labels=args.if_output_all_labels,
show_batches=args.show_batches,
n_epochs=args.n_epochs,
batch_size=args.batch_size,
update_rule=args.update_rule,
learning_rate=args.learning_rate,
pretrained_model=args.pretrained_model_path,
model_path=args.folder_path + args.model + '/',
test_path=args.folder_path + args.model + '/')
# train
if args.train:
print('================= begin training...')
solver.train(args.folder_path + args.model + '/outcome.txt')
# test
if args.test:
print('================= begin testing...')
solver.test(args.folder_path + args.model + '/' + args.pretrained_model_path, args.folder_path + args.model + '/test_outcome.txt')
# predict
if args.predict:
print('================= begin predicting...')
predict_path = args.folder_path+'model_save/'+args.model+'/'
solver.predict(trained_model_path=predict_path,
output_file_path=predict_path+'predict_outcome.txt',
k=10, emb_saved=1, can_saved=1)
def main():
parse = argparse.ArgumentParser()
# ---------- environment setting: which gpu -------
parse.add_argument('-gpu', '--gpu', type=str, default='0', help='which gpu to use: 0 or 1')
# ---------- foler path of train/test data -------
parse.add_argument('-valid_labels', '--valid_labels', type=int,
default=0, help='-if remove invalid labels')
parse.add_argument('-data', '--data', type=str, default='eurlex',
help='dataset')
# ---------- model ----------
parse.add_argument('-word_embedding_dim', '--word_embedding_dim', type=int, default=100, help='dim of word embedding')
parse.add_argument('-aggr_type', '--aggr_type', type=str, default='sum', help='aggregation type in embedding layer')
parse.add_argument('-vocab_size', '--vocab_size', type=int, default=5000, help='vocabulary size')
parse.add_argument('-max_seq_len', '--max_seq_len', type=int, default=500, help='maximum sequence length')
parse.add_argument('-model', '--model', type=str, default='NN', help='model: NN, LSTM, biLSTM, CNN')
parse.add_argument('-pretrained_model', '--pretrained_model_path', type=str, default=None, help='path to the pretrained model')
parse.add_argument('-dropout_keep_prob', '--dropout_keep_prob', type=float,
default=0.5, help='keep probability in dropout layer')
parse.add_argument('-saved_dis', '--saved_dis', type=int, default=0,
help='whether to use saved dis matrix')
parse.add_argument('-use_sne', '--use_sne', type=int, default=1,
help='whether to use sne regularization')
parse.add_argument('-ac_lbl_ratio', '--ac_lbl_ratio', type=float, default=0.5,
help='ratio of active labels in sne regularization')
parse.add_argument('-use_propensity', '--use_propensity', type=int, default=1,
help='whether to use propensity loss')
parse.add_argument('-use_bi_inter', '--use_bi_inter', type=int, default=0,
help='whether to add bi-interactive layer')
parse.add_argument('-use_comp', '--use_comp', type=int, default=0,
help='whether to add competitive layer')
parse.add_argument('-topk', '--topk', type=int, default=10,
help='top k neurons in competitive layer')
parse.add_argument('-factor', '--factor', type=float, default=0.01,
help='factor in competitive layer')
parse.add_argument('-lamb', '--lamb', type=float, default=0.002,
help='lambda for weight regularization')
# ---------- training parameters --------
parse.add_argument('-n_epochs', '--n_epochs', type=int, default=10, help='number of epochs')
parse.add_argument('-batch_size', '--batch_size', type=int, default=32, help='batch size for training')
parse.add_argument('-lr', '--learning_rate', type=float, default=0.002, help='learning rate')
parse.add_argument('-update_rule', '--update_rule', type=str, default='adam', help='update rule')
# ------ train or predict -------
parse.add_argument('-train', '--train', type=int, default=1, help='if training')
parse.add_argument('-test', '--test', type=int, default=0, help='if testing')
parse.add_argument('-predict', '--predict', type=int, default=0, help='if predicting')
args = parse.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
print '-------------- load labels ------------------------'
# default='datasets/eurlex/trn_tst_data/'
if args.valid_labels:
args.folder_path = 'datasets/' + args.data + '/trn_tst_data/valid_label_data/'
else:
args.folder_path = 'datasets/' + args.data + '/trn_tst_data/all_label_data/'
#label_prop_dict = load_pickle(args.folder_path + 'inv_prop_dict.pkl')
label_prop = load_pickle(args.folder_path + 'inv_prop.pkl')
label_dict = load_pickle(args.folder_path + 'label_dict.pkl')
all_labels = load_pickle(args.folder_path + 'index_label.pkl')
num_labels = len(all_labels)
print 'real number of labels: ' + str(num_labels)
print 'maximum label: ' + str(np.max(all_labels))
print 'minimum label: ' + str(np.min(all_labels))
print 'number of labels: ' + str(num_labels)
print '-------------- load train/test data -------------------------'
train_doc = load_pickle(args.folder_path + 'train_doc_wordID.pkl')
test_doc = load_pickle(args.folder_path + 'test_doc_wordID.pkl')
train_label = load_pickle(args.folder_path + 'train_label.pkl')
test_label = load_pickle(args.folder_path + 'test_label.pkl')
print '============== create train/test data loader ...'
train_loader = DataLoader_all(train_doc, train_label, label_dict, label_prop,
batch_size=args.batch_size, max_seq_len=args.max_seq_len, ac_lbl_ratio=args.ac_lbl_ratio, folder_path=args.folder_path)
test_loader = DataLoader_all(test_doc, test_label, label_dict, label_prop,
batch_size=args.batch_size, max_seq_len=args.max_seq_len, folder_path=args.folder_path)
print '============== build model ...'
print 'build NN model ...'
model = NN(args.max_seq_len, args.vocab_size, args.word_embedding_dim, num_labels, label_prop, 32, args)
args.if_use_seq_len = 1
print '================= model solver ...'
solver = ModelSolver(model, train_loader, test_loader,
n_epochs=args.n_epochs,
batch_size=args.batch_size,
update_rule=args.update_rule,
learning_rate=args.learning_rate,
pretrained_model=args.pretrained_model_path,
model_path=args.folder_path + args.model + '/',
log_path=args.folder_path + args.model + '/',
test_path=args.folder_path + args.model + '/',
use_sne=args.use_sne,
saved_dis=args.saved_dis
)
# train
if args.train:
print '================= begin training...'
solver.train(args.folder_path + args.model + '/outcome.txt')
# test
if args.test:
print '================= begin testing...'
solver.test(args.folder_path + args.model + '/' + args.pretrained_model_path, args.folder_path + args.model + '/test_outcome.txt')
# predict
if args.predict:
print '================= begin predicting...'
predict_path = args.folder_path+'model_save/'+args.model+'/'
solver.predict(trained_model_path=predict_path,
output_file_path=predict_path+'predict_outcome.txt',
k=10, emb_saved=1, can_saved=1)