tensorflow、keras使用记录

1. 网络可视化

writer = tf.summary.FileWriter('logs/',sess.graph)

tensorboard --logdir=logs

2. keras进行训练

keras fit 和 fit_generator的区别，以及如何使用fit_generator进行训练：

https://blog.csdn.net/learning_tortosie/article/details/85243310​

已测试，可用；下面是保存模型的方法

best_weights_filepath = 'a.h5'

saveBestModel = ModelCheckpoint(best_weights_filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='auto')

# train the network

print("[INFO] training w/ generator...")

H = model.fit_generator(    trainGen,    steps_per_epoch=NUM_TRAIN_IMAGES // BS,    validation_data=testGen,    validation_steps=NUM_TEST_IMAGES // BS,    epochs=NUM_EPOCHS,    callbacks=[saveBestModel])​

或者采用：

model.save_weights('a.h5')

hdf5与h5的区别在于hdf5只保存weights，而h5会同时保存weights和structure。​

3. 基于代码层的maskrcnn解释 https://www.cnblogs.com/wangyong/p/10614898.html​

4. error:module keras.backend has no attribute control_flow_ops ​

在使用YOLO v3训练自己单目标检测的数据集时.

出现了backend’ has no attribute ‘control_flow_ops’错误，

which python (找到python的路径)

找到lib/python3.7/site-packages/keras​/backend/__init__.py

在底部添加两项(需要在tensorflow的分支底下增加):

    from .load_backend import control_flow_ops

     from .load_backend import set_image_dim_ordering

​5. tensorflow GPU使用说明

watch nvidia-smi 备注：实时返回GPU使用情况​

1）指定GPU训练：

方法一、在python程序中设置：

代码：os.environ[‘CUDA_VISIBLE_DEVICES’] = ‘0’ 备注：使用 GPU 0

代码：os.environ[‘CUDA_VISIBLE_DEVICES’] = ‘0,1’ 备注：使用 GPU 0，1

方法二、在执行python程序时候：

指令：CUDA_VISIBLE_DEVICES=2 python yourcode.py

指令：CUDA_VISIBLE_DEVICES=0,1 python yourcode.py

备注：‘=’的左右不允许有空格

注：**TensorFlow会默认直接占满我们模型部署的GPU的存储资源，只允许一个小内存的程序也会占用所有GPU资源。因此有的时候我们通过nvidia-smi查看GPU状态的时候，会发现有些GPU的计算利用率很低或者计算利用率为0，但是存储被占满了，而这个时候其他人也不能使用这块GPU。但是现在公司的问题是模型多，卡不够用，所有只能“文明”使用GPU，如果设置为允许动态增长的话，这样这个GPU没有被占的存储还可以被其他人使用。​

2）​两种限定GPU占用量的方法：

方法一、设置定量的GPU显存使用量:

config = tf.ConfigProto()

config.gpu_options.per_process_gpu_memory_fraction = 0.4 # 占用GPU40%的显存

session = tf.Session(config=config)

方法二、设置最小的GPU显存使用量，动态申请显存:（建议）
 

config = tf.ConfigProto()

config.gpu_options.allow_growth = True

session = tf.Session(config=config)

6. 打印pb的layer名字以及保存网络模型

import tensorflow as tf
import sys
from tensorflow.python.framework import graph_util
tf.reset_default_graph()
output_graph_path =  sys.argv[1]
with tf.Session() as sess:
 
    tf.global_variables_initializer().run()
    output_graph_def = tf.GraphDef()

    graph = tf.get_default_graph()
    with open(output_graph_path, "rb") as f:
        output_graph_def.ParseFromString(f.read())
        _ = tf.import_graph_def(output_graph_def, name="")

        print("%d ops in the final graph." % len(output_graph_def.node))
 
        tensor_name = [tensor.name for tensor in output_graph_def.node]
        # print every layer
        for tensor_name_1 in tensor_name:
        	print(tensor_name_1)
        print('---------------------------')
        # save graph for tensorboard to view the graph
        summaryWriter = tf.summary.FileWriter(sys.argv[2], graph) 
 
        for op in graph.get_operations():
            print(op.name, op.values())

7. 读取pb，并打印模型中间结果

detection_graph = tf.Graph()
with detection_graph.as_default():
  od_graph_def = tf.GraphDef()
  with tf.gfile.GFile(PATH_TO_FROZEN_GRAPH, 'rb') as fid:
    serialized_graph = fid.read()
    od_graph_def.ParseFromString(serialized_graph)
    tf.import_graph_def(od_graph_def, name='')
    print('pb import sucess')
with graph.as_default():
    # with tf.Session() as sess:
    sess = tf.Session()
    # Get handles to input and output tensors
    ops = tf.get_default_graph().get_operations()
    all_tensor_names = {output.name for op in ops for output in op.outputs}
    tensor_dict = {}
    for key in [
        'Preprocessor/map/TensorArrayStack_1/TensorArrayGatherV3', 
        'detection_classes', 'detection_masks'
    ]:
    tensor_name = key + ':0'
    if tensor_name in all_tensor_names:
        tensor_dict[key] = tf.get_default_graph().get_tensor_by_name(
        tensor_name)
     image_tensor = tf.get_default_graph().get_tensor_by_name('image_tensor:0')

    output_dict = sess.run(tensor_dict,
                             feed_dict={image_tensor: image})

