pyhon源码阅读积累(持续更新)

1. sess.run() 的时候还可以[]组合到一起，一组写在一行里面。

    rpn_box_loss, rpn_cls_loss, refined_box_loss, refined_cls_loss, mask_loss, \
            gt_boxesnp, \
            rpn_batch_pos, rpn_batch, refine_batch_pos, refine_batch, mask_batch_pos, mask_batch = \
                         sess.run([update_op, total_loss, reg_loss,  img_id] + 
                                  losses + 
                                  [gt_boxes] + 
                                  batch_info)

2. 每1000轮迭代或者最大次数迭代时，保存ckpt

if (step % 1000 == 0 or step + 1 == FLAGS.max_iters) and step != 0:
	saver.save(sess,path,glabal_step = step)

3. 训练框架

saver = tf.train.Saver(max_to_keep=20)  #保存20个ckpt文件（默认是5）
config = tf.ConfigProto()    
config.gpu_options.allow_growth = True

with tf.Session(config = config) as sess:
    sess.run(tf.global_variables_initializer())
    sess.run(tf.local_variables_initializer())

    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord = coord)
    step = 0
    summary_writer = tf.summary.FileWriter('output', sess.graph)
    while (True):
        step += 1
        try:
            #这里写主程序
            
            if step % 100 == 0:
                summary_writer.add_summary(summary_, step)
                summary_writer.flush()
                save_path = saver.save(sess, os.path.join(design.saver_folder, design.path_ckpt), global_step=step)
                    
                if step % 10000 == 0:
                    coord.request_stop()
                    coord.join(threads)

        except tf.errors.OutOfRangeError:
            break

    # Finish off the fiqlename queue coordinator.
    coord.request_stop()
    coord.join(threads)

4.条件判断函数tf.cond()，满足条件执行第一个表达式，不满足条件执行第二个表达式

imsize = tf.size(image)
image = tf.cond(tf.equal(imsize, ih * iw),  # tf.cond()条件表达式，满足条件执行第一个表达式，否则执行第二个lambda表达式
               lambda: tf.image.grayscale_to_rgb(tf.reshape(image, (ih, iw, 1))),
               lambda: tf.reshape(image, (ih, iw, 3)))

5.找label的操作（已知label的shape为（N，））

keeps = tf.where(tf.greater_equal(labels, 0))   #tf.where()函数返回矩阵中true的位置
keeps = tf.reshape(keeps, [-1])  #输出正label的索引（过滤掉了值为负的labels，只输出了值为正的labels）

顺便提一下tf.where()的另一个操作：将bool型的,True OR False值转换成了数字0.0或1.0，加起来求了个和～
res.append(tf.reduce_sum(tf.cast(tf.greater_equal(labels, 0), tf.float32)))

6.拆分bbox坐标：[boxes[:,i] for i in range(5)]

boxes = np.array([[10,10,40,80,0.5],[30,25,65,78,0.8],[25,30,70,80,0.9]])
x1,y1,x2,y2,score = [boxes[:,i] for i in range(5)]
print(x1,y1,x2,y2,score)
#输出：[10. 30. 25.] [10. 25. 30.] [40. 65. 70.] [80. 78. 80.] [0.5 0.8 0.9]

7.python手动垃圾回收：gc.collect()命令。

8.将多个图像拼接成一个图像的方法：

from skimage.util.montage import montage2d as montage
montage_rgb = lambda x: np.stack([montage(x[:, :, :, i]) for i in range(x.shape[3])], -1)