def main():
with tf.device('/gpu:' + str(GPU_ID)):
dis_param = pickle.load(open(DIS_MODEL_PRETRAIN_FILE, 'rb'))
# gen_param = cPickle.load(open(GEN_MODEL_PRETRAIN_FILE))
discriminator = DIS(IMAGE_DIM,
TEXT_DIM,
HIDDEN_DIM,
OUTPUT_DIM,
WEIGHT_DECAY,
D_LEARNING_RATE,
BETA,
GAMMA,
param=dis_param)
generator = GEN(IMAGE_DIM,
TEXT_DIM,
HIDDEN_DIM,
OUTPUT_DIM,
CLASS_DIM,
WEIGHT_DECAY,
G_LEARNING_RATE,
param=None)
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
print('start adversarial training')
map_best_val_gen = 0.0
map_best_val_dis = 0.0
for epoch in range(WHOLE_EPOCH):
print('Training D ...')
for d_epoch in range(D_EPOCH):
print('d_epoch: ' + str(d_epoch))
if d_epoch % GS_EPOCH == 0:
print('negative text sampling for d using g ...')
dis_train_i2t_list = generate_samples(
sess, generator, train_i2t, train_i2t_pos,
train_i2t_neg, 'i2t')
print('negative image sampling for d using g ...')
dis_train_t2i_list = generate_samples(
sess, generator, train_t2i, train_t2i_pos,
train_t2i_neg, 't2i')
# dis_train_i2i_list = generate_samples_pretrain(train_i2i_pos, train_i2i_neg, 'i2i')
# dis_train_t2t_list = generate_samples_pretrain(train_t2t_pos, train_t2t_neg, 't2i')
discriminator = train_discriminator(sess, discriminator,
dis_train_i2t_list, 'i2t')
discriminator = train_discriminator(sess, discriminator,
dis_train_t2i_list, 't2i')
# discriminator = train_discriminator(sess, discriminator, dis_train_i2i_list, 'i2i')
# discriminator = train_discriminator(sess, discriminator, dis_train_t2t_list, 't2t')
if (d_epoch + 1) % (D_DISPLAY) == 0:
i2t_test_map, t2i_test_map, i2i_test_map, t2t_test_map = MAP(
sess, discriminator)
print(
'---------------------------------------------------------------'
)
print('train_I2T_Test_MAP: %.4f' % i2t_test_map)
print('train_T2I_Test_MAP: %.4f' % t2i_test_map)
average_map = 0.5 * (i2t_test_map + t2i_test_map)
if average_map > map_best_val_dis:
map_best_val_dis = average_map
discriminator.save_model(sess, DIS_MODEL_BEST_FILE)
discriminator.save_model(sess, DIS_MODEL_NEWEST_FILE)
print('Training G ...')
for g_epoch in range(G_EPOCH):
print('g_epoch: ' + str(g_epoch))
generator = train_generator(sess, generator, discriminator,
train_i2t, train_i2t_pos, 'i2t')
generator = train_generator(sess, generator, discriminator,
train_t2i, train_t2i_pos, 't2i')
if (g_epoch + 1) % (G_DISPLAY) == 0:
i2t_test_map, t2i_test_map, i2i_test_map, t2t_test_map = MAP(
sess, generator)
print(
'---------------------------------------------------------------'
)
print('train_I2T_Test_MAP: %.4f' % i2t_test_map)
print('train_T2I_Test_MAP: %.4f' % t2i_test_map)
sess.close()
G_LEARNING_RATE = 0.001
TEMPERATURE = 0.2
workdir = 'MQ2008-semi'
GAN_MODEL_BEST_FILE = workdir + '/gan/gan_best_nn.model'
query_url_feature, _, _ =\
ut.load_all_query_url_feature(workdir + '/Large_norm.txt', FEATURE_SIZE)
query_pos_train = ut.get_query_pos(workdir + '/train.txt')
query_pos_test = ut.get_query_pos(workdir + '/test.txt')
param_best = cPickle.load(open(GAN_MODEL_BEST_FILE))
assert param_best is not None
generator_best = GEN(FEATURE_SIZE, HIDDEN_SIZE, WEIGHT_DECAY, G_LEARNING_RATE, temperature=TEMPERATURE, param=param_best)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
p_1_best = precision_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=1)
p_3_best = precision_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=3)
p_5_best = precision_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=5)
p_10_best = precision_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=10)
ndcg_1_best = ndcg_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=1)
ndcg_3_best = ndcg_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=3)
ndcg_5_best = ndcg_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=5)
ndcg_10_best = ndcg_at_k(sess, generator_best, query_pos_test, query_pos_train, query_url_feature, k=10)
def main():
print("load initial model ...")
