官方代码 Deeplab v3+ resnet101 做backbone

喜欢ヅ旅行 2023-07-02 03:25 11阅读 0赞

大年初一我居然在更博客。今年过年由于病毒横行，没有串门没有聚餐，整个人闲的没事干。。。医生真是不容易，忙得团团转还有生命危险，新希望他们平安。

**本篇不属于初级教程。如果完全看不懂请自行谷歌或搜索作者博客。**

deeplab官方提供了多种backbone，通过train.py中传递参数，

--model_variant="resnet_v1_101_beta" \

可以更改backbone。（resnet\_v1\_\{50,101\}\_beta: We modify the original ResNet-101 \[10\], similar to PSPNet \[11\] by replacing the first 7x7 convolution with three 3x3 convolutions. See resnet\_v1\_beta.py for more details.）

这个是官方对beta的说法。就是改了一个卷积核的尺寸。改小了，并且用了三个。

一共有如下可选：

# A map from feature extractor name to the network name scope used in the
    # ImageNet pretrained versions of these models.
    name_scope = {
        'mobilenet_v2': 'MobilenetV2',
        'resnet_v1_50': 'resnet_v1_50',
        'resnet_v1_50_beta': 'resnet_v1_50',
        'resnet_v1_101': 'resnet_v1_101',
        'resnet_v1_101_beta': 'resnet_v1_101',
        'xception_41': 'xception_41',
        'xception_65': 'xception_65',
        'xception_71': 'xception_71',
        'nas_pnasnet': 'pnasnet',
        'nas_hnasnet': 'hnasnet',
    }

当然backbone更改后，网络训练的参数比如decay什么的也有所不同。在feature\_extractor.py中307行开始就是在改参数，举个例子：

if 'resnet' in model_variant:
        arg_scope = arg_scopes_map[model_variant](
            weight_decay=weight_decay,
            batch_norm_decay=0.95,
            batch_norm_epsilon=1e-5,
            batch_norm_scale=True)

并且在train.py中：

# For weight_decay, use 0.00004 for MobileNet-V2 or Xcpetion model variants.
    # Use 0.0001 for ResNet model variants.
    flags.DEFINE_float('weight_decay', 0.00004,
                       'The value of the weight decay for training.')

所以，在我自己的bash文件中，我也要改

--weight_decay=0.0001 \

于是我完整的bash文件就是：

python "${WORK_DIR}"/train.py \
      --logtostderr \
      --train_split="train" \
      --model_variant="resnet_v1_50_beta" \
      --weight_decay=0.0001 \
      --atrous_rates=6 \
      --atrous_rates=12 \
      --atrous_rates=18 \
      --output_stride=16 \
      --decoder_output_stride=4 \
      --train_crop_size="513,513" \
      --train_batch_size=12 \
      --num_clones=6 \
      --training_number_of_steps=30000 \
      --fine_tune_batch_norm=True \
      --train_logdir="${TRAIN_LOGDIR}" \
      --dataset_dir="${PASCAL_DATASET}" \
      --tf_initial_checkpoint="${RES_WEIGHT}" \
      --base_learning_rate=0.007 \

说完了命令再说一下权重，官方给了resnet101，以及resnet50在imagenet，（[https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model\_zoo.md][https_github.com_tensorflow_models_blob_master_research_deeplab_g3doc_model_zoo.md]）resnet预训练的权重。从官网下载后，加载的过程中，我发现，如果使用

\--model\_variant="resnet\_v1\_101" \\

会出现加载错误。网络结构中在bottleneck上的参数设置，与checkpoint训练的网络结构不一样。同时，resnet在论文中提及的时候，作者说自己改过了。所以，这里大年初一更博客的笔者推测，beta版本才是真正的backbone。由于谷歌上不去，不想用镜像，所以笔者使用的是beta。使用后权重加载成功，并且有如下提示：

INFO:tensorflow:Initializing model from path: /home/DATA/liutian/tmp/tfdeeplab/resnet/model.ckpt
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/weights]
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/weights]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/depthwise_weights]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/weights]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/depthwise_weights]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/weights]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/depthwise_weights]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/weights]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/weights]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/weights]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/depthwise_weights]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/weights]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/depthwise_weights]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/weights]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/BatchNorm/gamma]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/BatchNorm/beta]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/BatchNorm/moving_mean]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/BatchNorm/moving_variance]
    WARNING:tensorflow:Checkpoint is missing variable [logits/semantic/weights]
    WARNING:tensorflow:Checkpoint is missing variable [logits/semantic/biases]
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [image_pooling/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp0/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/depthwise_weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_depthwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp1_pointwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/depthwise_weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_depthwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp2_pointwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/depthwise_weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_depthwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [aspp3_pointwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [concat_projection/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/feature_projection0/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/depthwise_weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_depthwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv0_pointwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/depthwise_weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_depthwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/BatchNorm/gamma/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [decoder/decoder_conv1_pointwise/BatchNorm/beta/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [logits/semantic/weights/Momentum]
    WARNING:tensorflow:Checkpoint is missing variable [logits/semantic/biases/Momentum]
    INFO:tensorflow:Create CheckpointSaverHook.

