Basic DeepLearning Models

I build it for study deep learning model with pytorch

Usage

All code started with train.py, we use config file to differentiate the model we used.

Just like: python train -c model_config.json

For your owner useage:

• write your owner dataSet under ./data_loader and add it to ./data_loader/data_loaders.py
• write your owner config file under ./configs to choose model and set parameters
• python train.py -c ./configs/config.json

For each task of your owner, you should build dataloader in ./data_loader and config the json file in ./configs

Sometimes you will get tensor type error between long\float\int, all you need is to change your dataset file __getitem__

Example

For Factorization Machine:

• write criteo_dataset.py under ./data_loader
• add it to ./data_loader/data_loaders.py

class CriteoDataLoader(BaseDataLoader):
    def __init__(self, data_dir, batch_size, shuffle=True, validation_split=0.0, num_workers=1):
        self.data_dir = data_dir
        self.dataset = CriteoDataset(self.data_dir)
        super().__init__(self.dataset, batch_size, shuffle, validation_split, num_workers)

• write config_fm.json under ./configs
• run python train.py -c ./configs/config.json

CTR Models

Model	Reference
Factorization Machine	S Rendle, Factorization Machines, 2010.
Field-aware Factorization Machine	Y Juan, et al. Field-aware Factorization Machines for CTR Prediction, 2015.
DeepFM	H Guo, et al. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction, 2017.
Wide&Deep	HT Cheng, et al. Wide & Deep Learning for Recommender Systems, 2016.
Deep Cross Network	R Wang, et al. Deep & Cross Network for Ad Click Predictions, 2017.
xDeepFM	J Lian, et al. xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems, 2018.

NLP Models

Model	Reference
fastText	Bag of Tricks for Efficient Text Classification
TextCNN	Convolutional Neural Networks for Sentence Classification

DataSet

ModelType	DataSet	Source
CTR Prediction	CriteoDataset	criteo
NLP Classify	ThucnewsDataset	THUCNews

Performance

Model	acc	loss
FM	0.854	0.68
FastText	0.998	0.02
TextCNN	0.954	0.18

Tensorboard

You can also see the tensorboard at localhost:6006 by running tensorboard --logdir='./saved/log/fm'

Reference

Pytorch template based on: pytorch-template

Rec based on：pytorch-fm