param_nn = cPickle.load(open(DIS_MODEL_FILE_NN))
assert param_nn is not None
discriminator = DIS(FEATURE_SIZE,
HIDDEN_SIZE,
D_WEIGHT_DECAY,
D_LEARNING_RATE,
loss='log',
param=param_nn)
generator = GEN(FEATURE_SIZE,
HIDDEN_SIZE,
G_WEIGHT_DECAY,
G_LEARNING_RATE,
param=param_nn)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
print('start adversarial training')
p_best_val = 0.0
ndcg_best_val = 0.0
for epoch in range(30):
if epoch > 0:
# G generate negative for D, then train D
print('Training D ...')
generate_for_d(sess, generator, DIS_TRAIN_FILE)
train_size = ut.file_len(DIS_TRAIN_FILE)
for d_epoch in range(30):
index = 1
while True:
if index > train_size:
break
if index + BATCH_SIZE <= train_size + 1:
input_pos, input_neg = ut.get_batch_data(
DIS_TRAIN_FILE, index, BATCH_SIZE)
else:
input_pos, input_neg = ut.get_batch_data(
DIS_TRAIN_FILE, index, train_size - index + 1)
index += BATCH_SIZE
_ = sess.run(discriminator.d_updates,
feed_dict={
discriminator.pos_data: input_pos,
discriminator.neg_data: input_neg
})
p_5 = precision_at_k(sess,
discriminator,
query_pos_test,
query_pos_train,
query_url_feature,
k=5)
ndcg_5 = ndcg_at_k(sess,
discriminator,
query_pos_test,
query_pos_train,
query_url_feature,
k=5)
if p_5 > p_best_val:
p_best_val = p_5
ndcg_best_val = ndcg_5
discriminator.save_model(sess, GAN_MODEL_BEST_FILE)
print("Best: ", "dis p@5 ", p_5, "dis ndcg@5 ", ndcg_5)
elif p_5 == p_best_val:
if ndcg_5 > ndcg_best_val:
ndcg_best_val = ndcg_5
discriminator.save_model(sess, GAN_MODEL_BEST_FILE)
print("Best: ", "dis p@5 ", p_5, "dis ndcg@5 ", ndcg_5)
# Train G
print('Training G ...')
for g_epoch in range(50): # 50
for query in query_pos_train.keys():
pos_list = query_pos_train[query]
# candidate_list = list(set(query_url_feature[query].keys()) - set(pos_list))
candidate_list = list(query_url_feature[query].keys())
if len(candidate_list) <= 0:
continue
candidate_list_feature = [
query_url_feature[query][url] for url in candidate_list
]
candidate_list_feature = np.asarray(candidate_list_feature)
candidate_list_score = sess.run(
generator.pred_score,
{generator.pred_data: candidate_list_feature})
# softmax for all
exp_rating = np.exp(candidate_list_score)
prob = exp_rating / np.sum(exp_rating)
neg_index = np.random.choice(np.arange(len(candidate_list)),
size=[len(pos_list)],
p=prob)
neg_list = np.array(candidate_list)[neg_index]
pos_list_feature = [
query_url_feature[query][url] for url in pos_list
]
neg_list_feature = [
query_url_feature[query][url] for url in neg_list
]
neg_index = np.asarray(neg_index)
# every negative samples have a reward
neg_reward = sess.run(discriminator.reward,
feed_dict={
discriminator.pos_data:
pos_list_feature,
discriminator.neg_data:
neg_list_feature
})
# Method 1: softmax before gather
_ = sess.run(generator.gan_updates,
feed_dict={
generator.pred_data: candidate_list_feature,
generator.sample_index: neg_index,
generator.reward: neg_reward
})
print('Best p@5: ', p_best_val, 'Best ndcg@5: ', ndcg_best_val)
def main():
discriminator = DIS(FEATURE_SIZE,
HIDDEN_SIZE,
WEIGHT_DECAY,
D_LEARNING_RATE,
param=None)
generator = GEN(FEATURE_SIZE,
HIDDEN_SIZE,
WEIGHT_DECAY,
G_LEARNING_RATE,
temperature=TEMPERATURE,
param=None)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
print('start adversarial training')
p_best_val = 0.0
ndcg_best_val = 0.0
for epoch in range(30):
if epoch >= 0:
# G generate negative for D, then train D
print('Training D ...')
