- Notifications
You must be signed in to change notification settings - Fork3
[ICML 2024] Outlier-Efficient Hopfield Layers for Large Transformer-Based Models
License
MAGICS-LAB/OutEffHop
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
This is the code of the paperOutEffHop. You can use this repo to reproduce the results in the paper.
You can set up the experimental environment by running the following command line:
Set locale variables and add the project root directory to your pythonpath:$export LC_ALL=C.UTF-8$export LANG=C.UTF-8$cd OutEffHop/$ pip install --upgrade --no-deps pip$export PYTHONPATH=${PYTHONPATH}:$(realpath"$PWD")
For experiment in paper section 4.1, outlier efficiency of BERT and OPT:
$ conda create -n outlier python==3.9# Run the pip module as a script.$ python -m pip install -r /your_path/OutEffHop/OutEffHop/requirements.txtFor the experiment in paper section 4.1 about STanHop :
$ conda create -n STHM python==3.8# Run the pip module as a script.$ python -m pip install -r /your_path/OutEffHop/STanHop_time_seeries/requirements.txtIf you want run the experiment of STanHop quantize, please install below enviroment:
$ conda create -n quantize_STHM python==3.8 $ python -m pip install -r /your_path/OutEffHop/OutEffHop/STanHop_outlier/quantize_requirements.txt
All the training scripts (batch size, etc.) are set up to fit on two single A100 80GB GPU on Slrum machine.
| Model | Softmax | Script |
|---|---|---|
| BERT-base | vanilla, clipped softmax, gated attention, gated OutEffHop, clipped OutEffHop, OutEffHop | OutEffHop_script/submit_outlier_bert.sh |
| OPT-125m | vanilla, clipped softmax, gated attention, gated OutEffHop, clipped OutEffHop, OutEffHop | OutEffHop_script/submit_outlier_opt.sh |
| STanHop | vanilla, clipped softmax, gated attention, gated OutEffHop, clipped OutEffHop, OutEffHop | OutEffHop_script/submit_STHM_outlier.sh |
After the model is trained, you can run evaluation (both floating point, and quantized) usingthe following commands.Make sure to pass the same softmax method arguments that were used for pre-training (e.g.,--attn_softmax vanilla,--attn_softmax "clipped(-.025:1)",--attn_softmax softmax1,--attn_gate_type conditional_per_token --attn_gate_mlp,--attn_gate_type conditional_per_token --attn_gate_init 0.25 etc.)
Run command:
$ accelerate launch --config_file accelerate_configs/1gpu_no_mp.yaml validate_mlm_config.py \--seed 3000 \--dataset_setup bookcorpus_and_wiki \--preprocessing_num_workers 8 \--model_type bert \--max_seq_length 128 \--mlm_probability 0.15 \--per_device_eval_batch_size 32 \--attn_softmax"clippedsoftmax1(-.025:1)" \--data_cache_dir .hf_data \--model_cache_dir .hf_cache \--model_name_or_path output/clipped_softmax1 \--output_dir /output_metrics/clipped_softmax1-3000Run command:
$ accelerate launch --config_file accelerate_configs/1gpu_no_mp.yaml validate_mlm_config.py \--quantize \--est_num_batches 16 \--seed 4000 \--dataset_setup bookcorpus_and_wiki \--preprocessing_num_workers 8 \--model_type bert \--max_seq_length 128 \--mlm_probability 0.15 \--per_device_eval_batch_size 32 \--attn_softmax"clippedsoftmax1(-.025:1)" \--data_cache_dir .hf_data \--model_cache_dir .hf_cache \--model_name_or_path output/clipped_softmax1 \--output_dir output_metrics/bert_quantize_clipped_softmax1-4000Run command:
$ accelerate launch --config_file accelerate_configs/1gpu_no_mp.yaml validate_clm.py \--seed 5678 \--dataset_setup bookcorpus_and_wiki \--preprocessing_num_workers 32 \--model_type opt \--block_size 512 \--per_device_eval_batch_size 4 \--attn_gate_type conditional_per_token \--attn_gate_init 0.25 \--data_cache_dir .hf_data \--model_cache_dir .hf_cache \--model_name_or_path output/gate_opt \--output_dir output_metrics/opt_gate_attention-5678
Run command:
$ accelerate launch --config_file accelerate_configs/1gpu_no_mp.yaml validate_clm.py \--quantize \--quant_setup fp32_head \--ranges_acts running_minmax \--qmethod_acts asymmetric_uniform \--percentile 99.999 \--est_num_batches 4 \--seed 6789 \--dataset_setup bookcorpus_and_wiki \--preprocessing_num_workers 32 \--model_type opt \--block_size 512 \--per_device_eval_batch_size 1 \--attn_gate_type conditional_per_token \--attn_gate_init 0.25 \--data_cache_dir .hf_data \--model_cache_dir .hf_cache \--model_name_or_path output/gate_opt \--output_dir output_metrics/opt_quantize_gate_attention-6789
Run command:
$ accelerate launch --config_file accelerate_configs/1gpu_no_mp.yaml validate_clm.py \--quantize \--quant_setup fp32_head \--ranges_acts running_minmax \--qmethod_acts asymmetric_uniform \--percentile 99.999 \--est_num_batches 4 \--seed 6789 \--dataset_setup bookcorpus_and_wiki \--preprocessing_num_workers 32 \--model_type opt \--block_size 512 \--per_device_eval_batch_size 1 \--attn_gate_type conditional_per_token \--attn_gate_init 0.25 \--data_cache_dir .hf_data \--model_cache_dir .hf_cache \--model_name_or_path output/gate_opt \--output_dir output_metrics/opt_quantize_gate_attention-6789
Run command:
