Improve training efficiency by integrating different sess.run()

[JingyunLiang]

Bilinear-CNN-TensorFlow/core/bcnn_finetuning.py

Line 391 in cc2df09

cost = sess.run(loss, feed_dict={imgs: batch_xs, target: batch_ys})

Original implementation seems to propagate for two(three) times, leading to large amount of time waste. Use
sess.run([optimizer,check_op,loss,accuracy], feed_dict={imgs: batch_xs, target: batch_ys})
It saves nearly half of training time.