for d_epoch in range(100):
if d_epoch % 30 == 0:
generate_for_d(sess, generator, DIS_TRAIN_FILE)
train_size = ut.file_len(DIS_TRAIN_FILE)
index = 1
while True:
if index > train_size:
break
if index + BATCH_SIZE <= train_size + 1:
input_pos, input_neg = ut.get_batch_data(
DIS_TRAIN_FILE, index, BATCH_SIZE)
else:
input_pos, input_neg = ut.get_batch_data(
DIS_TRAIN_FILE, index, train_size - index + 1)
index += BATCH_SIZE
pred_data = []
pred_data.extend(input_pos)
pred_data.extend(input_neg)
pred_data = np.asarray(pred_data)
pred_data_label = [1.0] * len(input_pos)
pred_data_label.extend([0.0] * len(input_neg))
pred_data_label = np.asarray(pred_data_label)
_ = sess.run(discriminator.d_updates,
feed_dict={
discriminator.pred_data: pred_data,
discriminator.pred_data_label:
pred_data_label
})
# Train G
print('Training G ...')
for g_epoch in range(30):
for query in query_pos_train.keys():
pos_list = query_pos_train[query]
pos_set = set(pos_list)
all_list = query_index_url[query]
all_list_feature = [
query_url_feature[query][url] for url in all_list
]
all_list_feature = np.asarray(all_list_feature)
all_list_score = sess.run(
generator.pred_score,
{generator.pred_data: all_list_feature})
# softmax for all
exp_rating = np.exp(all_list_score - np.max(all_list_score))
prob = exp_rating / np.sum(exp_rating)
prob_IS = prob * (1.0 - LAMBDA)
for i in range(len(all_list)):
if all_list[i] in pos_set:
prob_IS[i] += (LAMBDA / (1.0 * len(pos_list)))
choose_index = np.random.choice(np.arange(len(all_list)),
[5 * len(pos_list)],
p=prob_IS)
choose_list = np.array(all_list)[choose_index]
choose_feature = [
query_url_feature[query][url] for url in choose_list
]
choose_IS = np.array(prob)[choose_index] / np.array(
prob_IS)[choose_index]
choose_index = np.asarray(choose_index)
choose_feature = np.asarray(choose_feature)
choose_IS = np.asarray(choose_IS)
choose_reward = sess.run(
discriminator.reward,
feed_dict={discriminator.pred_data: choose_feature})
_ = sess.run(generator.g_updates,
feed_dict={
generator.pred_data: all_list_feature,
generator.sample_index: choose_index,
generator.reward: choose_reward,
generator.important_sampling: choose_IS
})
p_5 = precision_at_k(sess,
generator,
query_pos_test,
query_pos_train,
query_url_feature,
k=5)
ndcg_5 = ndcg_at_k(sess,
generator,
query_pos_test,
query_pos_train,
query_url_feature,
k=5)
if p_5 > p_best_val:
p_best_val = p_5
ndcg_best_val = ndcg_5
generator.save_model(sess, GAN_MODEL_BEST_FILE)
print("Best:", "gen p@5 ", p_5, "gen ndcg@5 ", ndcg_5)
elif p_5 == p_best_val:
if ndcg_5 > ndcg_best_val:
ndcg_best_val = ndcg_5
generator.save_model(sess, GAN_MODEL_BEST_FILE)
print("Best:", "gen p@5 ", p_5, "gen ndcg@5 ", ndcg_5)
sess.close()
param_best = cPickle.load(open(GAN_MODEL_BEST_FILE))
assert param_best is not None
generator_best = GEN(FEATURE_SIZE,
HIDDEN_SIZE,
WEIGHT_DECAY,
G_LEARNING_RATE,
temperature=TEMPERATURE,
param=param_best)
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
p_1_best = precision_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=1)
p_3_best = precision_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=3)
p_5_best = precision_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=5)
p_10_best = precision_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=10)
ndcg_1_best = ndcg_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=1)
ndcg_3_best = ndcg_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=3)
ndcg_5_best = ndcg_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=5)
ndcg_10_best = ndcg_at_k(sess,
generator_best,
query_pos_test,
query_pos_train,
query_url_feature,
k=10)
map_best = MAP(sess, generator_best, query_pos_test, query_pos_train,
query_url_feature)
mrr_best = MRR(sess, generator_best, query_pos_test, query_pos_train,
query_url_feature)
print("Best ", "p@1 ", p_1_best, "p@3 ", p_3_best, "p@5 ", p_5_best,
"p@10 ", p_10_best)
print("Best ", "ndcg@1 ", ndcg_1_best, "ndcg@3 ", ndcg_3_best, "ndcg@5 ",
ndcg_5_best, "p@10 ", ndcg_10_best)
print("Best MAP ", map_best)
print("Best MRR ", mrr_best)
def main():
#call discriminator, generator
discriminator = DIS(FEATURE_SIZE, HIDDEN_SIZE, WEIGHT_DECAY, D_LEARNING_RATE)
generator = GEN(FEATURE_SIZE, HIDDEN_SIZE, WEIGHT_DECAY, G_LEARNING_RATE, temperature=TEMPERATURE)
print('start adversarial training')
p_best_val = 0.0
ndcg_best_val = 0.0
for epoch in range(30):
if epoch >= 0:
# G generate negative for D, then train D
print('Training D ...')
