MTCNN训练代码的相关问题

最新推荐文章于 2023-08-05 15:49:23 发布

原创

最新推荐文章于 2023-08-05 15:49:23 发布 · 2k 阅读

CC 4.0 BY-SA版权

文章标签：

本文记录了在训练MTCNN过程中遇到的错误和解决方案，包括数据下载、目录结构、Python环境问题以及训练脚本的错误处理。通过解决显卡显存、模块导入、版本兼容、CUDA配置等问题，逐步进行数据准备和模型训练。

最近在整MTCNN相关的东西，于是从github大佬处clone了一份，无奈本人手残，跑起来处处是报错。帖子上的询问均是石沉大海，只有边搞边记录了，希望能坚持到跑通的那一天吧= =。

Download Wider Face Training part only from Official Website , unzip to replace WIDER_train and put it into prepare_data folder.
Download landmark training data from [here]((http://mmlab.ie.cuhk.edu.hk/archive/CNN_FacePoint.htm ))，unzip and put them into prepare_data folder.
Run prepare_data/gen_12net_data.py to generate training data(Face Detection Part) for PNet.
Run gen_landmark_aug_12.py to generate training data(Face Landmark Detection Part) for PNet.
Run gen_imglist_pnet.py to merge two parts of training data.
Run gen_PNet_tfrecords.py to generate tfrecord for PNet.
After training PNet, run gen_hard_example to generate training data(Face Detection Part) for RNet.
Run gen_landmark_aug_24.py to generate training data(Face Landmark Detection Part) for RNet.
Run gen_imglist_rnet.py to merge two parts of training data.
Run gen_RNet_tfrecords.py to generate tfrecords for RNet.(you should run this script four times to generate tfrecords of neg,pos,part and landmark respectively)
After training RNet, run gen_hard_example to generate training data(Face Detection Part) for ONet.
Run gen_landmark_aug_48.py to generate training data(Face Landmark Detection Part) for ONet.
Run gen_imglist_onet.py to merge two parts of training data.

我们根据文章先去下有关数据集(官网就有，觉得慢就用迅雷之类的下，注意这里要下载的文件数量)，然后放到对应文件夹位置(prepare_data)。

有需要可以去下阉割版的：https://pan.baidu.com/s/1mf0hM5VqtpdMfTpf2VRELA 提取码：kuqw (感谢网上大佬的资源)(不过我还是推荐下原版文件，虽然很大)

在遇见具体的代码问题之前，有几个常规报错需要了解：

这里是显卡的显存的问题，手动调整显卡的占用率：

import tensorflow as tf
gpu_options = tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=0.333)
sess = tf.compat.v1.S