      # all outputs are float32 numpy arrays, so convert types as appropriate

      output_dict['Preprocessor/map/TensorArrayStack_1/TensorArrayGatherV3'] = output_dict['Preprocessor/map/TensorArrayStack_1/TensorArrayGatherV3']
      output_dict['detection_scores'] = output_dict['detection_scores'][0]
      if 'detection_masks' in output_dict:
        output_dict['detection_masks'] = output_dict['detection_masks'][0]

8. 数据全量打印

需要使用numpy并设置一下内容

import numpy as np
np.set_printoptions(threshold=np.inf)

之后再使用print打印就会将数组中所有的内容都打印出来。

9. 关于tf.image.resize_images与opencv的resize不一致的问题：

提出一个解决方案（此处针对的是bilinear方式的resize函数）：

void Resize_Image(cv::Mat& src, cv::Mat& dst, int target_height, int target_width, int channel)
{
    auto clip = [](float in, float low, float high) -> float { return (in < low) ? low : (in > high ? high : in); };
    int original_height = src.rows;
    int original_width = src.cols;
    assert(dst.rows == target_height);
    assert(dst.cols == target_width);
    float ratio_h = static_cast<float>(original_height - 1.0f) / static_cast<float>(target_height - 1.0f);
    float ratio_w = static_cast<float>(original_width - 1.0f) / static_cast<float>(target_width - 1.0f);

    for (int y = 0; y < target_height; ++y){
        for (int x = 0; x < target_width; ++x){
            float x0 = static_cast<float>(x) * ratio_w;
            float y0 = static_cast<float>(y) * ratio_h;
            int left = static_cast<int>(clip(std::floor(x0), 0.0f, static_cast<float>(original_width - 1.0f)));
            int top = static_cast<int>(clip(std::floor(y0), 0.0f, static_cast<float>(original_height - 1.0f)));
            int right = static_cast<int>(clip(std::ceil(x0), 0.0f, static_cast<float>(original_width - 1.0f)));
            int bottom = static_cast<int>(clip(std::ceil(y0), 0.0f, static_cast<float>(original_height - 1.0f)));

            for (int c = 0; c < channel; ++c){
                // H, W, C ordering
                float left_top_val = src.at<cv::Vec3f>(top, left)[c];
                float right_top_val = src.at<cv::Vec3f>(top, right)[c];
                float left_bottom_val = src.at<cv::Vec3f>(bottom, left)[c];
                float right_bottom_val = src.at<cv::Vec3f>(bottom, right)[c];
                float top_lerp = left_top_val + (right_top_val - left_top_val) * (x0 - left);
                float bottom_lerp = left_bottom_val + (right_bottom_val - left_bottom_val) * (x0 - left);
                float lerp = clip(std::round(top_lerp + (bottom_lerp - top_lerp) * (y0 - top)), 0.0f, 255.0f);
                dst.at<cv::Vec3f>(y,x)[c] = lerp;
            }
        }
    }
}

10. 更换pip install 源

pip install xxxx -i https://pypi.tuna.tsinghua.edu.cn/simple

11. 修改pb的某一层value

将pb转化为pbtxt，然后用文本打开pbtxt文件，手动修改某一层layer的值然后使用reverse，进行转回pb文件

import numpy as np
import tensorflow as tf
import argparse


def convert(pb_path, pbtxt_path):
    with open(pb_path, 'rb') as f:
        input_graph_def = tf.GraphDef()
        input_graph_def.ParseFromString(f.read())

    with tf.Graph().as_default() as graph:
        tf.import_graph_def(input_graph_def, name='')
        with tf.Session(graph=graph) as sess:
            tf.train.write_graph(graph, 'PBs', pbtxt_path, as_text=True)


def convert_back(pbtxt_path, pb_path):
    from google.protobuf import text_format as pbtf
    input_graph_def = tf.GraphDef()
    with open(pbtxt_path, 'r') as f:
        graph_str = f.read()
    pbtf.Parse(graph_str, input_graph_def)

    with tf.Graph().as_default() as graph:
        tf.import_graph_def(input_graph_def, name='')
        with tf.Session(graph=graph) as sess:
            tf.train.write_graph(graph, 'PBs', pb_path, as_text=False)


if __name__ == '__main__':
    """Parse input arguments."""
    parser = argparse.ArgumentParser(description='compare_graphs')

    parser.add_argument('--pb', dest='pb_path', help='input pb file path')
    parser.add_argument('--pbtxt', dest='pbtxt_path', help='output pbtxt file path')
    parser.add_argument('--reverse', action='store_true', help='reverse convert')
    args = parser.parse_args()

    if args.pb_path is None:
        print("empty input arguments")
    elif args.reverse:
        convert_back(args.pbtxt_path, args.pb_path)
    else:
        convert(args.pb_path, args.pbtxt_path)