ljh415 / hierarchical_cnn

Hierarchical_CNN

Reference paper

• Hierarchical image classification in CNNs

Environment

• torch : 1.8.2+cu111
• torchvision : 0.9.2+cu111
• torchmetrics : 0.10.0
• numpy : 1.21.2
• wandb : 0.13.2

pip install -r requirements.txt

Dataset

• Cifar100
  • Use Superclass for Coarse label
  • Use Classes for Fine label

python preprocess.py --dataset cifar100 --data_dir "data_dir" --types train,test

Architecture

• Backbone : Wide_ResNet_50
  • Coarse : Using WRN50 Backbone expcet for layer4
  • Fine : Using whole WRN50 Backbone
• Main Goal is training Fine Label Classification
• Concatenating Coarse feature and Fine feature
• Using 1x1conv layer for matching channel after concatenating
• Simply Using Fc layer for Classifier

Training (in progress)

The following hyperparameters are not the best ones.
The model is still training.

• Training w/ wandb.sweep

train.py --epochs 15 --batch_size 32 --lr 0.008 --backbone wide_resnet_50 --wandb --sweep --sweep_count 5

• Training w/o wandb.sweep

python train.py --epochs 15 --batch_size 32 --lr 0.008 --backbone wide_resnet_50 --wandb --milestones 7,13

• If you want to train only_fine model, just add --only_fine above command

Default Hyperparameter Setting

• optimizer : SGD
  • or you can use Adam
• batch_size : 32
• coarse : 1, fine : 2
  • setting different Loss Weight
  • Coarse : Fine = 1 : 2
• epochs : 15
• lr : 8e-3
• lr_scheduler : ReduceLROnPlateau
  • or you can use MultiStepLR with using milestones argument

optimizing model

• Using wandb.sweep
• Process
  1. Optimizing "only fine model" using wandb.sweep
     • batch_size :32
     • epochs : 15
     • lr : 0.0079
     • milestones : [7, 12]
  2. Traning "coarse+fine model" using above hyper parameter
• Graph only fine vs. coarse+fine model
• Accuracy Comparison (max)

  	only_fine	coarse+fine
  acc	0.787	0.7832

• wandb report Link

To do

• Find best parameter for Coarse+fine model
  • coarse and fine loss weight
  • lr
  • etc..
• Change the layer to freeze on the Coarse Backbone
• Apply other dataset with hierarchical label
  • Deep Fashion
  • etc..
• Apply other Backbone
  • BiT
• Apply different feature synthesis instead of concatenating