Skip to content

Latest commit

 

History

History
537 lines (439 loc) · 18.3 KB

README.md

File metadata and controls

537 lines (439 loc) · 18.3 KB
Raw

Step-by-Step

This document is used to list steps of reproducing PyTorch BERT tuning zoo result. Original BERT documents please refer to BERT README and README.

Note

Dynamic Quantization is the recommended method for huggingface models.

Prerequisite

1. Installation

Python Version

Recommend python 3.6 or higher version.

Install BERT model

cd examples/pytorch/eager/huggingface_models
python setup.py install

Note

Please don't install public transformers package.

Install dependency

cd examples/text-classification
pip install -r requirements.txt

cd ../seq2seq
pip install -r requirements.txt

Install PyTorch

# Install Dependencies
conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi
# Install pytorch from source
git clone https://github.com/pytorch/pytorch.git
cd pytorch
git reset --hard 24aac321718d58791c4e6b7cfa50788a124dae23
git submodule sync
git submodule update --init --recursive
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
python setup.py install
pip install torchvision==0.7.0 --no-deps

2. Prepare pretrained model

Before use Intel® Neural Compressor, you should fine tune the model to get pretrained model, You should also install the additional packages required by the examples:

Text-classification

BERT

  • For BERT base and glue tasks(task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI...)
export TASK_NAME=MRPC

python run_glue.py \
  --model_name_or_path bert-base-cased \
  --task_name $TASK_NAME \
  --do_train \
  --do_eval \
  --max_seq_length 128 \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3 \
  --output_dir /tmp/$TASK_NAME/

NOTE

model_name_or_path : Path to pretrained model or model identifier from huggingface.co/models

task_name : where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.

The dev set results will be present within the text file 'eval_results.txt' in the specified output_dir. In case of MNLI, since there are two separate dev sets, matched and mismatched, there will be a separate output folder called '/tmp/MNLI-MM/' in addition to '/tmp/MNLI/'.

please refer to [BERT base scripts and instructions](examples/text-classification/README.md#PyTorch version).

  • After fine tuning, you can get a checkpoint dir which include pretrained model, tokenizer and training arguments. This checkpoint dir will be used by neural_compressor tuning as below.

XLM-RoBERTa

For BERT base and glue tasks(task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI...)

cd examples/text-classification
export TASK_NAME=MRPC

python run_glue.py 
  --model_name_or_path xlm-roberta-base 
  --task_name $TASK_NAME 
  --do_train 
  --do_eval 
  --max_seq_length 128 
  --per_device_train_batch_size 32 
  --learning_rate 2e-5 
  --num_train_epochs 2 
  --output_dir /path/to/checkpoint/dir

XLNet

For BERT base and glue tasks(task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI...)

cd examples/text-classification
export TASK_NAME=MRPC

python run_glue.py 
  --model_name_or_path xlnet-base-cased 
  --task_name $TASK_NAME  
  --do_train 
  --do_eval 
  --max_seq_length 256 
  --per_device_train_batch_size 32 
  --learning_rate 5e-5 
  --num_train_epochs 5 
  --output_dir /path/to/checkpoint/dir

GPT2

For BERT base and glue tasks(task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI...). First you need to create a model with appropriate number of labels. For example, MRPC is a two-way sentence pair classification task.

from transformers import GPT2ForSequenceClassification, GPT2Tokenizer

model = GPT2ForSequenceClassification.from_pretrained("gpt2", num_labels=2)
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")

# Define a padding token
tokenizer.pad_token = tokenizer.eos_token
model.config.pad_token_id = tokenizer.pad_token_id

# Save the model and tokenizer in a directory
model.save_pretrained("directory")
tokenizer.save_pretrained("directory")

fine tuning command like this.

cd examples/text-classification
export MODEL_PATH=directory
export TASK_NAME=MRPC

python run_glue.py 
  --model_name_or_path $MODEL_PATH
  --task_name $TASK_NAME 
  --do_train 
  --do_eval 
  --max_seq_length 128 
  --per_device_train_batch_size 1 
  --learning_rate 2e-5 
  --num_train_epochs 1 
  --output_dir /path/to/checkpoint/dir

