deeplabv3+ 从零开始训练

Dear 丶 2024-04-17 15:35 61阅读 0赞

首先需要下载官方提供的imagenet以及coco的预训练权重：

[http://download.tensorflow.org/models/xception\_65\_coco\_pretrained\_2018\_10\_02.tar.gz][http_download.tensorflow.org_models_xception_65_coco_pretrained_2018_10_02.tar.gz]

然后将其解压到一个文件夹比如：

.../deeplab/datasets/pascal\_voc\_seg/coco\_pretrain

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70][]

train\_aug的处理：[https://blog.csdn.net/u013249853/article/details/100136780][https_blog.csdn.net_u013249853_article_details_100136780]，只需要下载一分钟，仅下载标注，一定要看清楚。

然后我直接放我的训练配置，注意，python代码是一句没改的。一定要将train\_split改成train。否则会将val也一并加入训练。比如下面的保存成train.sh

#!/bin/bash
    # Copyright 2018 The TensorFlow Authors All Rights Reserved.
    #
    # Licensed under the Apache License, Version 2.0 (the "License");
    # you may not use this file except in compliance with the License.
    # You may obtain a copy of the License at
    #
    #     http://www.apache.org/licenses/LICENSE-2.0
    #
    # Unless required by applicable law or agreed to in writing, software
    # distributed under the License is distributed on an "AS IS" BASIS,
    # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    # See the License for the specific language governing permissions and
    # limitations under the License.
    # ==============================================================================
    #
    # This script is used to run local test on PASCAL VOC 2012. Users could also
    # modify from this script for their use case.
    #
    # Usage:
    #   # From the tensorflow/models/research/deeplab directory.
    #   sh ./local_test.sh
    #
    #
    
    # Exit immediately if a command exits with a non-zero status.
    set -e
    
    # Move one-level up to tensorflow/models/research directory.
    cd ..
    
    # Update PYTHONPATH.
    export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
    
    # Set up the working environment.
    CURRENT_DIR=$(pwd)
    WORK_DIR="${CURRENT_DIR}/deeplab"
    #export CUDA_VISIBLE_DEVICES=3
    # Run model_test first to make sure the PYTHONPATH is correctly set.
    #python "${WORK_DIR}"/model_test.py -v
    
    # Go to datasets folder and download PASCAL VOC 2012 segmentation dataset.
    DATASET_DIR="datasets"
    cd "${WORK_DIR}/${DATASET_DIR}"
    #sh download_and_convert_voc2012.sh
    
    # Go back to original directory.
    cd "${CURRENT_DIR}"
    
    # Set up the working directories.
    PASCAL_FOLDER="pascal_voc_seg"
    EXP_FOLDER="/result/coco"
    INIT_FOLDER="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/init_models"
    COCO_PRE="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/coco_pretrain"
    TRAIN_LOGDIR="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/${EXP_FOLDER}/train"
    EVAL_LOGDIR="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/${EXP_FOLDER}/eval"
    VIS_LOGDIR="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/${EXP_FOLDER}/vis"
    EXPORT_DIR="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/${EXP_FOLDER}/export"
    mkdir -p "${INIT_FOLDER}"
    mkdir -p "${TRAIN_LOGDIR}"
    mkdir -p "${EVAL_LOGDIR}"
    mkdir -p "${VIS_LOGDIR}"
    mkdir -p "${EXPORT_DIR}"
    
    # Copy locally the trained checkpoint as the initial checkpoint.
    TF_INIT_ROOT="http://download.tensorflow.org/models"
    TF_INIT_CKPT="deeplabv3_pascal_train_aug_2018_01_04.tar.gz"
    cd "${INIT_FOLDER}"
    #wget -nd -c "${TF_INIT_ROOT}/${TF_INIT_CKPT}"
    tar -xf "${TF_INIT_CKPT}"
    cd "${CURRENT_DIR}"
    
    PASCAL_DATASET="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/tfrecord"
    
