该模型是使用 OpenVINO/NNCF 对 vuiseng9/bert-base-squadv1-block-pruning-hybrid-filled-lt 进行下游优化的结果。应用的优化包括：

1. 初始化时进行60%的幅度稀疏化。参数通过其绝对范数进行全局排序。仅针对自注意力层和前馈神经网络（FFNN）的线性层。
2. NNCF量化感知训练 - 对所有可学习层的权重和激活值进行对称8位量化。
3. 使用大模型 bert-large-uncased-whole-word-masking-finetuned-squad 进行自定义蒸馏

  eval_exact_match = 80.3122
  eval_f1          = 87.6162
  eval_samples     =   10784

环境配置

# OpenVINO/NNCF
git clone https://github.com/vuiseng9/nncf && cd nncf
git checkout tld-poc
git reset --hard 1dec7afe7a4b567c059fcf287ea2c234980fded2
python setup.py develop
pip install -r examples/torch/requirements.txt

# Huggingface nn_pruning
git clone https://github.com/vuiseng9/nn_pruning && cd nn_pruning
git checkout reproduce-evaluation
git reset --hard 2d4e196d694c465e43e5fbce6c3836d0a60e1446
pip install -e \".[dev]\"

# Huggingface Transformers
git clone https://github.com/vuiseng9/transformers && cd transformers
git checkout tld-poc
git reset --hard 10a1e29d84484e48fd106f58957d9ffc89dc43c5
pip install -e .
head -n 1 examples/pytorch/question-answering/requirements.txt | xargs -i pip install {}

# 其他依赖
pip install onnx

训练

git clone https://huggingface.co/vuiseng9/bert-base-squadv1-block-pruning-hybrid-filled-lt
BASE_MODEL=/path/to/cloned_repo_above #需修改

wget https://huggingface.co/vuiseng9/bert-base-squadv1-block-pruning-hybrid-filled-lt-nncf-60.0sparse-qat-lt/raw/main/nncf_bert_squad_sparsity.json
NNCF_CFG=/path/to/downloaded_nncf_cfg_above #需修改

OUTROOT=/path/to/train_output_root #需修改
WORKDIR=transformers/examples/pytorch/question-answering #需修改
RUNID=bert-base-squadv1-block-pruning-hybrid-filled-lt-nncf-60.0sparse-qat-lt

cd $WORKDIR

OUTDIR=$OUTROOT/$RUNID
mkdir -p $OUTDIR

export CUDA_VISIBLE_DEVICES=0
NEPOCH=5

python run_qa.py \
    --model_name_or_path vuiseng9/bert-base-squadv1-block-pruning-hybrid \
    --optimize_model_before_eval \
    --optimized_checkpoint $BASE_MODEL \
    --dataset_name squad \
    --do_eval \
    --do_train \
    --evaluation_strategy steps \
    --eval_steps 250 \
    --learning_rate 3e-5 \
    --lr_scheduler_type cosine_with_restarts \
    --warmup_ratio 0.25 \
    --cosine_cycles 1 \
    --teacher bert-large-uncased-whole-word-masking-finetuned-squad \
    --teacher_ratio 0.9 \
    --num_train_epochs $NEPOCH \
    --per_device_eval_batch_size 128 \
    --per_device_train_batch_size 16 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --save_steps 250 \
    --nncf_config $NNCF_CFG \
    --logging_steps 1 \
    --overwrite_output_dir \
    --run_name $RUNID \
    --output_dir $OUTDIR

评估

此仓库必须克隆到本地。

git clone https://huggingface.co/vuiseng9/bert-base-squadv1-block-pruning-hybrid-filled-lt-nncf-60.0sparse-qat-lt
MODELROOT=/path/to/cloned_repo_above #需修改

export CUDA_VISIBLE_DEVICES=0

OUTDIR=eval-bert-base-squadv1-block-pruning-hybrid-filled-lt-nncf-60.0sparse-qat-lt
WORKDIR=transformers/examples/pytorch/question-answering #需修改
cd $WORKDIR
mkdir $OUTDIR

nohup python run_qa.py  \
    --model_name_or_path vuiseng9/bert-base-squadv1-block-pruning-hybrid  \
    --dataset_name squad  \
    --optimize_model_before_eval \
    --qat_checkpoint $MODELROOT/checkpoint-22000  \
    --nncf_config $MODELROOT/nncf_bert_squad_sparsity.json  \
    --to_onnx $OUTDIR/bert-base-squadv1-block-pruning-hybrid-filled-lt-nncf-60.0sparse-qat-lt.onnx  \
    --do_eval  \
    --per_device_eval_batch_size 128  \
    --max_seq_length 384  \
    --doc_stride 128  \
    --overwrite_output_dir \
    --output_dir $OUTDIR 2>&1 | tee $OUTDIR/run.log &