Seq2seq

Install dependency

cd examples/pytorch/eager/huggingface_models/examples/seq2seq
pip install -r requirements.txt

summarization_billsum task

Prepare dataset
wget https://cdn-datasets.huggingface.co/summarization/pegasus_data/billsum.tar.gz
tar -xzvf billsum.tar.gz
Finetune command
export TASK_NAME=summarization_billsum

python run_seq2seq_tune.py \
  --model_name_or_path google/pegasus-billsum \
  --do_train \
  --do_eval \
  --task $TASK_NAME \
  --data_dir /path/to/data/dir \
  --output_dir /path/to/checkpoint/dir \
  --overwrite_output_dir \
  --predict_with_generate \
  --max_source_length 1024 \
  --max_target_length=256 \
  --val_max_target_length=256 \
  --test_max_target_length=256

for example /path/to/data/dir can be ./billsum

translation_en_ro task

Finetune command
export TASK_NAME=translation

python run_seq2seq_tune.py \
  --model_name_or_path t5-small \
  --do_train \
  --do_eval \
  --task $TASK_NAME \
  --data_dir /path/to/data/dir \
  --output_dir /path/to/checkpoint/dir \
  --overwrite_output_dir \
  --predict_with_generate

for example /path/to/data/dir can be ../test_data/wmt_en_ro

NOTE

model_name_or_path : Path to pretrained model or model identifier from huggingface.co/models

task : Task name, summarization (or summarization_{dataset} for pegasus) or translation

Where task name can be one of summarization_{summarization dataset name},translation_{language}_to_{language}.

Where summarization dataset can be one of xsum,billsum etc.

Where output_dir is path of checkpoint which be created by fine tuning.

  • After fine tuning, you can get a checkpoint dir which include pretrained model, tokenizer and training arguments. This checkpoint dir will be used by neural_compressor tuning as below.

Language-modeling

Finetune command
cd examples/pytorch/eager/huggingface_models/examples/language-modeling
python run_clm.py \
  --model_name_or_path microsoft/DialoGPT-small \
  --dataset_name wikitext\
  --dataset_config_name wikitext-2-raw-v1 \
  --do_train \
  --do_eval \
  --output_dir /path/to/checkpoint/dir

NOTE

Until now, enabled dialogpt, reformer.

model_name_or_path : Path to pretrained model or model identifier from huggingface.co/models

dataset_name : Dataset name, can be one of {wikitext, crime_and_punish}.

dataset_config_name : just for dialogpt: wikitext-2-raw-v1.

Start to neural_compressor tune for Model Quantization

cd examples/pytorch/eager/huggingface_models

Glue task

sh run_tuning.sh --topology=topology_name --dataset_location=/path/to/glue/data/dir --input_model=/path/to/checkpoint/dir

NOTE

topology_name can be:{"bert_base_MRPC", "distilbert_base_MRPC", "albert_base_MRPC", "funnel_MRPC", "bart_WNLI", "mbart_WNLI", "xlm_roberta_MRPC", "gpt2_MRPC", "xlnet_base_MRPC", "transfo_xl_MRPC", "ctrl_MRPC", "xlm_MRPC"}

/path/to/checkpoint/dir is the path to finetune output_dir

Seq2seq task

sh run_tuning.sh --topology=topology_name --dataset_location=/path/to/seq2seq/data/dir --input_model=/path/to/checkpoint/dir

NOTE

topology_name can be:{"t5_WMT_en_ro", "marianmt_WMT_en_ro", "pegasus_billsum"}

/path/to/checkpoint/dir is the path to output_dir set in finetune.

/path/to/seq2seq/data/dir is the path to data_dir set in finetune.

for example,

examples/test_data/wmt_en_ro for translation task

examples/seq2seq/billsum for summarization task

Language-modeling task

sh run_tuning.sh --topology=topology_name --input_model=/path/to/checkpoint/dir

NOTE

topology_name can be one of {dialogpt_wikitext, reformer_crime_and_punishment}

/path/to/checkpoint/dir is the path to finetune output_dir

Start to neural_compressor tune for Model Pruning

Below are example NLP tasks for model pruning together with task specific fine-tuning. It requires the pre-trained bert-base sparsity model Intel/bert-base-uncased-sparse-70-unstructured from Intel Huggingface portal. The pruning configuration is specified in yaml file i.e. prune.yaml.

