Source codes and datasets for Emergent Modularity in Pre-trained Transformers (Findings of ACL 2023).
Download the checkpoints and evaluation data.
wget https://thunlp.oss-cn-qingdao.aliyuncs.com/zzy/NeuronPredictivityData.tar
tar -xvf NeuronPredictivityData.tar
rm -r NeuronPredictivityData.tar
wget https://thunlp.oss-cn-qingdao.aliyuncs.com/zzy/NeuronPredictivityCheckpoint.tar
tar -xvf NeuronPredictivityCheckpoint.tar
rm -r NeuronPredictivityCheckpoint.tarThen run cd NeuronPredictivity && bash scripts.sh to get the neuron predictivity.
Or directly download the neuron predictivity.
wget https://thunlp.oss-cn-qingdao.aliyuncs.com/zzy/NeuronPredictivityPred.tar
tar -xvf NeuronPredictivityPred.tar
rm -r NeuronPredictivityPred.tarAt each transformer layer, we compute the best predictivity of neurons for each sub-function and then calculate the average best predictivity among all sub-function in each function. We can get the figures (Figure 1 Left) by the following commands.
cd NeuronFunctionalSpecialization
python Figure1Left_maker.py switchT5 semantic
python Figure1Left_maker.py switchT5 knowledge
python Figure1Left_maker.py switchT5 task
python Figure1Left_maker.py T5 semantic
python Figure1Left_maker.py T5 knowledge
python Figure1Left_maker.py T5 task
# We already put the results in the folder.We also calculate the average distribution similarity between different functions of different layers for the same pre-trained models. We can get the figures (Figure 1 Right) by the following commands.
cd NeuronFunctionalSpecialization
python Figure1Right_maker.py T5
python Figure1Right_maker.py switchT5
# We already put the results in the folder.We first get the neurons with the highest 1% predictivities for each sub-function.
cd FunctionalExpert
python TopK_neurons.py --model T5
python TopK_neurons.py --model switchT5
python TopK_neurons_step.py --model T5_step
python TopK_neurons_step.py --model switchT5_stepOr directly download the result.
wget https://thunlp.oss-cn-qingdao.aliyuncs.com/zzy/FunctionalExpertTopKNeurons.tar
tar -xvf FunctionalExpertTopKNeurons.tar
rm -r FunctionalExpertTopKNeurons.tarThen we conduct hypothesis testing to get the functional experts.
cd FunctionalExpert
python HypothesisTesting.py --model T5
python HypothesisTesting.py --model switchT5
python HypothesisTesting_step.py --model T5_step
python HypothesisTesting_step.py --model switchT5_step
python HypothesisTesting4RandomPartitioning.py --model T5
python HypothesisTesting4RandomPartitioning.py --model switchT5Or directly download the result.
wget https://thunlp.oss-cn-qingdao.aliyuncs.com/zzy/FunctionalExpertHypothesisTesting.tar
tar -xvf FunctionalExpertHypothesisTesting.tar
rm -r FunctionalExpertHypothesisTesting.tarFinally, we run cd FunctionalExpert && python Table1and4_maker.py to get the numbers in Table 1 and Table 4.
cd Perturbation4T5We first run python average_predictivity.py to get the average predictivities on all datasets (we already put the results in the folder). We then run python run_perturbation.py to conduct the perturbation experiments for T5 (we already put the results in the folder). Finally, we run python Figure2_maker.py to get the figures (Figure 2).
cd Perturbation4SwitchTransformerWe run bash scripts.sh to conduct the perturbation experiments for Switch Transformer (we already put the results in the folder). Finally, we run python Table3_maker.py to get the numbers in Table 3.
cd Stabilization
python ExpertPredictivity.py --model T5_step
python ExpertPredictivity.py --model switchT5_stepOr directly download the result.
wget https://thunlp.oss-cn-qingdao.aliyuncs.com/zzy/StabilizationExpertPredictivity.tar
tar -xvf StabilizationExpertPredictivity.tar
rm -r StabilizationExpertPredictivity.tarThen calculate the the similarity between a layer of two model checkpoints w.r.t. a particular sub-function, at the expert/neuron level.
cd Stabilization
python ExpertAndNeuron_Stabilization.py --model T5_step
python ExpertAndNeuron_Stabilization.py --model switchT5_step
python RandomPartitioning_Stabilization.py --model T5_step --epoch 50
python RandomPartitioning_Stabilization.py --model switchT5_step --epoch 30
# We already put the results in the folder.Finally we can get the figures (Figure 4) by the following commands.
cd Stabilization
python Figure4_maker.py --model T5_step
python Figure4_maker.py --model switchT5_step
# We already put the results in the folder.If you use the code, please cite this paper:
@inproceedings{zhang2023emergent,
title={Emergent Modularity in Pre-trained Transformers},
author={Zhang, Zhengyan and Zeng, Zhiyuan and Lin, Yankai and Xiao, Chaojun and Wang, Xiaozhi and Han, Xu and Liu, Zhiyuan and Xie, Ruobing and Sun, Maosong and Zhou, Jie},
booktitle={Findings of ACL},
year={2023}
}