Jinming-Su / DeepLab-v3-plus-cityscapes-Res50

DeepLab V3plus

The implementation of Deeplab_v3plus. This repository is based on the dataset of cityscapes and the mIOU is 80.12.

I am working with python3.5 and pytorch1.0.0 built from source. Other environments are not tested, but you need at least pytorch1.0 since I use torch.distributed to manipulate my gpus. I use two 1080ti to train my model, so you also need two gpus each of which should have at least 9G memory.

Dataset

The experiment is done with the dataset of CityScapes. You need to register on the website and download the dataset images and annotations. Then you create a data directory and then decompress.

    $ cd DeepLabv3plus
    $ mkdir -p data
    $ mv /path/to/leftImg8bit_trainvaltest.zip data
    $ mv /path/to/gtFine_trainvaltest.zip data
    $ cd data
    $ unzip leftImg8bit_trainvaltest.zip
    $ unzip gtFine_trainvaltest.zip

Train && Eval

After creating the dataset, you can train on the cityscapes train set and evaluate on the validation set.
Train:

    $ cd DeepLabv3plus
    $ CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 train.py

This will take around 13 hours on two 1080ti gpus. After training, the model will be evaluated on the val set automatically, and you will see a mIOU of 80.12%(backbone res101, evaluated with mutiple scale and flip). 78.477%(backbone res101, single scale, no flip) after 40000 iters

Eval: If you want to evaluate a trained model, you can also do this:

    $ python evaluate.py

or if you want to evaluate on multi-gpus, you can also do this:

    $ CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 evaluate.py

Configurations

• If you want to use your own dataset, you may implement you dataset file as does with my cityscapes.py.

• As for the hyper-parameters, you may change them in the configuration file configs/configurations.py.

Outreach

We re-trained the model and got a mIoU of 78.457% after ~60000 iters(backbone res101, evaluated with single scale and no flip), but we got an unbelievable mIoU of 78.398 after ~60000 iters when using res50 as backbone (evaluated with single scale and no flip).

我们使用了别人(coin)实现的pytorch版本的deeplabv3+(https://github.com/CoinCheung/DeepLab-v3-plus-cityscapes) 在cityscape上进行训练，结果和coin几乎一致. 但是我们将代码的backbone换成了res50，其余所有不变，精度几乎达到了res101的精度，单尺度evl集达到了很诧异的78.398(只比deeplabv3+使用Res101的backbone低了0.06%，比同样使用res50 backbone的PSPNet精度高了1个点 https://github.com/hszhao/semseg). 更让人意外的是deeplabv3+原作tensorflow (https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md) 中使用xception65 backbone的结果只有78.79%，xception71为80.31%

所以是说deeplabv3+在cityscape上其实backbone使用res50的就可以达到接近于res101、xception65 的精度？？？？？还是怎么

相关记录见slurm-11413.out(60000次迭代版本)