Skip to content

deep-spin/infinite-former

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Infinite-former

Implementation of the Infinite-former model.

Sorting

To perform experiments on the sorting task, you first need to generate the data. You can do that as (you will need to select the vocabulary size and the sequence length on the file ./sorting/generate_data.py):

python ./sorting/generate_data.py

Then, you simply need to run (changing the sequence length on the script run_sort_inftyformer.sh):

bash run_sort_inftyformer.sh train

Fine-tuning GPT-2

To fine-tune the GPT-2 with the long-term memory, you first need to install the Transformers library as:

pip install --editable ./finetune_gpt2 

Then, to fine-tune the model run the command:

python ./finetune_gpt2/examples/language-modeling/run_clm.py 	\
	--model_name_or_path=gpt2 \
	--model_type=gpt2 \
	--config_name=infinite_memory_transformer_sticky_mem \
	--per_device_eval_batch_size=1 \
	--per_device_train_batch_size=1 \
	--train_file=</path/to/train/data/file> \
	--validation_file=</path/to/val/data/file> \
	--do_train \
	--do_eval \
	--block_size=512 \
	--output_dir=</path/to/output/dir>
	--kl_regularizer \
	--kl_m=.000001

To evaluate the model do:

python ./finetune_gpt2/examples/language-modeling/run_clm.py \
	--model_name_or_path=</path/to/output/dir>
	--model_type=gpt2 \
	--config_name=infinite_memory_transformer_sticky_mem \ 
	--per_device_eval_batch_size=1 \
	--validation_file=</path/to/test/data/file> \
	--do_eval \
	--block_size=512 \
	--output_dir=</path/to/output/dir> 

Citation

@inproceedings{martins2022infinite,
  author    = {Martins, Pedro Henrique and Marinho, Zita and  Martins, Andr{\'e} FT},
  title     = {$\infty $-former: Infinite Memory Transformer},
  booktitle = {Proc. ACL},
  year      = {2022}
}

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published