FasterSeg: Searching for Faster Real-time Semantic Segmentation [PDF]

Wuyang Chen, Xinyu Gong, Xianming Liu, Qian Zhang, Yuan Li, Zhangyang Wang

In ICLR 2020.

Overview

Our predictions on Cityscapes Stuttgart demo video #0

We present FasterSeg, an automatically designed semantic segmentation network with not only state-of-the-art performance but also faster speed than current methods.

Highlights:

• Novel search space: support multi-resolution branches.
• Fine-grained latency regularization: alleviate the "architecture collapse" problem.
• Teacher-student co-searching: distill the teacher to the student for further accuracy boost.
• SOTA: FasterSeg achieves extremely fast speed (over 30\% faster than the closest manually designed competitor on CityScapes) and maintains competitive accuracy.
  • see our Cityscapes submission here.

Methods

Prerequisites

• Ubuntu 16.04
• Python 3.6.8
• CUDA 10.1 (lower versions may work but were not tested)
• NVIDIA GPU (>= 11G graphic memory) + CuDNN v7.3

This repository has been tested on GTX 1080Ti. Configurations (e.g batch size, image patch size) may need to be changed on different platforms.

Installation

• Clone this repo:

  git clone https://github.com/chenwydj/FasterSeg.git
  cd FasterSeg

• Install dependencies:

  pip install -r requirements.txt

• Install PyCuda which is a dependency of TensorRT.
• Install TensorRT (v5.1.5.0): a library for high performance inference on NVIDIA GPUs with Python API.

Usage

• Work flow: pretrain the supernet → search the archtecture → train the teacher → train the student.
• You can monitor the whole process in the Tensorboard.

0. Prepare the dataset

• Download the leftImg8bit_trainvaltest.zip and gtFine_trainvaltest.zip from the Cityscapes.
• Prepare the annotations by using the createTrainIdLabelImgs.py.
• Put the file of image list into where you save the dataset.
• Remember to properly set the C.dataset_path in the config files mentioned below.

1. Search

cd search

1.1 Pretrain the supernet

We first pretrain the supernet without updating the architecture parameter for 20 epochs.

• Set C.pretrain = True in config_search.py.
• Start the pretrain process:

  CUDA_VISIBLE_DEVICES=0 python train_search.py

• The pretrained weight will be saved in a folder like FasterSeg/search/search-pretrain-256x512_F12.L16_batch3-20200101-012345.

1.2 Search the architecture

We start the architecture searching for 30 epochs.

• Set the name of your pretrained folder (see above) C.pretrain = "search-pretrain-256x512_F12.L16_batch3-20200101-012345" in config_search.py.
• Start the search process:

  CUDA_VISIBLE_DEVICES=0 python train_search.py

• The searched architecture will be saved in a folder like FasterSeg/search/search-224x448_F12.L16_batch2-20200102-123456.
• arch_0 and arch_1 contains architectures for teacher and student networks, respectively.

2. Train from scratch

• cd FasterSeg/train
• Copy the folder which contains the searched architecture into FasterSeg/train/ or create a symlink via ln -s ../search/search-224x448_F12.L16_batch2-20200102-123456 ./

  2.1 Train the teacher network

• Set C.mode = "teacher" in config_train.py.
• Set the name of your searched folder (see above) C.load_path = "search-224x448_F12.L16_batch2-20200102-123456" in config_train.py. This folder contains arch_0.pt and arch_1.pth for teacher and student's architectures.
• Start the teacher's training process:

  CUDA_VISIBLE_DEVICES=0 python train.py

• The trained teacher will be saved in a folder like train-512x1024_teacher_batch12-20200103-234501

  2.2 Train the student network (FasterSeg)

• Set C.mode = "student" in config_train.py.
• Set the name of your searched folder (see above) C.load_path = "search-224x448_F12.L16_batch2-20200102-123456" in config_train.py. This folder contains arch_0.pt and arch_1.pth for teacher and student's architectures.
• Set the name of your teacher's folder (see above) C.teacher_path = "train-512x1024_teacher_batch12-20200103-234501" in config_train.py. This folder contains the weights0.pt which is teacher's pretrained weights.
• Start the student's training process:

  CUDA_VISIBLE_DEVICES=0 python train.py

3. Evaluation

Here we use our pretrained FasterSeg as an example for the evaluation.

cd train

• Set C.is_eval = True in config_train.py.
• Set the name of the searched folder as C.load_path = "fasterseg" in config_train.py.
• Download the pretrained weights of the teacher and student and put them into folder train/fasterseg.
• Start the evaluation process:

  CUDA_VISIBLE_DEVICES=0 python train.py

• You can switch the evaluation of teacher or student by changing C.mode in config_train.py.

4. Test

We support generating prediction files (masks as images) during training.

• Set C.is_test = True in config_train.py.
• During the training process, the prediction files will be periodically saved in a folder like train-512x1024_student_batch12-20200104-012345/test_1_#epoch.
• Simply zip the prediction folder and submit to the Cityscapes submission page.

5. Latency

5.0 Latency measurement tools

• If you have successfully installed TensorRT, you will automatically use TensorRT for the following latency tests (see function here).
• Otherwise you will be switched to use Pytorch for the latency tests (see function here).

5.1 Measure the latency of the FasterSeg

• Run the script:

  CUDA_VISIBLE_DEVICES=0 python run_latency.py

5.2 Generate the latency lookup table:

• cd FasterSeg/latency
• Run the script:

  CUDA_VISIBLE_DEVICES=0 python latency_lookup_table.py

  which will generate an .npy file. Be careful not to overwrite the provided latency_lookup_table.npy in this repo.

• The .npy contains a python dictionary mapping from an operator to its latency (in ms) under specific conditions (input size, stride, channel number etc.)

Citation

@inproceedings{chen2020fasterseg,
  title={FasterSeg: Searching for Faster Real-time Semantic Segmentation},
  author={Chen, Wuyang and Gong, Xinyu and Liu, Xianming and Zhang, Qian and Li, Yuan and Wang, Zhangyang},
  booktitle={International Conference on Learning Representations},
  year={2020}
}

Acknowledgement

• Segmentation training and evaluation code from BiSeNet.
• Search method from the DARTS.
• slimmable_ops from the Slimmable Networks.
• Segmentation metrics code from PyTorch-Encoding.