git clone https://github.com/feymanpriv/delf.git
cd training
bash install.sh
refer training/DELG.md
python train.py \
--train_file_pattern=gldv2_dataset/tfrecord/train* \
--validation_file_pattern=gldv2_dataset/tfrecord/validation* \
--imagenet_checkpoint=/path/to/pretrained_model \
--dataset_version=gld_v2_clean \
--batch_size=256 \
--logdir=output/
training curve shows below.
![](./examples/images/loss_desc_crossentropy.svg)
![](./examples/images/loss_attn_crossentropy.svg)
![](./examples/images/train_accuracy_desc.svg)
![](./examples/images/train_accuracy_attn.svg)
tensorboard --logdir ./output/train_logs/ --port 8198 --host /your/ip/address