ONNX 模型库
返回模型

说明文档

可复现性

# 1. 安装 nncf
# 检出 nncf 9c2845eeb38b4ab1b6d4ca19e31a1886e5bdf17c
# 补丁 b/nncf/torch/sparsity/magnitude/algo.py
 
    def sparsify_params(self):
        from collections import OrderedDict
        sparse_sd = OrderedDict()
        with torch.no_grad():    
            for sparse_info in self.sparsified_module_info:
                for n, m in self.model.named_modules():
                    if m == sparse_info.module:
                        sparse_sd[n+'.weight'] = m.weight*sparse_info.operand.binary_mask

        model_sd = self.model.state_dict()
        for k, v in sparse_sd.items():
            assert k in model_sd, \"key not exists!\"
            model_sd[k] = sparse_sd[k]
        self.model.load_state_dict(model_sd)


# 2. transformers 分支
git clone https://github.com/vuiseng9/transformers
cd transformers && git checkout gen-hybrid-sparse

# 3. 参考
# transformers/examples/pytorch/question-answering/gen-hybrid-sparse/vscode-launch.json

关键内容

gen_hybrid_sparse
├── 90pc_sparse-02_head-0512_ffnn
│   ├── 90pc_sparse-02_head-0512_ffnn-8bit.onnx
│   ├── ir
│   │   ├── 90pc_sparse-02_head-0512_ffnn-8bit.bin
│   │   ├── 90pc_sparse-02_head-0512_ffnn-8bit.mapping
│   │   ├── 90pc_sparse-02_head-0512_ffnn-8bit.xml
├── 90pc_sparse-04_head-1024_ffnn
│   ├── 90pc_sparse-04_head-1024_ffnn-8bit.onnx
│   ├── ir
│   │   ├── 90pc_sparse-04_head-1024_ffnn-8bit.bin
│   │   ├── 90pc_sparse-04_head-1024_ffnn-8bit.mapping
│   │   ├── 90pc_sparse-04_head-1024_ffnn-8bit.xml
├── 90pc_sparse-06_head-1536_ffnn
│   ├── 90pc_sparse-06_head-1536_ffnn-8bit.onnx
│   ├── ir
│   │   ├── 90pc_sparse-06_head-1536_ffnn-8bit.bin
│   │   ├── 90pc_sparse-06_head-1536_ffnn-8bit.mapping
│   │   ├── 90pc_sparse-06_head-1536_ffnn-8bit.xml
├── 90pc_sparse-08_head-2048_ffnn
│   ├── 90pc_sparse-08_head-2048_ffnn-8bit.onnx
│   ├── ir
│   │   ├── 90pc_sparse-08_head-2048_ffnn-8bit.bin
│   │   ├── 90pc_sparse-08_head-2048_ffnn-8bit.mapping
│   │   ├── 90pc_sparse-08_head-2048_ffnn-8bit.xml
└── 90pc_unstructured_sparse
    ├── 90pc_unstructured_sparse-8bit.onnx
    └── ir
        ├── 90pc_unstructured_sparse-8bit.bin
        ├── 90pc_unstructured_sparse-8bit.mapping
        ├── 90pc_unstructured_sparse-8bit.xml

vuiseng9/bert-large-hybrid-sparse-onnx-ir

作者 vuiseng9

↓ 0 ♥ 0

创建时间: 2022-03-02 23:29:05+00:00

更新时间: 2022-02-17 07:39:00+00:00

在 Hugging Face 上查看

文件 (72)

