qa_data = load_data2()
print('done loading data...\n\n')
batch_no = 0
batch_size = 50
#while batch_no*batch_size < len(qa_data['training']):
for eidx in range(20):
#while batch_no < 6500:
#if eidx == 0:
# batch_no = 6500
#else:
batch_no = 0
while batch_no * batch_size < len(qa_data['training']) - batch_size:
print('batch = ' + str(batch_no))
(questions_in_true, answer_in_true,
im_feat_true) = get_training_batch(batch_no, batch_size, qa_data)
if batch_no * batch_size < len(qa_data['training']) - 1:
(questions_in_false, answer_in_false,
im_feat_false) = get_training_batch(batch_no + 1, batch_size,
qa_data)
else:
(questions_in_false, answer_in_false,
im_feat_false) = get_training_batch(0, batch_size, qa_data)
_, loss_val, sr, sw, sf = sess.run(
[train_op_d, loss, s_r, s_w, s_f],
feed_dict={
answers_true: answer_in_true,
image: im_feat_true,
question: questions_in_true,
saver.restore(sess, "./train_g_v3.3")
#quit()
print('answers 1000....')
print('loading data...\n\n')
qa_data = load_data2()
print('done loading data...\n\n')
batch_size = 50
#while batch_no*batch_size < len(qa_data['training']):
for train_loops in range(10):
batch_no = 0
while batch_no * batch_size < len(qa_data['training']):
print('batch = ' + str(batch_no))
(questions_in_true, answer_in_true,
im_feat_true) = get_training_batch(batch_no, batch_size, qa_data)
noise_in = np.random.normal(scale=0.3, size=[batch_size, 2048])
_, loss_val, g_out = sess.run(
[train_op, loss, scores_emb],
feed_dict={
noise: noise_in,
answers_true: answer_in_true,
image: im_feat_true,
question: questions_in_true,
})
print('loss = ' + str(loss_val))
loss_vals.append(loss_val)
#np.save('loss_vals_g_v' + save_ver, loss_vals)
train_step = tf.train.AdamOptimizer(1e-4).minimize(loss)
# init variables
sess.run(tf.global_variables_initializer())
print('loading data...\n\n')
qa_data = load_data()
print('done loading data...\n\n')
#batch_no = 0
batch_size = 50
while batch_no*batch_size < len(qa_data['training']):
print('batch = ' + str(batch_no))
(questions_in_true, answer_in_true, im_feat_true) = get_training_batch(batch_no, batch_size, qa_data)
noise_in = np.random.normal(size=[batch_size,2048])
train_step.run(feed_dict={
noise: noise_in,
answers_true: answer_in_true,
cnn_out_true: im_feat_true,
questions_true: questions_in_true,
})
loss_val = sess.run(loss, feed_dict={
noise: noise_in,
answers_true: answer_in_true,
cnn_out_true: im_feat_true,
questions_true: questions_in_true,