pic_test_x = pic_test_data['N_MNIST_pic_test'].astype('float32')
print('pic_test_x: ', pic_test_x.shape)
in_imgs = pic_test_x
#num_selected = 200
#test_idx = np.linspace(0,len(pic_test_x)-1,num_selected).astype('int32')
#in_imgs = pic_test_x[test_idx]
#gt_imgs = pic_test_y[test_idx]
# In[]:
# prediction
ind = 0
mean_cost = 0
time_cost = 0
reconstructed = np.zeros(in_imgs.shape, dtype='float32')
for batch_x, _ in my_io.batch_iter(test_batch_size,
in_imgs,
in_imgs,
shuffle=False):
x = batch_x.reshape((-1, *pic_size, 1))
feed_dict = {inputs_: x, keep_prob: 1.0, mask_prob: 0.0} #for dropout
# feed_dict = {inputs_: x, targets_: y} #for non dropout
time1 = time()
res_imgs = sess.run(outputs_, feed_dict=feed_dict)
time2 = time()
time_cost += (time2 - time1)
res_imgs = np.squeeze(res_imgs)
reconstructed[ind * test_batch_size:(ind + 1) * test_batch_size] = res_imgs
ind += 1
time_cost = time_cost / len(in_imgs)
print('\nmean time cost(ms):%f\n' % (time_cost * 1e3))
# In[]:
# train
test_feed_dict = {
inputs_: test_x1,
targets_: test_y1,
keep_prob: 1.0,
mask_prob: 0.0
}
summaryWriter(writer_tr, writer_te, merged, cost, test_feed_dict,
test_feed_dict, 0)
time_start = time()
for e in range(1, 1 + epochs):
for batch_x, batch_y in my_io.batch_iter(batch_size,
train_x,
train_y,
throw_insufficient=True):
x = batch_x.reshape((-1, *pic_size, 1))
y = batch_y.reshape((-1, *pic_size, 1))
train_feed_dict = {
inputs_: x,
targets_: y,
keep_prob: keep_prob_v,
mask_prob: mask_prob_v
}
sess.run(optimizer, feed_dict=train_feed_dict)
if e % 5 == 0:
time_cost = time() - time_start
# tensorboard
writer_tr = tf.summary.FileWriter(train_log_dir, sess.graph)
writer_te = tf.summary.FileWriter(test_log_dir)
merged = tf.summary.merge_all()
# saver
saver = tf.train.Saver()
# original performance
cost_v = sess.run(cost, feed_dict={inputs_: test_x1, targets_:test_y1, keep_prob:1.0, mask_prob:0.0})
print('\nfinetune mode:\n\nimport model: ', ckpt_states[model_ind],
'\ncurrent test cost: ', cost_v, '\n\nfinetuning start!\n')
for e in range(1, 1+epochs):
for batch_x, batch_y in my_io.batch_iter(batch_size, train_x, train_y):
x = batch_x.reshape((-1, *pic_size, 1))
y = batch_y.reshape((-1, *pic_size, 1))
sess.run(training_op, feed_dict={inputs_: x, targets_: y, keep_prob:keep_prob_v, mask_prob:mask_prob_v})
if e%1 == 0:
tr, tr_cost = sess.run([merged, cost], feed_dict={inputs_: x, targets_: y, keep_prob:keep_prob_v,mask_prob:mask_prob_v})
te, te_cost = sess.run([merged, cost], feed_dict={inputs_: test_x1, targets_: test_y1, keep_prob:1.0, mask_prob:0.0})
writer_tr.add_summary(tr, e)
writer_te.add_summary(te, e)
print(e,"Train cost:",tr_cost,"Test cost",te_cost)
if e%20 == 0 and e!=0:
saver.save(sess, model_path+'my_model',global_step=e, write_meta_graph=False)