    # Train 10 iterations.
    NUM_ITERATIONS=20
    echo ${PASCAL_DATASET}
    
    python "${WORK_DIR}"/train.py \
      --logtostderr \
      --train_split="train_aug" \
      --model_variant="xception_65" \
      --atrous_rates=6 \
      --atrous_rates=12 \
      --atrous_rates=18 \
      --output_stride=16 \
      --decoder_output_stride=4 \
      --train_crop_size="513,513" \
      --train_batch_size=24 \
      --base_learning_rate=0.007 \
      --training_number_of_steps=30000 \
      --fine_tune_batch_norm=true \
      --num_clones=8 \
      --tf_initial_checkpoint="${COCO_PRE}/x65-b2u1s2p-d48-2-3x256-sc-cr300k_init.ckpt" \
      --train_logdir="${TRAIN_LOGDIR}" \
      --dataset_dir="${PASCAL_DATASET}"\
      --initialize_last_layer=true
    
    python "${WORK_DIR}"/train.py \
      --logtostderr \
      --train_split="train" \
      --model_variant="xception_65" \
      --atrous_rates=6 \
      --atrous_rates=12 \
      --atrous_rates=18 \
      --output_stride=16 \
      --decoder_output_stride=4 \
      --train_crop_size="513,513" \
      --train_batch_size=24 \
      --training_number_of_steps=60000 \
      --fine_tune_batch_norm=false \
      --num_clones=8 \
      --base_learning_rate=0.01 \
      --tf_initial_checkpoint="${COCO_PRE}/x65-b2u1s2p-d48-2-3x256-sc-cr300k_init.ckpt" \
      --train_logdir="${TRAIN_LOGDIR}" \
      --dataset_dir="${PASCAL_DATASET}"\
      --initialize_last_layer=true
    
      python "${WORK_DIR}"/eval.py \
      --logtostderr \
      --eval_split="val" \
      --model_variant="xception_65" \
      --atrous_rates=6 \
      --atrous_rates=12 \
      --atrous_rates=18 \
      --output_stride=16 \
      --decoder_output_stride=4 \
      --eval_crop_size="513,513" \
      --checkpoint_dir="${TRAIN_LOGDIR}" \
      --eval_logdir="${EVAL_LOGDIR}" \
      --dataset_dir="${PASCAL_DATASET}" \
      --max_number_of_evaluations=1
    #echo ${PASCAL_DATASET}
    # Run evaluation. This performs eval over the full val split (1449 images) and
    # will take a while.
    # Using the provided checkpoint, one should expect mIOU=82.20%.
    #--tf_initial_checkpoint="${INIT_FOLDER}/deeplabv3_pascal_train_aug/model.ckpt" \

那么参数其实没怎么改，主要是training\_batch\_size, fine\_tune,checkpoint\_dir(通过更改EXP\_FOLDER).

先通过train\_aug,以及比较大（至少大于12，我用的24）的batchsize得到优秀的BN参数（fine\_rune\_batch\_norm=true）。之后即可通过将BN参数固定（fine\_rune\_batch\_norm=false），再只使用train的情况下进行训练，然后就能复现了。

真的是奇怪，为什么train\_aug不能提供更好的结果呢，我看train\_aug图片还挺准确的啊。奇怪。train\_aug大多数图片没有白边的确是。也许没有白边，所以背景会有一部分被认为是前景。

另外也没有对于weight和loss的bias处理，毕竟背景更多不是么。

启动的命令是 bash train.sh

参数介绍看[https://blog.csdn.net/u013249853/article/details/100084147][https_blog.csdn.net_u013249853_article_details_100084147]后半部分

参考：[https://github.com/tensorflow/models/issues/5116][https_github.com_tensorflow_models_issues_5116]

[http_download.tensorflow.org_models_xception_65_coco_pretrained_2018_10_02.tar.gz]: http://download.tensorflow.org/models/xception_65_coco_pretrained_2018_10_02.tar.gz
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70]: https://image.dandelioncloud.cn/pgy_files/images/2024/04/17/83e93a30b6cd47e1bf8c7b46411de5b8.png
[https_blog.csdn.net_u013249853_article_details_100136780]: https://blog.csdn.net/u013249853/article/details/100136780
[https_blog.csdn.net_u013249853_article_details_100084147]: https://blog.csdn.net/u013249853/article/details/100084147
[https_github.com_tensorflow_models_issues_5116]: https://github.com/tensorflow/models/issues/5116