- 安装教程：https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/installation.md
- cityscapes训练：https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/cityscapes.md
遇到的坑：
1. 环境：
– tensorflow1.8+CUDA9.0+cudnn7.0+annaconda3+py3.5
– 使用最新的tensorflow1.12或者1.10都不行，报错：报错不造卷积算法（convolution algorithm…）
2. 数据集转换
# Exit immediately if a command exits with a non-zero status. set -e CURRENT_DIR=$(pwd) WORK_DIR="." # Root path for Cityscapes dataset. CITYSCAPES_ROOT="${WORK_DIR}/cityscapes" # Create training labels. python "${CITYSCAPES_ROOT}/cityscapesscripts/preparation/createTrainIdLabelImgs.py" # Build TFRecords of the dataset. # First, create output directory for storing TFRecords. OUTPUT_DIR="${CITYSCAPES_ROOT}/tfrecord" mkdir -p "${OUTPUT_DIR}" BUILD_SCRIPT="${CURRENT_DIR}/build_cityscapes_data.py" echo "Converting Cityscapes dataset..." python "${BUILD_SCRIPT}" \ --cityscapes_root="${CITYSCAPES_ROOT}" \ --output_dir="${OUTPUT_DIR}" \
– 首先当前conda环境下安装cityscapesScripts模块，要支持py3.5才行；
– 由于cityscapesscripts/preparation/createTrainIdLabelImgs.py里面默认会把数据集gtFine下面的test,train,val文件夹json文件都转为TrainIdlandelImgs.png;然而在test文件下有很多json文件编码格式是错误的，大约十几张，每次报错，然后将其剔除！！！
– 然后执行build_cityscapes_data.py将img,lable转换为tfrecord格式。
3. 训练cityscapes代码
– 将训练代码写成脚本文件：train_deeplab_cityscapes.sh
#!/bin/bash # CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --backbone resnet --lr 0.01 --workers 4 --epochs 40 --batch-size 16 --gpu-ids 0,1,2,3 --checkname deeplab-resnet --eval-interval 1 --dataset coco PATH_TO_INITIAL_CHECKPOINT='/home/rjw/tf-models/research/deeplab/pretrain_models/deeplabv3_cityscapes_train/model.ckpt' PATH_TO_TRAIN_DIR='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/train/' PATH_TO_DATASET='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/tfrecord' WORK_DIR='/home/rjw/tf-models/research/deeplab' # From tensorflow/models/research/ python "${WORK_DIR}"/train.py \ --logtostderr \ --training_number_of_steps=40000 \ --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 \ --train_crop_size=513 \ --train_batch_size=1 \ --fine_tune_batch_norm=False \ --dataset="cityscapes" \ --tf_initial_checkpoint=${PATH_TO_INITIAL_CHECKPOINT} \ --train_logdir=${PATH_TO_TRAIN_DIR} \ --dataset_dir=${PATH_TO_DATASET}
参数分析：
training_number_of_steps： 训练迭代次数；
train_crop_size：训练图片的裁剪大小，因为我的GPU只有8G，故我将这个设置为513了；
train_batch_size: 训练的batchsize，也是因为硬件条件，故保持1；
fine_tune_batch_norm=False ：是否使用batch_norm，官方建议，如果训练的batch_size小于12的话，须将该参数设置为False，这个设置很重要，否则的话训练时会在2000步左右报错
tf_initial_checkpoint：预训练的初始checkpoint，这里设置的即是前面下载的../research/deeplab/backbone/deeplabv3_cityscapes_train/model.ckpt.index
train_logdir: 保存训练权重的目录，注意在开始的创建工程目录的时候就创建了，这里设置为”../research/deeplab/exp/train_on_train_set/train/”
dataset_dir:数据集的地址，前面创建的TFRecords目录。这里设置为”../dataset/cityscapes/tfrecord”
4.验证测试
– 验证脚本：
#!/bin/bash # CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --backbone resnet --lr 0.01 --workers 4 --epochs 40 --batch-size 16 --gpu-ids 0,1,2,3 --checkname deeplab-resnet --eval-interval 1 --dataset coco PATH_TO_INITIAL_CHECKPOINT='/home/rjw/tf-models/research/deeplab/pretrain_models/deeplabv3_cityscapes_train/' PATH_TO_CHECKPOINT='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/train/' PATH_TO_EVAL_DIR='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/eval/' PATH_TO_DATASET='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/tfrecord' WORK_DIR='/home/rjw/tf-models/research/deeplab' # From tensorflow/models/research/ python "${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=1025 \ --eval_crop_size=2049 \ --dataset="cityscapes" \ --checkpoint_dir=${PATH_TO_INITIAL_CHECKPOINT} \ --eval_logdir=${PATH_TO_EVAL_DIR} \ --dataset_dir=${PATH_TO_DATASET}