MNLI task

python examples/text-classification/run_glue_no_trainer_prune.py --task_name mnli --max_length 128
       --model_name_or_path Intel/bert-base-uncased-sparse-70-unstructured
       --per_device_train_batch_size 32 --learning_rate 5e-5 --num_train_epochs 3 --output_dir /path/to/output_dir
       --prune --config prune.yaml --output_model /path/to/output/model.pt --seed 5143

SST-2 task

python examples/text-classification/run_glue_no_trainer_prune.py --task_name sst2 --max_length 128
       --model_name_or_path Intel/bert-base-uncased-sparse-70-unstructured
       --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 3 --output_dir /path/to/output_dir
       --prune --config prune.yaml --output_model /path/to/output/model.pt --seed 5143

QQP task

python examples/text-classification/run_glue_no_trainer_prune.py --task_name qqp --max_length 128
       --model_name_or_path Intel/bert-base-uncased-mnli-sparse-70-unstructured-no-classifier
       --seed 42 --per_device_train_batch_size 32 --learning_rate 5e-5 --num_train_epochs 3
       --weight_decay 0.1 --num_warmup_steps 1700 --output_dir /path/to/output_dir
       --prune --config prune.yaml --output_model /path/to/output/model.pt

QNLI task

python examples/text-classification/run_glue_no_trainer_prune.py --task_name qnli --max_length 128
       --model_name_or_path Intel/bert-base-uncased-mnli-sparse-70-unstructured-no-classifier
       --seed 42 --per_device_train_batch_size 32 --learning_rate 5e-5 --num_train_epochs 3
       --weight_decay 0.1 --num_warmup_steps 500 --output_dir /path/to/output_dir
       --prune --config prune.yaml --output_model /path/to/output/model.pt

QnA task

python examples/question-answering/run_qa_no_trainer_prune.py
       --model_name_or_path Intel/bert-base-uncased-sparse-70-unstructured --dataset_name squad
       --max_seq_length 384 --doc_stride 128 --weight_decay 0.01 --num_warmup_steps 900 --learning_rate 1e-4
       --num_train_epochs 2 --per_device_train_batch_size 12 --output_dir /path/to/output_dir
       --prune --config prune.yaml --output_model /path/to/output/model.pt --seed 5143

For gradient sensitivity pruning methods, an example yaml is provide as head_conf.yaml to support the algorithm from https://arxiv.org/abs/2010.13382. SST-2 example is enabled by:

python examples/text-classification/run_glue_no_trainer_gradient_prune.py
      --task_name sst2 --max_length 128 --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 3
      --seed 5143 --model_name_or_path bert-base-uncased --config ./head_conf.yaml
      --do_prune --do_eval --output_model /path/to/output/

Start to neural_compressor tune for Model Distillation

Below are example NLP tasks for model distillation from a task specific fine-tuned large model to a smaller model. It requires the pre-trained task specific model such as textattack/roberta-base-SST-2 from textattack Huggingface portal. The distillation configuration is specified in yaml file i.e. distillation.yaml.

SST-2 task

wget -P examples/text-classification/ http://nlp.stanford.edu/data/glove.6B.zip
unzip examples/text-classification/glove.6B.zip glove.6B.50d.txt -d examples/text-classification/
python examples/text-classification/run_glue_no_trainer_distillation.py \
--task_name sst2 --max_seq_length 64 \
--model_name_or_path textattack/roberta-base-SST-2 \
--do_distillation --per_device_train_batch_size 32 \
--learning_rate 0.02 --num_train_epochs 20 \
--output_dir /path/to/output_dir --config distillation.yaml --seed 5143

Examples of enabling Intel® Neural Compressor

This is a tutorial of how to enable BERT model with Intel® Neural Compressor.

User Code Analysis

Intel® Neural Compressor supports two usages:

  1. User specifies fp32 'model', calibration dataset 'q_dataloader', evaluation dataset "eval_dataloader" and metrics in tuning.metrics field of model-specific yaml config file.
  2. User specifies fp32 'model', calibration dataset 'q_dataloader' and a custom "eval_func" which encapsulates the evaluation dataset and metrics by itself.

As BERT's matricses are 'f1', 'acc_and_f1', mcc', 'spearmanr', 'acc', so customer should provide evaluation function 'eval_func', it's suitable for the second use case.

Write Yaml config file

In examples directory, there is conf.yaml. We could remove most of the items and only keep mandatory item for tuning.

model:
  name: bert
  framework: pytorch

device: cpu

quantization:
  approach: post_training_dynamic_quant

tuning:
  accuracy_criterion:
    relative: 0.01
  exit_policy:
    timeout: 0
    max_trials: 300
  random_seed: 9527

Here we set accuracy target as tolerating 0.01 relative accuracy loss of baseline. The default tuning strategy is basic strategy. The timeout 0 means early stop as well as a tuning config meet accuracy target.

