Combined Convolutional and Recurrent Neural Networks for Hierarchical Classification of Images
Author: Jaehoon Koo
This directory contains sample scripts of DHNN paper (https://ieeexplore.ieee.org/abstract/document/9378237) pusblished in IEEE BigData 2020.
Scripts are written in Python3 with Tensorflow based on TF-Slim library (https://github.com/google-research/tf-slim), SENet (https://github.com/kobiso/SENet-tensorflow-slim), CBAM (https://github.com/kobiso/CBAM-keras), and bidirectional RNN (https://github.com/aymericdamien/TensorFlow-Examples/blob/master/examples/3_NeuralNetworks/bidirectional_rnn.py).
Dataset:
https://northwestern.app.box.com/s/i23bm7eee5irjt164sn1zqzxvyyyrc37/folder/70446923518
Files:
code/hc-s2s-oi-v1-2-pr-run-val-se.py: General tree model with SE net
code/hc-s2s-oi-v1-2-pr-run-val-att.py: General tree model with CBAM net
code/PR-node-bulk-val-te.py: Compute node accuracy of general tree model
code/PR-path-bulk-ori-val-te.py: Compute path accuracy of general tree model
To run:
python hc-s2s-oi-v1-2-pr-run-val-se.py -server="nu" -model="resnet-50" -idx="idx-0" -batch=32 -lr=0.0001 -restore="" -prt="full" -opt="adam" -epoch=15 -base="res" -seed=1234
python hc-s2s-oi-v1-2-pr-run-val-se.py -server="nu" -model="hc" -idx="idx-0" -batch=32 -lr=0.0001 -input=2048 -hidden=1024 -conversion=3.1 -average="1;2;3 4" -restore="" -prt="full" -k=2 -opt="adam" -epoch=20 -alt1=-1 -alt2=20 -base="res" -seed=1234
python hc-s2s-oi-v1-2-pr-run-val-att.py -server="nu" -model="resnet-50" -idx="idx-0" -batch=32 -lr=0.0001 -restore="" -prt="full" -opt="adam" -epoch=15 -base="res" -seed=1234 -attention=cbam_block
python hc-s2s-oi-v1-2-pr-run-val-att.py -server="nu" -model="hc" -idx="idx-0" -batch=32 -lr=0.0001 -input=256 -hidden=512 -conversion=3.1 -average="1;2;3 4" -restore="" -prt="full" -k=2 -opt="adam" -epoch=20 -alt1=-1 -alt2=20 -base="res" -seed=1234 -attention=cbam_block