.gitattributes
README.md
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/90pc_sparse-02_head-0512_ffnn-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/90pc_sparse-02_head-0512_ffnn-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/90pc_sparse-02_head-0512_ffnn-8bit.onnx ONNX
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/config.json
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/ir/90pc_sparse-02_head-0512_ffnn-8bit.bin
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/ir/90pc_sparse-02_head-0512_ffnn-8bit.mapping
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/ir/90pc_sparse-02_head-0512_ffnn-8bit.xml
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/ir/mo.log
gen_hybrid_sparse/90pc_sparse-02_head-0512_ffnn/pytorch_model.bin
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/90pc_sparse-04_head-1024_ffnn-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/90pc_sparse-04_head-1024_ffnn-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/90pc_sparse-04_head-1024_ffnn-8bit.onnx ONNX
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/config.json
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/ir/90pc_sparse-04_head-1024_ffnn-8bit.bin
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/ir/90pc_sparse-04_head-1024_ffnn-8bit.mapping
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/ir/90pc_sparse-04_head-1024_ffnn-8bit.xml
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/ir/mo.log
gen_hybrid_sparse/90pc_sparse-04_head-1024_ffnn/pytorch_model.bin
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/90pc_sparse-06_head-1536_ffnn-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/90pc_sparse-06_head-1536_ffnn-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/90pc_sparse-06_head-1536_ffnn-8bit.onnx ONNX
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/config.json
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/ir/90pc_sparse-06_head-1536_ffnn-8bit.bin
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/ir/90pc_sparse-06_head-1536_ffnn-8bit.mapping
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/ir/90pc_sparse-06_head-1536_ffnn-8bit.xml
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/ir/mo.log
gen_hybrid_sparse/90pc_sparse-06_head-1536_ffnn/pytorch_model.bin
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/90pc_sparse-08_head-2048_ffnn-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/90pc_sparse-08_head-2048_ffnn-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/90pc_sparse-08_head-2048_ffnn-8bit.onnx ONNX
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/config.json
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/ir/90pc_sparse-08_head-2048_ffnn-8bit.bin
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/ir/90pc_sparse-08_head-2048_ffnn-8bit.mapping
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/ir/90pc_sparse-08_head-2048_ffnn-8bit.xml
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/ir/mo.log
gen_hybrid_sparse/90pc_sparse-08_head-2048_ffnn/pytorch_model.bin
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/90pc_sparse-10_head-2560_ffnn-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/90pc_sparse-10_head-2560_ffnn-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/90pc_sparse-10_head-2560_ffnn-8bit.onnx ONNX
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/config.json
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/ir/90pc_sparse-10_head-2560_ffnn-8bit.bin
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/ir/90pc_sparse-10_head-2560_ffnn-8bit.mapping
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/ir/90pc_sparse-10_head-2560_ffnn-8bit.xml
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/ir/mo.log
gen_hybrid_sparse/90pc_sparse-10_head-2560_ffnn/pytorch_model.bin
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/90pc_sparse-12_head-3072_ffnn-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/90pc_sparse-12_head-3072_ffnn-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/90pc_sparse-12_head-3072_ffnn-8bit.onnx ONNX
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/config.json
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/ir/90pc_sparse-12_head-3072_ffnn-8bit.bin
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/ir/90pc_sparse-12_head-3072_ffnn-8bit.mapping
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/ir/90pc_sparse-12_head-3072_ffnn-8bit.xml
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/ir/mo.log
gen_hybrid_sparse/90pc_sparse-12_head-3072_ffnn/pytorch_model.bin
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/90pc_sparse-14_head-3584_ffnn-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/90pc_sparse-14_head-3584_ffnn-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/90pc_sparse-14_head-3584_ffnn-8bit.onnx ONNX
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/config.json
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/ir/90pc_sparse-14_head-3584_ffnn-8bit.bin
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/ir/90pc_sparse-14_head-3584_ffnn-8bit.mapping
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/ir/90pc_sparse-14_head-3584_ffnn-8bit.xml
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/ir/mo.log
gen_hybrid_sparse/90pc_sparse-14_head-3584_ffnn/pytorch_model.bin
gen_hybrid_sparse/90pc_unstructured_sparse/90pc_unstructured_sparse-8bit-onnx_sparsity.csv
gen_hybrid_sparse/90pc_unstructured_sparse/90pc_unstructured_sparse-8bit-onnx_sparsity.md
gen_hybrid_sparse/90pc_unstructured_sparse/90pc_unstructured_sparse-8bit.onnx ONNX
gen_hybrid_sparse/90pc_unstructured_sparse/ir/90pc_unstructured_sparse-8bit.bin
gen_hybrid_sparse/90pc_unstructured_sparse/ir/90pc_unstructured_sparse-8bit.mapping
gen_hybrid_sparse/90pc_unstructured_sparse/ir/90pc_unstructured_sparse-8bit.xml
gen_hybrid_sparse/90pc_unstructured_sparse/ir/mo.log