def autotrain(train_data_list, val_data_list, test_data_list, vgg_features_neb, bow_neb): import model BATCH_SIZE = 128 LAMBDA = 1 #data_list = np.load('data_list.npy') bow = np.load('bow.npy') labels = np.load('label.npy') vgg_features = np.load('feature.npy') vgg_features = np.tanh((vgg_features) / np.std(vgg_features)) images = tf.placeholder(tf.float32, [None, 4096]) target_embedding = tf.placeholder(tf.float32, [None, 64]) # images_target = tf.placeholder(tf.float32, [BATCH_SIZE, 4096]) images_drop = tf.nn.dropout(images, 0.5) texts = tf.placeholder(tf.float32, [None, 2000]) # texts_target = tf.placeholder(tf.float32, [BATCH_SIZE, 1386]) texts_drop = tf.nn.dropout(texts, 0.5) model = model.luo(BATCH_SIZE, 64) text_embedding = model.txt_encoder(texts_drop) image_embedding = model.image_encoder(images_drop) text_embedding_t = model.txt_encoder(texts, reuse=True) image_embedding_t = model.image_encoder(images, reuse=True) mixed_embedding = (text_embedding + image_embedding) / 2 mixed_embedding = tf.nn.l2_normalize(mixed_embedding, 1) pre_distance = [] for i in range(BATCH_SIZE): pre_distance.append(tf.reduce_sum(mixed_embedding[i] * mixed_embedding, 1)) pre_distance = tf.stack(pre_distance) mean = tf.reduce_mean(pre_distance) var = tf.reduce_mean(tf.square(pre_distance - mean)) var_loss = tf.square(tf.clip_by_value(0.04- var, 0, 0.04)) re_image_text = model.txt_decoder_image(text_embedding, 4096) re_text_image = model.image_decoder_text(image_embedding, 2000) re_image_text_drop = tf.nn.dropout(re_image_text, 0.5) re_text_image_drop = tf.nn.dropout(re_text_image, 0.5) # disc_real = model.disc(images, texts) disc_fake_image = model.disc(re_image_text_drop, texts, reuse=False , drop_rate=1.0) disc_fake_text = model.disc(images, re_text_image_drop, reuse=True, drop_rate=1.0) loss_image = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake_image, labels=tf.ones_like(disc_fake_image))) loss_text = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake_text, labels=tf.ones_like(disc_fake_text))) gen_cost = loss_image + loss_text consistency_loss = tf.reduce_mean(tf.square(text_embedding - image_embedding)) vars = tf.trainable_variables() gen_params = [v for v in vars if 'coder' in v.name] disc_params = [v for v in vars if 'disc' in v.name] gen_train_op = tf.train.AdamOptimizer(learning_rate=1e-4, beta1=0.5, beta2=0.9).minimize( consistency_loss + gen_cost + 100 * var_loss, var_list=gen_params) def test(): mAP_n = 10000 P_n = 1000 def check(x, y): flag = bool(np.sum(np.logical_and(x, y))) return flag def countMAP(result, train_label, po): AP = 0 total_relevant = 0 buffer_yes = np.zeros(mAP_n) Ns = np.arange(1, mAP_n + 1, 1) for i in range(mAP_n): if check(train_label[result[i]], po): buffer_yes[i] = 1 total_relevant += 1 P = np.cumsum(buffer_yes) / Ns if sum(buffer_yes) != 0: AP += sum(P * buffer_yes) / sum(buffer_yes) return AP def countP(result, train_label, po): P = 0 for i in range(P_n): if check(train_label[result[i]], po): P += 1 return P / P_n def eval_total(binary_codes_train, binary_codes_test, train_label, test_label): lens1 = len(binary_codes_test) lens2 = len(binary_codes_train) P = 0 mAP = 0 dist = np.zeros((lens1, lens2), dtype=np.float32) binary_codes_train_norm = np.linalg.norm(binary_codes_train, axis=1) binary_codes_test_norm = np.linalg.norm(binary_codes_test, axis=1) for i in range(lens1): dist[i] = -np.sum(binary_codes_test[i] * binary_codes_train, 1) / binary_codes_train_norm / \ binary_codes_test_norm[i] results = [] for i in