返回模型
说明文档
使用 JPQD 压缩的 BERT-base(正则化因子 0.03)
F1: 87.66
EM: 80.23
重要文件说明
├── r0.030-squad-bert-b-mvmt-8bit
│ ├── 8bit_ref_bert_squad_nncf_mvmt.json (配合 ssbs-feb 分支使用的 nncf 配置)
│ ├── checkpoint-110000 (用于生成的训练检查点)
│ ├── ir
│ │ ├── sparsity_structures.csv
│ │ ├── sparsity_structures.md (分层稀疏度报告,仅针对 transformer 块中的线性层)
│ │ ├── sparsity_structures.pkl (包含被剪枝的结构 ID,例如 MHSA 中的特定注意力头或 FFN 中的维度,用于调试)
│ │ └── squad-BertForQuestionAnswering.cropped.8bit.xml (自定义丢弃剪枝维度并导出 onnx,随后进行 ir 转换
│ ├── ir_uncropped
│ │ ├── mo-pruned-ir
│ │ │ ├── mo.log (可在此处查看 Model Optimizer 版本)
│ │ │ └── squad-BertForQuestionAnswering.8bit.xml (使用 Model Optimizer --transform=Pruning 移除剪枝结构)
│ │ └── squad-BertForQuestionAnswering.8bit.xml (剪枝结构仅被稀疏化/置零)
vuiseng9/bert-base-squad-jpqd-r0.030
作者 vuiseng9
↓ 0
♥ 0
创建时间: 2022-03-17 04:56:32+00:00
更新时间: 2022-11-15 17:56:26+00:00
在 Hugging Face 上查看文件 (34)
.gitattributes
README.md
r0.030-squad-bert-b-mvmt-8bit/8bit_ref_bert_squad_nncf_mvmt.json
r0.030-squad-bert-b-mvmt-8bit/all_results.json
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/config.json
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/nncf-mvmt-p3.json
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/pytorch_model.bin
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/rng_state.pth
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/scheduler.pt
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/special_tokens_map.json
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/tokenizer.json
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/tokenizer_config.json
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/training_args.bin
r0.030-squad-bert-b-mvmt-8bit/checkpoint-110000/vocab.txt
r0.030-squad-bert-b-mvmt-8bit/compressed_graph.dot
r0.030-squad-bert-b-mvmt-8bit/eval_nbest_predictions.json
r0.030-squad-bert-b-mvmt-8bit/eval_predictions.json
r0.030-squad-bert-b-mvmt-8bit/eval_results.json
r0.030-squad-bert-b-mvmt-8bit/ir/sparsity_structures.csv
r0.030-squad-bert-b-mvmt-8bit/ir/sparsity_structures.md
r0.030-squad-bert-b-mvmt-8bit/ir/sparsity_structures.pkl
r0.030-squad-bert-b-mvmt-8bit/ir/squad-BertForQuestionAnswering.cropped.8bit.bin
r0.030-squad-bert-b-mvmt-8bit/ir/squad-BertForQuestionAnswering.cropped.8bit.mapping
r0.030-squad-bert-b-mvmt-8bit/ir/squad-BertForQuestionAnswering.cropped.8bit.onnx
ONNX
r0.030-squad-bert-b-mvmt-8bit/ir/squad-BertForQuestionAnswering.cropped.8bit.xml
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/mo-pruned-ir/mo.log
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/mo-pruned-ir/squad-BertForQuestionAnswering.8bit.bin
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/mo-pruned-ir/squad-BertForQuestionAnswering.8bit.mapping
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/mo-pruned-ir/squad-BertForQuestionAnswering.8bit.xml
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/squad-BertForQuestionAnswering.8bit.bin
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/squad-BertForQuestionAnswering.8bit.mapping
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/squad-BertForQuestionAnswering.8bit.onnx
ONNX
r0.030-squad-bert-b-mvmt-8bit/ir_uncropped/squad-BertForQuestionAnswering.8bit.xml
r0.030-squad-bert-b-mvmt-8bit/original_graph.dot