daqige / token_selection Goto Github PK

Official implementation of the paper: Efficient Backpropagation with Variance Controlled Adaptive Sampling accepted by ICLR 2024.

VCAS is an efficient unbiased sampling method to accelerate back propagation (BP). It can achieve an up to 73.87% FLOPs reduction of BP with lossless performance under theoretical guarantee. Example Results.

Requirements: Python >= 3.7 + CUDA >= 11.0 + torch >= 1.12.0 + transformers >= 4.21.0

To install VCAS only, run the following:

pip install -v -e .

To further run the examples, run the following :

conda create -n vcas python=3.9
conda activate vcas
pip install -r requirements.txt # install dependencies
pip install -v -e . # install vcas

There are four steps to use VCAS in your project:

1. Add VcasSampleArguments to the argument parser
2. Initialize VcasSampleScheme with VcasSampleArguments
3. Process the original model with VcasModelProcessor
4. Substitute the original huggingface Trainer with VcasTrainer

from vcas import VcasSampleArguments, VcasSampleScheme, VcasModelProcessor, VcasTrainer

......

# CHANGES #1: Add VcasSampleArguments to the argument parser 
parser = HfArgumentParser((..., VcasSampleArguments))
..., sample_args = parser.parse_args_into_dataclasses()

# CHANGES #2: Initialize VcasSampleScheme with VcasSampleArguments
sample_scheme = VcasSampleScheme(sample_args)

# CHANGES #3: Process the original model with VcasModelProcessor
model = ...
act_sampling_repetend = ... # basic block of the model, after which VCAS insert activation sampler, eg. BertLayer in BertModel
processor = VcasModelProcessor(model, act_sampling_repetend, sample_scheme)
processor.process()

# CHANGES #4: Substitute the original huggingface Trainer with VcasTrainer
trainer = VcasTrainer(..., sample_scheme=sample_scheme)

......

For research studies, you can refer to the sampling layers in vcas/layers and the adaptive sampling strategy in vcas/vcas_trainer.py. Help yourself to integrate them into your model and pipeline directly.

We provide an example of BERT fintuning on SST-2 dataset in examples/run_glue.py. Run the following:

cd examples
bash run.sh

We record the running results above and compare it with the exact training baseline (run with bash run_baseline.sh) on NVIDIA 2080Ti. Check the results through https://wandb.ai/thuwzt/vcas.

Help yourself to compare examples/run_glue.py with examples/run_glue_baseline.py to see the modifications and change model_name_or_path and task_name in examples/run.sh to try other models and datasets using VCAS.

VCAS provides a set of hyperparameters which are proven to be insensitive to the performance. However, you can still modify them to fit your specific task. Please check vcas/sample_args.py for more details.

The main hyperparameters are:

• act_var_tau: The acceptable ratio of the activation sampling variance to the SGD variance (default: 0.025), higher value means more aggressive sampling.
• w_var_tau: The acceptable ratio of the weight sampling variance to the SGD variance (default: 0.025), higher value means more aggressive sampling.
• cal_var_freq: The frequency to calculate variance (default: 100), higher value brings less overhead but less sufficient adaptation. We recommend to keep it at least 1/50 of total training steps for thorough adaptation.