FastFCN-paddle

Paddle implementation of FastFCN : A Faster, Stronger and Lighter framework for semantic segmentation

• Original Paper: FastFCN: Rethinking Dilated Convolution in the Backbone for Semantic Segmentation

by Huikai Wu, Junge Zhang, Kaiqi Huang, Kongming Liang, Yizhou Yu.

• Original Pytorch Implementation: https://github.com/wuhuikai/FastFCN

Introduction

We implement the Paper reproduction with Paddle . The results are close to the orginal paper on the ade20k datasets.

A brief introduction about important folders:

diff: log and fake data from recommended procedures of reproducing research papers 论文复现指南

FastFCN-paddle: paddle version of FastFCN

FastFCN-python: pytorch version of FastFCN

Training:

%cd FastFCN_paddle 
!python -m experiments.segmentation.train --dataset ade20k \ 
            --model encnet --jpu JPU --aux --se-loss \ 
            --backbone resnet50 --checkname encnet_res50_ade20k_train 

Evaluating:

%cd FastFCN_paddle 
!python -m experiments.segmentation.test --dataset ade20k \
    --model encnet --jpu JPU --aux --se-loss \
    --backbone resnet50 --resume {MODEL} --split val --mode testval

Results:

Method	DataSet	Environment	Model	Epoch	Miou
FastFCN	ade20k	Tesla V100	EncNet+JPU	96	34.4

Reprod_Log:

forward_diff: forward_diff.log
metric_diff : metric_diff.log
loss_diff : loss_diff.log
bp_align_diff : bp_align_diff.log
train_align_diff : train_align_diff.log
train_log : train.log

AI studio:

• AI studio link : https://aistudio.baidu.com/aistudio/projectdetail/2565388