我认为这个说明checkpoint少了decode与aspp的网络。而beta是有的。因为在代码中，aspp是否使用是通过参数空置的（model.py 397行：

model_options.aspp_with_batch_norm

），decode是否使用也是通过参数控制的（

decoder_output_stride如果不给参数就直接删除整个decoder。

），modelvariant只会影响encoder阶段，也就是feature\_extractor.

那么我的确认方式是1，debug2，看tensorboard

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70][]

ValueError: Total size of new array must be unchanged for resnet\_v1\_101/block1/unit\_1/bottleneck\_v1/shortcut/weights lh\_shape: \[(1, 1, 128, 256)\], rh\_shape: \[(1, 1, 64, 256)\]

之所以废这么多话是想说，复现可能会有一定问题，因为你需要先用coco预训练，再用voc2012 trainaug set预训练，得到的权重才可以和论文比。

我是没用coco。

更-------------------------

这个病毒越演愈烈，正好在北京直面过非典，实际上非典时期结束有一部分是因为天气变热，病毒传播困难。再加上隔离严格。所以武汉肺炎终究会过去。就是医护人员在湖北人手不足，新闻上全家感染的例子不在少数。致死率没有非典严重，大多数是并发症。但是传染的速度真的是太快了。虽然不能恐慌，但是也要严肃对待。未感染的需要提高免疫力，摄取维生素c，注意保暖，不要感冒发烧，以免给医疗系统增加压力。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 1][]

这里贴的是resnet101在voc的结果，

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 2][]

这个贴的是xception，可以看到，这里采用了coco以及JFT两个数据集预训练作为变量。而resnet则没有，所以很有可能是没用coco，只用了imagenet。

所以我的总的bash代码是这样的。

#!/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"
    RES_DIR="${CURRENT_DIR}/resnet"
    RES_WEIGHT="${CURRENT_DIR}/resnet/model.ckpt"
    #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/offres_freze"
    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"
    VOC_LOGDIR="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/${EXP_FOLDER}/train_voc"
    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}"
    
    cd "${CURRENT_DIR}"
    
    PASCAL_DATASET="${WORK_DIR}/${DATASET_DIR}/${PASCAL_FOLDER}/tfrecord"
    
    # Train 10 iterations.
    NUM_ITERATIONS=20
    echo ${PASCAL_DATASET}
    
    
    python "${WORK_DIR}"/train.py \
      --logtostderr \
      --train_split="train_aug" \
      --model_variant="resnet_v1_101_beta" \
      --weight_decay=0.001 \
      --atrous_rates=6 \
      --atrous_rates=12 \
      --atrous_rates=18 \
      --output_stride=16 \
      --decoder_output_stride=4 \
      --train_crop_size="513,513" \
      --train_batch_size=32 \
      --num_clones=8 \
      --training_number_of_steps=30000 \
      --fine_tune_batch_norm=True \
      --train_logdir="${TRAIN_LOGDIR}" \
      --dataset_dir="${PASCAL_DATASET}" \
      --tf_initial_checkpoint="${RES_WEIGHT}" \
    
    python "${WORK_DIR}"/train.py \
      --logtostderr \
      --train_split="train" \
      --model_variant="resnet_v1_101_beta" \
      --weight_decay=0.0001 \
      --atrous_rates=6 \
      --atrous_rates=12 \
      --atrous_rates=18 \
      --output_stride=16 \
      --decoder_output_stride=4 \
      --train_crop_size="513,513" \
      --train_batch_size=32 \
      --num_clones=8 \
      --training_number_of_steps=30000 \
      --fine_tune_batch_norm=True \
      --train_logdir="${VOC_LOGDIR}" \
      --dataset_dir="${PASCAL_DATASET}" \
      --tf_initial_checkpoint="${TRAIN_LOGDIR}/model.ckpt-30000" \

论文较真：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 3][]

由于官方没写resnet作为backbone怎么训练的，所以这里放一个同样用deeplabv3+，resnet，cvpr的文章：

Decoders Matter for Semantic Segmentation:Data-Dependent Decoding Enables Flexible Feature Aggregation

在val的pascal voc数据集大概是72,73的准确率：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 4][]

[https_github.com_tensorflow_models_blob_master_research_deeplab_g3doc_model_zoo.md]: https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20200125171201487.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20200203185858360.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/2020020318595943.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20200220202048867.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM_size_16_color_FFFFFF_t_70 4]: https://img-blog.csdnimg.cn/20200220202307573.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTMyNDk4NTM=,size_16,color_FFFFFF,t_70