for d_epoch in range(100):
if d_epoch % 30 == 0:
generate_for_d(generator, DIS_TRAIN_FILE)
train_size = ut.file_len(DIS_TRAIN_FILE)
index = 1
while True:
if index > train_size:
break
if index + BATCH_SIZE <= train_size + 1:
input_pos, input_neg = ut.get_batch_data(DIS_TRAIN_FILE, index, BATCH_SIZE)
else:
input_pos, input_neg = ut.get_batch_data(DIS_TRAIN_FILE, index, train_size - index + 1)
index += BATCH_SIZE
pred_data = []
#prepare pos and neg data
pred_data.extend(input_pos)
pred_data.extend(input_neg)
pred_data = np.asarray(pred_data)
#prepara pos and neg label
pred_data_label = [1.0] * len(input_pos)
pred_data_label.extend([0.0] * len(input_neg))
pred_data_label = np.asarray(pred_data_label)
#train
discriminator.train(pred_data, pred_data_label)
# Train G
print('Training G ...')
for g_epoch in range(10):
start_time = time.time()
print ('now_ G_epoch : ', str(g_epoch))
for query in query_pos_train.keys():
pos_list = query_pos_train[query]
pos_set = set(pos_list)
#all url
all_list = query_index_url[query]
#all feature
all_list_feature = [query_url_feature[query][url] for url in all_list]
all_list_feature = np.asarray(all_list_feature)
# G generate all url prob
prob = generator.get_prob(all_list_feature[np.newaxis, :])
prob = prob[0]
prob = prob.reshape([-1])
#important sampling, change doc prob
prob_IS = prob * (1.0 - LAMBDA)
for i in range(len(all_list)):
if all_list[i] in pos_set:
prob_IS[i] += (LAMBDA / (1.0 * len(pos_list)))
# G generate some url (5 * postive doc num)
choose_index = np.random.choice(np.arange(len(all_list)), [5 * len(pos_list)], p=prob_IS)
#choose url
choose_list = np.array(all_list)[choose_index]
#choose feature
choose_feature = [query_url_feature[query][url] for url in choose_list]
#prob / importan sampling prob (loss => prob * reward * prob / importan sampling prob)
choose_IS = np.array(prob)[choose_index] / np.array(prob_IS)[choose_index]
choose_index = np.asarray(choose_index)
choose_feature = np.asarray(choose_feature)
choose_IS = np.asarray(choose_IS)
#get reward((prob - 0.5) * 2 )
choose_reward = discriminator.get_preresult(choose_feature)
#train
generator.train(choose_feature[np.newaxis, :], choose_reward.reshape([-1])[np.newaxis, :], choose_IS[np.newaxis, :])
print("train end--- %s seconds ---" % (time.time() - start_time))
p_5 = precision_at_k(generator, query_pos_test, query_pos_train, query_url_feature, k=5)
ndcg_5 = ndcg_at_k(generator, query_pos_test, query_pos_train, query_url_feature, k=5)
if p_5 > p_best_val:
p_best_val = p_5
ndcg_best_val = ndcg_5
generator.save_model(GAN_MODEL_BEST_FILE)
print("Best:", "gen p@5 ", p_5, "gen ndcg@5 ", ndcg_5)
elif p_5 == p_best_val:
if ndcg_5 > ndcg_best_val:
ndcg_best_val = ndcg_5
generator.save_model(GAN_MODEL_BEST_FILE)
print("Best:", "gen p@5 ", p_5, "gen ndcg@5 ", ndcg_5)
