Few-Shot Image Classification System (FSICS)

a graduation project

Requirements

• Windows 10

Python packages

• torch 1.10.2
• torchaudio 0.10.2
• torchvision 0.11.3
• timm 0.5.4
• pandas 1.4.1
• numpy 1.19.5

Directory structure

• checkpoint pre-trained backbone used to generate feature of images, you can acqiure open source pre-trained backbones at CONFIG.md
• config command-line argument files, e.g., the confing of feature.py is feature.ymal
• csv output of feature.py, <image_path, label> pair list of dataset
• dataset dataset of images after preprocessing (e.g. all images are RGB three-channel). The folder structure has no special requirements, modify your code according to feature.py
• featute output of feature.py, i.e., the features extracted by backbone
• picture reasearch results of my project
• misc miscellaneous python scripts for experimenting training tricks
• moco-v3 open source repo forked from https://github.com/facebookresearch/moco-v3
• utils common function implementations,e.g., linear, transforms etc.
• feature.py python script used to extract feature of dataset with specific backbone
• real.py python script used to train a liner classifier, then classify query set

Usage

Step 1: train a feature extractor (RECOMMENDED: Skipping this step by using a pre-trained checkpoint resided in .\checkpoint)

Build a vit-samll using single-node GPU by training on ImageNet

python main_moco.py \
  -a vit_small -b 1024 \
  --optimizer=adamw --lr=1.5e-4 --weight-decay=.1 \
  --epochs=300 --warmup-epochs=40 \
  --stop-grad-conv1 --moco-m-cos --moco-t=.2 \
  --dist-url 'tcp://localhost:10001' \
  --multiprocessing-distributed --world-size 1 --rank 0 \
  [your imagenet-folder with train and val folders]

Then, move the checkpoint to ./checkpoint/

See CONFIG.md for more examples

Step 2: Compute feture of support set and query set

Check if checkpoint you used is added to the path list in .\feature.py, if ture, run

python feature.py \
    --image_size 224 \
    --batch_size 128 \
    --workers 8 \
    --dataset_path "dataset//support_real" \
    --set "supp"

and run

python feature.py \
    --image_size 224 \
    --batch_size 128 \
    --workers 8 \
    --dataset_path "dataset//query" \
    --set "query"

$(set_name).csv and $(set_name).npy are stored in ./csv and ./feature respectively

See config\feature.ymal for more config parameters about feature.py

Step 3: train a linear classifier

python real.py \
      --logistic_batch_size 1 \
      --support_init_enabled True \
      --mixup_enabled False \
      --logistic_epochs 500 \
      --way 5 \
      --shot 10
      --support_npy "feature//support_real-mocov3-vit_small_s.npy" \
      --query_npy "feature//query-mocov3-vit_small.npy" \
      --out_path "real_result"  \
      --query_csv "csv//query.csv" \

See config\real.ymal for more config parameters about real.py

the results of classification will be stored in ./real_result/$(label_no)/$(img_no).jpg

An example

a 5way-10shot support set

five classes, and 10 images per class:

• high rise
• low rise
• bird eye
• drawing
• indoor

result on query set

high-rise

low-rise

bird-eye

drawing

indoor