– rusult:model.ckpt-40000为在初始化模型上训练40000次迭代的模型；后面用初始化模型测试miou_1.0还是很低，不知道是不是有什么参数设置的问题！！！
– 注意，如果使用官方提供的checkpoint，压缩包中是没有checkpoint文件的，需要手动添加一个checkpoint文件；初始化模型中是没有提供chekpoint文件的。
INFO:tensorflow:Restoring parameters from /home/rjw/tf-models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/train/model.ckpt-40000 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Starting evaluation at 2018-12-18-07:13:08 INFO:tensorflow:Evaluation [50/500] INFO:tensorflow:Evaluation [100/500] INFO:tensorflow:Evaluation [150/500] INFO:tensorflow:Evaluation [200/500] INFO:tensorflow:Evaluation [250/500] INFO:tensorflow:Evaluation [300/500] INFO:tensorflow:Evaluation [350/500] INFO:tensorflow:Evaluation [400/500] INFO:tensorflow:Evaluation [450/500] miou_1.0[0.478293568] INFO:tensorflow:Waiting for new checkpoint at /home/rjw/tf-models/research/deeplab/pretrain_models/deeplabv3_cityscapes_train/ INFO:tensorflow:Found new checkpoint at /home/rjw/tf-models/research/deeplab/pretrain_models/deeplabv3_cityscapes_train/model.ckpt INFO:tensorflow:Graph was finalized. 2018-12-18 15:18:05.210957: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1435] Adding visible gpu devices: 0 2018-12-18 15:18:05.211047: I tensorflow/core/common_runtime/gpu/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix: 2018-12-18 15:18:05.211077: I tensorflow/core/common_runtime/gpu/gpu_device.cc:929] 0 2018-12-18 15:18:05.211100: I tensorflow/core/common_runtime/gpu/gpu_device.cc:942] 0: N 2018-12-18 15:18:05.211645: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 9404 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1) INFO:tensorflow:Restoring parameters from /home/rjw/tf-models/research/deeplab/pretrain_models/deeplabv3_cityscapes_train/model.ckpt INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Starting evaluation at 2018-12-18-07:18:06 INFO:tensorflow:Evaluation [50/500] INFO:tensorflow:Evaluation [100/500] INFO:tensorflow:Evaluation [150/500] INFO:tensorflow:Evaluation [200/500] INFO:tensorflow:Evaluation [250/500] INFO:tensorflow:Evaluation [300/500] INFO:tensorflow:Evaluation [350/500] INFO:tensorflow:Evaluation [400/500] INFO:tensorflow:Evaluation [450/500] miou_1.0[0.496331513]
5.可视化测试
– 在vis目录下生成分割结果图
#!/bin/bash # CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --backbone resnet --lr 0.01 --workers 4 --epochs 40 --batch-size 16 --gpu-ids 0,1,2,3 --checkname deeplab-resnet --eval-interval 1 --dataset coco PATH_TO_CHECKPOINT='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/train/' PATH_TO_VIS_DIR='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/vis/' PATH_TO_DATASET='/home/rjw/tf-models/research/deeplab/datasets/cityscapes/tfrecord' WORK_DIR='/home/rjw/tf-models/research/deeplab' # From tensorflow/models/research/ python "${WORK_DIR}"/vis.py \ --logtostderr \ --vis_split="val" \ --model_variant="xception_65" \ --atrous_rates=6 \ --atrous_rates=12 \ --atrous_rates=18 \ --output_stride=16 \ --decoder_output_stride=4 \ --vis_crop_size=1025 \ --vis_crop_size=2049 \ --dataset="cityscapes" \ --colormap_type="cityscapes" \ --checkpoint_dir=${PATH_TO_CHECKPOINT} \ --vis_logdir=${PATH_TO_VIS_DIR} \ --dataset_dir=${PATH_TO_DATASET}