$ python main_stanhop.py --data ETTh1 --in_len 168 --out_len 24 --seg_len 6 --learning_rate 1e-4 --itr 1 --mode softmax1 --use_gpu --gpu 0 --batch_size 128 --run_name STHM_softmax1 --e_layers 11 --save_np --with_tracking
Run command:
$ python quantized_main_stanhop.py \ --data ETTh1 \ --in_len 168 \ --out_len 24 \ --seg_len 6 \ --learning_rate 1e-4 \ --itr 1 \ --mode softmax \ --use_gpu \ --gpu 0 \ --batch_size 128 \ --run_name STHM_softmax \ --e_layers 11 \ --quantize \ --quantize_model_path OutEffHop/OutEffHop/STanHop_outlier/checkpoints1/stanhop_ETTh1_il168_ol24_sl6_win1_fa10_dm256_nh4_el11_itr0_softmax/checkpoint.pth \ --seed$((i*1000))> OutEffHop/OutEffHop/STanHop_outlier/results/stanhop_ETTh1_quantized/softmax_seq24_$i.txt
You can set up the experimental environment by running the following command line:
$cd STanHop_time_seeries$ pip3 install -r requirements.txt$export PYTHONPATH=$PYTHONPATH:$PWD
Put datasets to conduct experiments into folder
datasets/. We have already putETTh1andETTm1into it.WTHandECLcan be downloaded fromhttps://github.com/zhouhaoyi/Informer2020.ILIandTrafficcan be downloaded fromhttps://github.com/thuml/Autoformer. Note that theWTHwe used in the paper is the one with 12 dimensions from Informer, not the one with 21 dimensions from Autoformer.To get results of Crossformer with
$T=168, \tau = 48, L_{seg} = 6$ on ETTh1 dataset, run:
python main_stanhop.py --data ETTh1 --in_len 168 --out_len 48 --seg_len 6 --learning_rate 1e-4 --itr 1 --mode softmax1 --use_gpu --gpu 0 --batch_size 128 --run_name STHM_softmax1 --e_layers 11The model will be automatically trained and tested. The trained model will be saved in foldercheckpoints/ and evaluated metrics will be saved in folderresults/.
- To reproduce all results in the paper, run following scripts to get corresponding results:
batch OutEffHop_script/submit_STHM.shmain_stanhop is the entry point of our model and there are other parameters that can be tuned. Here we describe them in detail:
| Parameter name | Description of parameter |
|---|---|
| data | The dataset name |
| root_path | The root path of the data file (defaults to./datasets/) |
| data_path | The data file name (defaults toETTh1.csv) |
| data_split | Train/Val/Test split, can be ratio (e.g.0.7,0.1,0.2) or number (e.g.16800,2880,2880), (defaults to0.7,0.1,0.2) |
| checkpoints | Location to store the trained model (defaults to./checkpoints/) |
| in_len | Length of input/history sequence, i.e. |
| out_len | Length of output/future sequence, i.e. |
| seg_len | Length of each segment in DSW embedding, i.e. |
| win_size | How many adjacent segments to be merged into one in segment merging of HED (defaults to 2) |
| factor | Number of routers in Cross-Dimension Stage of TSA, i.e. |
| data_dim | Number of dimensions of the MTS data, i.e. |
| d_model | Dimension of hidden states, i.e. |
| d_ff | Dimension of MLP in MSA (defaults to 512) |
| n_heads | Num of heads in MSA (defaults to 4) |
| e_layers | Num of encoder layers, i.e. |
| dropout | The probability of dropout (defaults to 0.2) |
| weight_decay | The weight decay |
| num_workers | The num_works of Data loader (defaults to 0) |
| batch_size | The batch size for training and testing (defaults to 32) |
| train_epochs | Train epochs (defaults to 20) |
| patience | Early stopping patience (defaults to 3) |
| learning_rate | The initial learning rate for the optimizer (defaults to 1e-4) |
| lradj | Ways to adjust the learning rate (defaults totype1) |
| itr | Experiments times (defaults to 1) |
| save_pred | Whether to save the predicted results. If True, the predicted results will be saved in folderresults in numpy array form. This will cost a lot time and memory for datasets with largeFalse). |
| use_gpu | Whether to use gpu (defaults toTrue) |
| gpu | The gpu no, used for training and inference (defaults to 0) |
| use_multi_gpu | Whether to use multiple gpus (defaults toFalse) |
| devices | Device ids of multile gpus (defaults to0,1,2,3) |
| mode | The type of the Hopfield Network (Hopfield, SparseHopfield, STanHop, OutEffHop) |
| run_name | The name of experiment |
| eta | The eta value of Entmax |
| gamma | The gamma value of Entmax |
You can set up the experimental environment by running the following command line:
$ conda create -n theory_verify python=3.8$ conda activate theory_verify$cd theory_verification$ pip3 install -r requirements.txt$ python3 plotting.py
The experiments in this work benefit from the following open-source codes:
https://github.com/zhouhaoyi/Informer2020
https://github.com/thuml/Autoformer
https://github.com/MAGICS-LAB/STanHop
https://github.com/Qualcomm-AI-research/outlier-free-transformers
If you find our work useful, please consider citing our paper:
@inproceedings{hu2024outlier, title={Outlier-Efficient Hopfield Layers for Large Transformer-Based Models}, author={Hu, Jerry Yao-Chieh and Chang, Pei-Hsuan and Luo, Robin and Chen, Hong-Yu and Li, Weijian and Wang, Wei-Po and Liu, Han}, booktitle={Forty-first International Conference on Machine Learning (ICML)}, year={2024}}About
[ICML 2024] Outlier-Efficient Hopfield Layers for Large Transformer-Based Models