深度学习训练自己的模型

1.数据集标注

ssd训练自己的模型

参考https://blog.csdn.net/u014696921/article/details/53353896

2.用别的模型进行微调，并根据自己的数据类别调整参数

   如果仅仅调整程序参数，这时调用预训练模型是会出错的（从新开始训练不会报错），这是因为预训练模型的类别与调整后的c类别不一样，导致某些层输出张量维度不一样，因此出错，修正方法有两种;

方法一:

不加载这些层的参数：

 tf.app.flags.DEFINE_string(
     'checkpoint_exclude_scopes', 'ssd_300_vgg/block11_box/conv_cls/biases,ssd_300_vgg/block11_box/conv_cls/weights,'
                                  'ssd_300_vgg/block10_box/conv_cls/biases,ssd_300_vgg/block10_box/conv_cls/weights,'
                                  'ssd_300_vgg/block9_box/conv_cls/biases,ssd_300_vgg/block9_box/conv_cls/weights,'
                                  'ssd_300_vgg/block8_box/conv_cls/biases,ssd_300_vgg/block8_box/conv_cls/weights,'
                                  'ssd_300_vgg/block7_box/conv_cls/biases,ssd_300_vgg/block7_box/conv_cls/weights,'
                                  'ssd_300_vgg/block6_box/conv_cls/biases,ssd_300_vgg/block6_box/conv_cls/weights,'
                                  'ssd_300_vgg/block5_box/conv_cls/biases,ssd_300_vgg/block5_box/conv_cls/weights,'
                                  'ssd_300_vgg/block4_box/conv_cls/biases,ssd_300_vgg/block4_box/conv_cls/weights,'
                                  'ssd_300_vgg/block3_box/conv_cls/biases,ssd_300_vgg/block3_box/conv_cls/weights,'
                                  'ssd_300_vgg/block2_box/conv_cls/biases,ssd_300_vgg/block2_box/conv_cls/weights,'
                                  'ssd_300_vgg/block1_box/conv_cls/biases,ssd_300_vgg/block1_box/conv_cls/weights',
     'Comma-separated list of scopes of variables to exclude when restoring '
     'from a checkpoint.')

方法二：

修改模型中的参数使其张量维数保持一致：

import os
import tensorflow as tf
from tensorflow.python import pywrap_tensorflow
def readcheckpoint(model_dir="../checkpoints/ssd_300_vgg.ckpt"):
    # model_dir="../checkpoints/ssd_300_vgg.ckpt" #checkpoint的文件位置
    # Read data from checkpoint file
    reader = pywrap_tensorflow.NewCheckpointReader(model_dir)
    var_to_shape_map = reader.get_variable_to_shape_map()
    # Print tensor name and values
    for key in var_to_shape_map:
        print("tensor_name: ", key)  #输出变量名
        # print(reader.get_tensor(key))   #输出变量值
        print(reader.get_tensor(key).shape)

def savecheckpoint():
    ckpt_path="../checkpoints/ssd_300_vgg.ckpt"
    with tf.Session() as sess:
        for var_name, _ in tf.contrib.framework.list_variables(ckpt_path):
            # Load the variable
            var = tf.contrib.framework.load_variable(ckpt_path, var_name)

            # Set the new name
            new_name = var_name


            print('Renaming %s to %s.' % (var_name, new_name))
            # Rename the variable
            # print(var)
            if new_name.__contains__('_box/conv_cls/biases'):
                if new_name.__contains__('block7_box/conv_cls/biases') or new_name.__contains__('block8_box/conv_cls/biases') or new_name.__contains__('block9_box/conv_cls/biases'):
                    var=var[0:15*6]
                else:
                    var = var[0:15 * 4]
            if new_name.__contains__('conv_cls/weights'):
                if new_name.__contains__('block7_box/conv_cls/weights') or new_name.__contains__('block8_box/conv_cls/weights') or new_name.__contains__('block9_box/conv_cls/weights'):
                    var=var[:,:,:,0:15*6]
                else:
                    var = var[:, :, :, 0:15 * 4]
            var = tf.Variable(var, name=new_name)

        # Save the variables
        saver = tf.train.Saver()
        sess.run(tf.global_variables_initializer())
        saver.save(sess, './test.ckpt')
savecheckpoint()
# readcheckpoint()