CpRec

KO | EN

gitlite — search

#javascript #python #ai #android #deep-learning #linux #nodejs #java #windows #awesome #vue #pytorch #security #openai

(★ 20)

A Generic Network Compression Framework for Sequential Recommender Systems

CpRec Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

You can email me if you have any questions about the performance and efficiency of CpRec.

```
you can run nextitrec_eval.py (python nextitrec_eval.py) directly, which includes training and testing.  
Each training epoch, test once test set.  
```

## The import configuration:  
```
--softmax_type: This is used to control whether block-wise embedding is used. There are 5 classes:  
FullSoftmax：The input layer and output layer use basic embedding and softmax respectively.  
Block_Input_Full：Only the input layer uses block-wise embedding, and the output layer uses basic softmax.  
Block_for_Softmax：Only the output layer uses block-wise embedding, and the output layer uses basic softmax.  
Block_Input_Softmax：Both the input layer and the output layer use block-wise embedding.  
Block_Input_Softmax_Inference： Both the input layer and the output layer use block-wise embedding, but fast inference is used in the prediction stage. it should be noted that the use of this type requires batch_size_test = 1  
```
```
--use_parametersharing_type: This is used to control whether to use layer-wise parameter sharing, there are 5 classes:  
original：No parameter sharing.  
cross-layer：use the cross-layer parameter sharing.  
cross-block：use the cross-block parameter sharing.  
adjacent-layer：use the adjacent-layer parameter sharing.  
adjacent-block：use the adjacent-block parameter sharing.  
```
you should cross-valid the number of blocks and the embedding size of each block, but we strongly suggest you use 3~5 blocks.  
--use_embedding_type_factor is the attenuation factor of block dimension in block-wise embedding, the default is 4. You can also manually set the dimension of each block, e.g. [300, 200, 100].

## Dataset
you can download a large sequential dataset of movielen that has been pre-processed: https://pan.baidu.com/s/1XZCUoRwWFa8fpRWpEA2HPQ  
code (提取码):fo51

# References
https://github.com/fajieyuan/nextitnet code that including Caser and GRURec  
https://github.com/graytowne/caser code

# CpRec
A Generic Network Compression Framework for Sequential Recommender Systems  
```
@article{sun2020generic,
  title={A Generic Network Compression Framework for Sequential Recommender Systems},
  author={Sun, Yang and Yuan, Fajie and Yang, Min and Wei, Guoao and Zhao, Zhou and Liu, Duo},
  journal={arXiv preprint arXiv:2004.13139},
  year={2020}
}
```
Please cite this paper if you find our code is useful.

CpRec

CpRec Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

CpRec

CpRec Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README