Code for the paper "Are Sixteen Heads Really Better than One?"

Paul Michel

Last update: Dec 14, 2022

Related tags

Overview

Are Sixteen Heads Really Better than One?

This repository contains code to reproduce the experiments in our paper Are Sixteen Heads Really Better than One?.

Prerequisite

First, you will need python >=3.6 with pytorch>=1.0. Then, clone our forks of fairseq (for MT experiments) and pytorch-pretrained-BERT (for BERT):

# Fairseq
git clone https://github.com/pmichel31415/fairseq
# Pytorch pretrained BERT
git clone https://github.com/pmichel31415/pytorch-pretrained-BERT
cd pytorch-pretrained-BERT
git checkout paul
cd ..

If you are running into issues with pytorch-pretrained-BERT (because you have another version installed globally for instance), check out this work around (thanks @insop).

You will also need sacrebleu to evaluate BLEU score (pip install sacrebleu).

Ablation experiments

BERT

Running

bash experiments/BERT/heads_ablation.sh MNLI

Will fine-tune a pretrained BERT on MNLI (stored in ./models/MNLI) and perform the individual head ablation experiment from Section 3.1 in the paper alternatively you can run the experiment with CoLA, MRCP or SST-2 as a task in place of MNLI.

MT

You can obtain the pretrained WMT model from ~~this link from the fairseq repo~~ now this link. Use the Moses tokenizer and subword-nmt in conjunction to the BPE codes provided with the pretrained model to prepair any input file you want. Then run:

bash experiments/MT/wmt_ablation.sh $BPE_SEGMENTED_SRC_FILE $DETOKENIZED_REF_FILE

Systematic Pruning Experiments

BERT

To iteratively prune 10% heads in order of increasing importance run

bash experiments/BERT/heads_pruning.sh MNLI --normalize_pruning_by_layer

This will reuse the BERT model fine-tuned if you have run the ablation experiment before (otherwise it'll just fine-tune it for you). The output of this is very verbose, but you can get the gist of the result by calling grep "strategy\|results" -A1 on the output.

WMT

Similarly, just run:

bash experiments/MT/prune_wmt.sh $BPE_SEGMENTED_SRC_FILE $DETOKENIZED_REF_FILE

You might want to change the paths in the experiment files to point to the binarized fairseq dataset on whic you want to estimate importance scores.

Comments

Not able to prune the BERT model

Hi, I am running the command

bash experiments/BERT/heads_ablation.sh MNLI

I am getting the following error

Traceback (most recent call last):
  File "pytorch-pretrained-BERT/examples/run_classifier.py", line 578, in <module>
    main()
  File "pytorch-pretrained-BERT/examples/run_classifier.py", line 275, in main
    model.bert.mask_heads(to_prune)
  File "/home/pdguest/ishita/py-env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 535, in __getattr__
    type(self).__name__, name))
AttributeError: 'BertModel' object has no attribute 'mask_heads'

need more info

opened by ishita1995 7

Is BERT finetuned after pruning?

Hi, I'm currently working on attention head pruning on models. I think in your reported experiments, you fine-tuned bert when training downstream MNLI task, right? But does it also work to fix the bert representation after pruning and train downstream MNLI task? I appreciate your answer

opened by Huan80805 2
Not able to obtain pretrained WMT model

Hello,

I am trying to run the MT ablation experiments. When I ran the command

wget https://s3.amazonaws.com/fairseq-py/models/wmt14.en-fr.joined-dict.transformer.tar.bz2

I get the following error

--2020-03-31 17:58:31-- https://s3.amazonaws.com/fairseq-py/models/wmt14.en-fr.joined-dict.transformer.tar.bz2 Resolving s3.amazonaws.com (s3.amazonaws.com)... 52.216.207.205 Connecting to s3.amazonaws.com (s3.amazonaws.com)|52.216.207.205|:443... connected. HTTP request sent, awaiting response... 404 Not Found 2020-03-31 17:58:31 ERROR 404: Not Found.

opened by marwash25 2
Why do we need different normalization for all the layers compared to the last layer in BERT during importance score calculation?

Hi,

I am trying to understand why did we need different normalization factors for the last layer of the BERT compared to all other layers?

https://github.com/pmichel31415/pytorch-pretrained-BERT/blob/18a86a7035cf8a48d16c101a66e439bf6ab342f1/examples/classifier_eval.py#L246 vs https://github.com/pmichel31415/pytorch-pretrained-BERT/blob/18a86a7035cf8a48d16c101a66e439bf6ab342f1/examples/classifier_eval.py#L247

opened by Hritikbansal 1
a question about run_classifier.py
1、 (1)I do this and get a pruned model: model.bert.prune_heads(to_prune) (2) I set n_retrain_steps_after_pruning a value greater than 0 next： then:

to retrain my pruned model, that is ok?

