zhenglab / BioTL

Deep Transfer Learning for Biology Cross-domain Image Classification

BioTL

This repository contains code for the paper Deep Transfer Learning for Biology Cross-domain Image Classification.

Datasets

Five datasets used in our experiments:

• Flowers17
• Flowers102
• Plant Seedlings
• PlanktonSet 1.0
• QUT Fish

The scripts we used for splitting the datasets can be found in utils.

How to use

System requirement:

• PyTorch>=0.3.0
• TorchVision>=0.2.0
• PyTorchNet (up to date)

Train from scratch:

DATASET='flowers17'

python main.py \
    --dataset $DATASET \
    --model alexnet \
    --lr 0.01 \
    --weight-decay 1e-4 \
    --batchsize 32 \
    --print-freq 10 \
    --expname AlexNet \
    --tensorboard \
    --gpu_ids 1 \
    --epochs 300

Fine-tuning on ImageNet:

DATASET='flowers17'

python main.py \
    --pretrained \
    --dataset $DATASET \
    --model alexnet \
    --lr 0.01 \
    --weight-decay 1e-4 \
    --batchsize 32 \
    --print-freq 10 \
    --expname AlexNet \
    --tensorboard \
    --gpu_ids 1 \
    --epochs 300

Transfer learning:

SRC_DATASET='flowers17'
DST_DATASET='flowers102'

python transfer.py \
    --src_dataset $SRC_DATASET \
    --dst_dataset $DST_DATASET \
    --model resnet18 \
    --lr 0.01 \
    --weight-decay 1e-4 \
    --batchsize 16 \
    --print-freq 10 \
    --expname ResNet-18 \
    --tensorboard \
    --gpu_ids 3 \
    --epochs 300 \
    --basemodel '/path/to/'$SRC_DATASET'_checkpoints/ResNet-18/model_best.pth.tar'