This repo is an unofficial pytorch implementation of DFANet:Deep Feature Aggregation for Real-Time Semantic Segmentation

log

• 2019.4.16 After 483 epoches it rases RuntimeError: value cannot be converted to type float without overflow: (9.85073e-06,-3.2007e-06).According to the direction of the stackoverflow the error can be fixed by modifying "self.scheduler.step()" to "self.scheduler.step(loss.cpu().data.numpy())" in train.py.

• 2019.4.24 An function has been writed to load the pretrained model which was trained on imagenet-1k.The project of training the backbone can be Downloaded from here -https://github.com/huaifeng1993/ILSVRC2012. Limited to my computing resources(only have one RTX2080),I trained the backbone on ILSVRC2012 with only 22 epochs.But it have a great impact on the results.

• 2019.5.23 It's hard to improve the performance of the model.May be the model's details are different from the original paper's or the hyperparameters ....or the training strategy...or something else...

• 2020.3.7 rewrited some code of this rep which make the code more modular(model/dfanet.py).

Installation

• pytorch==1.0.0
• python==3.6
• numpy
• torchvision
• matplotlib
• opencv-python
• tensorflow
• tensorboardX

Dataset and pretrained model

Download CityScape dataset and unzip the dataset into data folder.Then run the command 'python utils/preprocess_data.py' to create labels.

Train the network without pretrained model.

Modify your configuration in main.py.

run the command  'python main.py'

curvs on CityScape set

#

inference speed

platform	input size	batch size	inference time /ms
rk3399	320*200	1	960
rk3399	200*100	1	589
rk3399	80*80	1	259
rk3399	72*72	1	161
2080	1024*1024	4	40
2080	1024*1024	1	16
2080	2048*1024	1	17
2080	2048*1024	2	39
2080	512*512	1	39
2080	512*512	16	44

Some experimental results was provided by @ShaoqingGong

To do

• Train the backbone xceptionA on the ImageNet-1k.

• Modify the network and improve the accuracy.

• Debug and report the performance.

• Schedule the lr

• ...

Thanks