2、I don't understand the difference between above method and retrain_pruned_heads（the following method）

THANK YOU !
opened by Ixuanzhang 1

BERT actually_prune option not working

Hi,

thanks for your code! The pruning works great using masking. However, when I tried to actually prune the model to see, if there's a speedup, it fails.

bash experiments/BERT/heads_pruning.sh SST-2 --actually_prune

The new_layer in prune_linear_layer has to be moved to the correct device.

new_layer.to(layer.weight.device)

The forward function fails, because the input and output shape of the previous layer do not seem to match:

13:09:27-INFO: Evaluating following pruning strategy
13:09:27-INFO: 9:3 10:10 11:3,7,8,9,10
13:09:27-INFO: ***** Running evaluation *****
13:09:27-INFO:   Num examples = 872
13:09:27-INFO:   Batch size = 32
Evaluating:   0%|                                                                | 0/28 [00:00<?, ?it/s]Traceback (most recent call last):
  File "pytorch-pretrained-BERT/examples/run_classifier.py", line 578, in <module>
    main()
  File "pytorch-pretrained-BERT/examples/run_classifier.py", line 514, in main
    scorer=processor.scorer,
  File "/home/glock/projects/are-16-heads-really-better-than-1/pytorch-pretrained-BERT/examples/classifier_eval.py", line 78, in evaluate
    input_ids, segment_ids, input_mask, label_ids)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/glock/projects/are-16-heads-really-better-than-1/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 1072, in forward
    output_all_encoded_layers=False, return_att=return_att)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/glock/projects/are-16-heads-really-better-than-1/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 769, in forward
    output_all_encoded_layers=output_all_encoded_layers)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/glock/projects/are-16-heads-really-better-than-1/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 458, in forward
    hidden_states, attn = layer_module(hidden_states, attention_mask)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/glock/projects/are-16-heads-really-better-than-1/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 441, in forward
    attention_output, attn = self.attention(hidden_states, attention_mask)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/glock/projects/are-16-heads-really-better-than-1/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 335, in forward
    self_output, attn = self.self(input_tensor, attention_mask)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/glock/projects/are-16-heads-really-better-than-1/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 274, in forward
    mixed_query_layer = self.query(hidden_states)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/modules/linear.py", line 92, in forward
    return F.linear(input, self.weight, self.bias)
  File "/home/glock/.pyenv/versions/pruning/lib/python3.7/site-packages/torch/nn/functional.py", line 1408, in linear
    output = input.matmul(weight.t())
RuntimeError: size mismatch, m1: [4096 x 768], m2: [576 x 768] at /opt/conda/conda-bld/pytorch_1556653114079/work/aten/src/THC/generic/THCTensorMathBlas.cu:268

opened by pglock 1

No code on master?

In the README, it asks us to checkout a branch not put on master. I don't think that branch is uploaded to master, as the github has no actual code in it.

opened by aninrusimha 1

Is the code still able to run?

Hi,

I am trying to reproduce your result of BERT. I followed the Prerequisite:

# Pytorch pretrained BERT
git clone https://github.com/pmichel31415/pytorch-pretrained-BERT
cd pytorch-pretrained-BERT
git checkout paul
cd ..

# Install the pytorch-pretrained_BERT:
cd pytorch-pretrained-BERT
pip install .
cd ..

# Run the code:
bash experiments/BERT/heads_ablation.sh MNLI

But got this error:

02:06:57-INFO: Weights of BertForSequenceClassification not initialized from pretrained model: ['classifier.weight', 'classifier.bias']
02:06:57-INFO: Weights from pretrained model not used in BertForSequenceClassification: ['cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias']
Traceback (most recent call last):
  File "pytorch-pretrained-BERT/examples/run_classifier.py", line 582, in <module>
    main()
  File "pytorch-pretrained-BERT/examples/run_classifier.py", line 275, in main
    model.bert.mask_heads(to_prune)
  File "/home/guest/anaconda3/envs/huggingface_env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 594, in __getattr__
    type(self).__name__, name))
AttributeError: 'DataParallel' object has no attribute 'bert'


1(standard_in) 2: syntax error
        0.00000(standard_in) 2: syntax error
        0.00000(standard_in) 2: syntax error
        0.00000(standard_in) 2: syntax error
        0.00000(standard_in) 2: syntax error
        0.00000(standard_in) 2: syntax error

Any idea or suggestion?