Note : neural_compressor does NOT support "mse" tuning strategy for pytorch framework

Code Prepare

We just need update run_squad_tune.py and run_glue_tune.py like below

if training_args.tune:
    def eval_func_for_nc(model_tuned):
        trainer = Trainer(
            model=model_tuned,
            args=training_args,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            compute_metrics=compute_metrics,
            tokenizer=tokenizer,
            data_collator=data_collator,
        )
        result = trainer.evaluate(eval_dataset=eval_dataset)
        bert_task_acc_keys = ['eval_f1', 'eval_accuracy', 'mcc', 'spearmanr', 'acc']
        for key in bert_task_acc_keys:
            if key in result.keys():
                logger.info("Finally Eval {}:{}".format(key, result[key]))
                acc = result[key]
                break
        return acc
    from neural_compressor.experimental import Quantization, common
    quantizer = Quantization("./conf.yaml")
    calibration_dataset = quantizer.dataset('bert', dataset=eval_dataset,
                                         task="classifier", model_type=config.model_type)
    quantizer.model = common.Model(model)
    quantizer.calib_dataloader = common.DataLoader(
        calibration_dataset, batch_size=training_args.per_device_eval_batch_size)
    quantizer.eval_func = eval_func_for_nc
    q_model = quantizer()
    q_model.save(training_args.tuned_checkpoint)
    exit(0)

For seq2seq task,We need update run_seq2seq_tune.py like below

if training_args.tune:
    def eval_func_for_nc(model):
        trainer.model = model
        results = trainer.evaluate(
            eval_dataset=eval_dataset,metric_key_prefix="val", max_length=data_args.val_max_target_length, num_beams=data_args.eval_beams
        )
        assert data_args.task.startswith("summarization") or data_args.task.startswith("translation") , \
            "data_args.task should startswith summarization or translation"
        task_metrics_keys = ['val_bleu','val_rouge1','val_rouge2','val_rougeL','val_rougeLsum']
        for key in task_metrics_keys:
            if key in results.keys():
                logger.info("Finally Eval {}:{}".format(key, results[key]))
                if 'bleu' in key:
                    acc = results[key]
                    break
                if 'rouge' in key:
                    acc = sum([v for k,v in results.items() if "rouge" in k])/4
                    break
        return acc
    from neural_compressor.experimental import Quantization, common
    quantizer = Quantization("./conf.yaml")
    quantizer.model = common.Model(model)
    quantizer.calib_dataloader = common.DataLoader(
                                            eval_dataset, 
                                            batch_size=training_args.eval_batch_size,
                                            collate_fn=Seq2SeqDataCollator_nc(tokenizer, data_args, training_args.tpu_num_cores)
                                            )
    quantizer.eval_func = eval_func_for_nc
    q_model = quantizer()
    q_model.save(training_args.tuned_checkpoint)
    exit(0)

For language modeling task,We need update run_clm_tune.py like below

if training_args.tune:
    def eval_func_for_nc(model_tuned):
        trainer = Trainer(
            model=model_tuned,
            args=training_args,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            tokenizer=tokenizer,
            data_collator=default_data_collator,
        )
        eval_output = trainer.evaluate(eval_dataset=eval_dataset)
        perplexity = math.exp(eval_output["eval_loss"])
        results = {"perplexity":perplexity,"eval_loss":eval_output["eval_loss"],\
                    "eval_samples_per_second":eval_output['eval_samples_per_second']}
        clm_task_metrics_keys = ["perplexity"]
        for key in clm_task_metrics_keys:
            if key in results.keys():
                logger.info("Finally Eval {}:{}".format(key, results[key]))
                if key=="perplexity":
                    perplexity = results[key]
                    break
        return 100-perplexity
    from neural_compressor.experimental import Quantization, common
    quantizer = Quantization("./conf.yaml")
    quantizer.model = common.Model(model)
    quantizer.calib_dataloader = common.DataLoader(
                                            eval_dataset, 
                                            batch_size=training_args.eval_batch_size,
                                            collate_fn=default_data_collator_nc
                                            )
    quantizer.eval_func = eval_func_for_nc
    q_model = quantizer()
    q_model.save(training_args.tuned_checkpoint)
    exit(0)