range(lens1): results.append(np.argsort(dist[i])) for i in range(lens1): mAP += countMAP(results[i], train_label, test_label[i]) P += countP(results[i], train_label, test_label[i]) return P / lens1, mAP / lens1 total_train_embedding = [] total_test_embedding = [] for i in range(int(len(train_data_list) / 1000)): train_images = vgg_features[train_data_list[i * 1000:min((i + 1) * 1000, len(train_data_list))]] _image_embedding = session.run(image_embedding_t, feed_dict={images: train_images}) total_train_embedding.append(_image_embedding) total_train_embedding = np.concatenate(total_train_embedding, axis=0) for i in range(int(len(val_data_list) / 1000)): test_texts = bow[val_data_list[i * 1000:min((i + 1) * 1000, len(val_data_list))]] _text_embedding = session.run(text_embedding_t, feed_dict={texts: test_texts}) total_test_embedding.append(_text_embedding) total_test_embedding = np.concatenate(total_test_embedding, axis=0) P, mAP = eval_total(total_train_embedding, total_test_embedding, labels[train_data_list], labels[val_data_list]) print('mAP') print(mAP) print('P') print(P) return mAP config = tf.ConfigProto() config.gpu_options.allow_growth = True with tf.Session(config=config) as session: session.run(tf.initialize_all_variables()) saver = tf.train.Saver() saver2 = tf.train.Saver(disc_params) saver2.restore(session, 'disc_block') old_map = 0 for i in range(50001): train_list = np.random.randint(0, len(train_data_list), BATCH_SIZE) train_images = vgg_features[train_data_list[train_list]] # _target_embedding = np.random.uniform(-1, 1, [BATCH_SIZE, 64]) changer = np.random.randint(0, 2, [BATCH_SIZE, 64], np.int8) _target_embedding = np.clip( changer * np.random.normal(0.7, 0.2, [BATCH_SIZE, 64]) + (1 - changer) * np.random.normal(-0.7, 0.2, [BATCH_SIZE, 64]), -1, 1) train_texts = bow[train_data_list[train_list]] # train_images_list = np.zeros_like(train_list, np.int32) session.run(gen_train_op, feed_dict={images: train_images, texts: train_texts, target_embedding: _target_embedding}) #session.run(disc_train_op, # feed_dict={images: train_images, texts: train_texts, target_embedding: _target_embedding}) if (i % 2000 == 0 and i >= 30000): new_map = test() if new_map > old_map: old_map = new_map saver.save(session, './referee_advKNN') if (i % 100 == 0): _closs, _gloss, _var, _t, _i, _d_loss = session.run( [consistency_loss, gen_cost, var, loss_text, loss_image, var_loss], feed_dict={images: train_images, texts: train_texts, target_embedding: _target_embedding}) print("consistency_loss:%f, gen_loss:%f, var:%f, var_loss:%f" % (_closs, _gloss, _var, _d_loss)) print('image:%f, text:%f' % (_i, _t)) print('best map') print(old_map) session.close() return old_map
def binary_train(code_len, rounds): import model BATCH_SIZE = 128 tf_idf = np.load('bow.npy') vgg_features = np.load('feature.npy') vgg_features = np.tanh((vgg_features) / np.std(vgg_features)) images = tf.placeholder(tf.float32, [None, 4096]) texts = tf.placeholder(tf.float32, [None, 2000]) embeddings = tf.placeholder(tf.float32, [BATCH_SIZE, 64]) dist = tf.placeholder(tf.float32, [BATCH_SIZE, BATCH_SIZE]) model = model.luo(BATCH_SIZE, 64) text_embedding = model.txt_encoder(texts) image_embedding = model.image_encoder(images) embeddings_drop = tf.nn.dropout(embeddings, 0.5) binary_vec_drop = model.binary_layer(embeddings_drop, code_len, rate=0.5) loss = 0 for i in range(BATCH_SIZE): loss += tf.reduce_mean(tf.square(dist[i, :] - tf.reduce_mean(binary_vec_drop[i] * binary_vec_drop, 1))) loss /= BATCH_SIZE lossb = tf.square(1 - tf.reduce_mean(binary_vec_drop * binary_vec_drop)) vars = tf.trainable_variables() b_params = [v for v in vars if 'binary' in v.name] o_params = [v for v in vars if 'binary' not in v.name] b_train_op = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(loss, var_list=b_params) config = tf.ConfigProto() config.gpu_options.allow_growth = True train_data_list = np.load('train_data_list.npy') total_train_embedding_images = [] total_train_embedding_texts = [] with tf.Session(config=config) as session: saver_old = tf.train.Saver(o_params) saver = tf.train.Saver() session.run(tf.initialize_all_variables()) saver_old.restore(session, './referee_advKNN') for i in range(int(len(train_data_list) / 1000)): train_images = vgg_features[train_data_list[i * 1000:min((i + 1) * 1000, len(train_data_list))]] _image_embedding = session.run(image_embedding, feed_dict={images: train_images}) total_train_embedding_images.append(_image_embedding) total_train_embedding_images = np.concatenate(total_train_embedding_images, axis=0) for i in range(int(len(train_data_list) / 1000)): train_texts = tf_idf[train_data_list[i * 1000:min((i + 1) * 1000, len(train_data_list))]] _text_embedding = session.run(text_embedding, feed_dict={texts: train_texts}) total_train_embedding_texts.append(_text_embedding) total_train_embedding_texts = np.concatenate(total_train_embedding_texts, axis=0) total_train_embedding = (total_train_embedding_texts + total_train_embedding_images) / 2 total_dist = np.zeros((len(train_data_list), len(train_data_list)), dtype=np.float32) norm = np.linalg.norm(total_train_embedding, axis=1) for i in range(len(train_data_list)): total_dist[i, i:] = np.sum(total_train_embedding[i] * total_train_embedding[i:], 1) / norm[i:] / norm[i] total_dist[i:, i] = total_dist[i, i:] tempstd2 = total_dist.std() total_dist = np.clip((-2/np.mean(np.min(total_dist-1, 1))) * (total_dist - 1) + 1, -1, 1) tempstd = total_dist.std() for i in range(rounds): train_list = np.random.randint(0, len(train_data_list), BATCH_SIZE) train_embeddings = total_train_embedding[train_list] train_dist = np.zeros([BATCH_SIZE, BATCH_SIZE], np.float32) for k in range(BATCH_SIZE): train_dist[k] = total_dist[train_list[k]][train_list] session.run(b_train_op, feed_dict={embeddings: train_embeddings, dist: train_dist}) if (i % 10000 == 0): saver.save(session, './referee_advKNNB') if (i % 100 == 0): _loss, _lossB = session.run([loss, lossb], feed_dict={embeddings: train_embeddings, dist: train_dist}) print(_loss) print(_lossB) session.close() return tempstd, tempstd2
def embeddingB(code_len): import model mAP_n = 10000 P_n = 1000 def HammingDistance(a, b): # c = np.logical_xor((np.sign(a)+1).astype(np.bool_),(np.sign(b)+1).astype(np.bool_)) # dis = np.sum(c.astype(np.int32)) dis = np.sum(np.square(a - b)) return dis def check(x, y): flag = bool(np.sum(np.logical_and(x, y))) return flag def countMAP(result, train_label, po): AP = 0 total_relevant = 0 buffer_yes = np.zeros(mAP_n) Ns = np.arange(1, mAP_n + 1, 1) for i in range(mAP_n): if check(train_label[result[i]], po): buffer_yes[i] = 1 total_relevant += 1 P = np.cumsum(buffer_yes) / Ns if sum(buffer_yes) != 0: AP += sum(P * buffer_yes) / sum(buffer_yes) return AP def countP(result, train_label, po): P = 0 for i in range(P_n): if check(train_label[result[i]], po): P += 1 return P / P_n def eval_total(binary_codes_train, binary_codes_test, train_label, test_label): lens1 = len(binary_codes_test) lens2 = len(binary_codes_train) P = 0 mAP = 0 dist = np.zeros((lens1, lens2), dtype=np.float32) binary_codes_train_norm = np.linalg.norm(binary_codes_train, axis=1) binary_codes_test_norm = np.linalg.norm(binary_codes_test, axis=1) for i in range(lens1): dist[i] = -np.sum(binary_codes_test[i] * binary_codes_train, 1) / binary_codes_train_norm / \ binary_codes_test_norm[i] results = [] for i in range(lens1): results.append(np.argsort(dist[i])) for i in range(lens1): mAP += countMAP(results[i], train_label, test_label[i]) P += countP(results[i], train_label, test_label[i]) return P / lens1, mAP / lens1 BATCH_SIZE = 128 bow = np.load('bow.npy') labels = np.load('label.npy') vgg_features = np.load('feature.npy') vgg_features = np.tanh((vgg_features) / np.std(vgg_features)) images = tf.placeholder(tf.float32, [None, 4096]) texts = tf.placeholder(tf.float32, [None, 2000]) model = model.luo(BATCH_SIZE, 64) text_embedding = model.txt_encoder(texts) text_embedding = tf.sign(model.binary_layer(text_embedding, code_len)) image_embedding = model.image_encoder(images) image_embedding = tf.sign(model.binary_layer(image_embedding, code_len, reuse=True)) config = tf.ConfigProto() config.gpu_options.allow_growth = True train_data_list = np.load('train_data_list.npy') test_data_list = np.load('test_data_list.npy') total_train_embedding = [] total_test_embedding = [] with tf.Session(config=config) as session: saver = tf.train.Saver() saver.restore(session, './referee_advKNNB') for i in range(int(len(train_data_list) / 1000)): train_images = vgg_features[train_data_list[i * 1000:min((i + 1) * 1000, len(train_data_list))]] _image_embedding = session.run(image_embedding, feed_dict={images: train_images}) total_train_embedding.append(_image_embedding) total_train_embedding = np.concatenate(total_train_embedding, axis=0) for i in range(int(len(test_data_list) / 1000)): test_texts = bow[test_data_list[i * 1000:min((i + 1) * 1000, len(test_data_list))]] _text_embedding = session.run(text_embedding, feed_dict={texts: test_texts}) total_test_embedding.append(_text_embedding) total_test_embedding = np.concatenate(total_test_embedding, axis=0) session.close() P, mAP = eval_total(total_train_embedding, total_test_embedding, labels[train_data_list], labels[test_data_list]) print('mAP') print(mAP) print('P') print(P) return P, mAP
results.append(np.argsort(dist[i])) for i in range(lens1): mAP += countMAP(results[i], train_label, test_label[i]) P += countP(results[i], train_label, test_label[i]) return P/lens1, mAP/lens1 BATCH_SIZE = 128 tf_idf = np.load('bow.npy') labels = np.load('label.npy') vgg_features = np.load('feature.npy') vgg_features = np.tanh((vgg_features)/np.std(vgg_features)) images = tf.placeholder(tf.float32, [None, 4096]) texts = tf.placeholder(tf.float32, [None, 2000]) model = model.luo(BATCH_SIZE,64) text_embedding = model.txt_encoder(texts) image_embedding = model.image_encoder(images) config = tf.ConfigProto() config.gpu_options.allow_growth = True train_data_list = np.load('train_data_list.npy') test_data_list = np.load('test_data_list.npy') total_train_embedding = [] total_test_embedding = [] with tf.Session(config=config) as session: saver = tf.train.Saver() saver.restore(session, './referee_ae1.50_0.05') for i in range(int(len(train_data_list)/1000)): train_images = vgg_features[train_data_list[i*1000:min((i+1)*1000, len(train_data_list))]] _image_embedding = session.run(image_embedding, feed_dict={images:train_images}) total_train_embedding.append(_image_embedding) total_train_embedding = np.concatenate(total_train_embedding, axis=0) for i in range(int(len(test_data_list)/1000)):