opened by bing0037 2

about the params: --raw-text and --transformer-mask-heads

Hi ! @pmichel31415 1.In are-16-heads-really-better-than-1/experiments/MT/prune_wmt.sh you have the --raw-text $EXTRA_OPTIONS, and I don't know the meaning. Can you tell me its explanation and how to use it? It is the origin ref text or something? 2. I don't know how to use the --transformer-mask-heads . Can you show me an example?

opened by LiangQiqi677 0
RuntimeError: can't retain_grad on Tensor that has requires_grad=False

Sorry to bother you. I met a bug druing runing the "heads_pruning.sh", and the error is：

12:21:27-INFO: ***** Running evaluation ***** 12:21:27-INFO: Num examples = 9815 12:21:27-INFO: Batch size = 32 Evaluating: 0% 0/307 [00:00<?, ?it/s]Traceback (most recent call last): File "pytorch-pretrained-BERT/examples/run_classifier.py", line 585, in main() File "pytorch-pretrained-BERT/examples/run_classifier.py", line 521, in main scorer=processor.scorer, File "/content/drive/My Drive/XAI in NLP/pytorch-pretrained-BERT/examples/classifier_eval.py", line 78, in evaluate input_ids, segment_ids, input_mask, label_ids) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/content/drive/My Drive/XAI in NLP/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 1072, in forward output_all_encoded_layers=False, return_att=return_att) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/content/drive/My Drive/XAI in NLP/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 769, in forward output_all_encoded_layers=output_all_encoded_layers) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/content/drive/My Drive/XAI in NLP/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 458, in forward hidden_states, attn = layer_module(hidden_states, attention_mask) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/content/drive/My Drive/XAI in NLP/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 441, in forward attention_output, attn = self.attention(hidden_states, attention_mask) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/content/drive/My Drive/XAI in NLP/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 335, in forward self_output, attn = self.self(input_tensor, attention_mask) File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/content/drive/My Drive/XAI in NLP/pytorch-pretrained-BERT/pytorch_pretrained_bert/modeling.py", line 307, in forward self.context_layer_val.retain_grad() File "/usr/local/lib/python3.6/dist-packages/torch/tensor.py", line 326, in retain_grad raise RuntimeError("can't retain_grad on Tensor that has requires_grad=False") RuntimeError: can't retain_grad on Tensor that has requires_grad=False Evaluating: 0% 0/307 [00:00<?, ?it/s]

I don't know how to fix it. Hope you can help me!

opened by YJiangcm 1

Code for the paper "Are Sixteen Heads Really Better than One?"

Related tags

Overview

Are Sixteen Heads Really Better than One?

Prerequisite

Ablation experiments

BERT

MT

Systematic Pruning Experiments

BERT

WMT

Comments

Owner

Paul Michel

Code for the paper "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer"

Code for the paper "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer"

Code associated with the "Data Augmentation using Pre-trained Transformer Models" paper

Code for CVPR 2021 paper: Revamping Cross-Modal Recipe Retrieval with Hierarchical Transformers and Self-supervised Learning

This repository will contain the code for the CVPR 2021 paper "GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields"

Code for ACL 2021 main conference paper "Conversations are not Flat: Modeling the Intrinsic Information Flow between Dialogue Utterances".

Code from the paper "High-Performance Brain-to-Text Communication via Handwriting"

source code for paper: WhiteningBERT: An Easy Unsupervised Sentence Embedding Approach.

Pytorch code for ICRA'21 paper: "Hierarchical Cross-Modal Agent for Robotics Vision-and-Language Navigation"

Code for our paper "Mask-Align: Self-Supervised Neural Word Alignment" in ACL 2021

Code for our ACL 2021 paper - ConSERT: A Contrastive Framework for Self-Supervised Sentence Representation Transfer

Code for our ACL 2021 (Findings) Paper - Fingerprinting Fine-tuned Language Models in the wild .

Code for our paper "Transfer Learning for Sequence Generation: from Single-source to Multi-source" in ACL 2021.

Code and datasets for our paper "PTR: Prompt Tuning with Rules for Text Classification"

This repository contains the code, data, and models of the paper titled "XL-Sum: Large-Scale Multilingual Abstractive Summarization for 44 Languages" published in Findings of the Association for Computational Linguistics: ACL 2021.

Code for paper "Which Training Methods for GANs do actually Converge? (ICML 2018)"

Code and checkpoints for training the transformer-based Table QA models introduced in the paper TAPAS: Weakly Supervised Table Parsing via Pre-training.

This is the code for the EMNLP 2021 paper AEDA: An Easier Data Augmentation Technique for Text Classification

This repository contains the code for EMNLP-2021 paper "Word-Level Coreference Resolution"