CHYangzzz / Amazon-DIN-TFrecord-estimator

Two-Tower Model and DIN model(without Dice) use TF-estimator API at Amazon Electronics dataset

batch_size	Model	max-AUC
32	Two-Tower	0.877
32	DIN(without Dice)	0.893

Requirements

• Python 3.6
• Numpy 1.18.5
• Pandas 1.1.3
• TensorFlow 2.3.1

Amazon Electronics dataset download

wget -c http://snap.stanford.edu/data/amazon/productGraph/categoryFiles/reviews_Electronics_5.json.gz  
gzip -d reviews_Electronics_5.json.gz  
wget -c http://snap.stanford.edu/data/amazon/productGraph/categoryFiles/meta_Electronics.json.gz  
gzip -d meta_Electronics.json.gz

Training and Evaluation

• Step 1: generate tfrecord dataset

python generate_tfrecord.py

or use spark to generate tfrecord
(I use zeppelin and code save in generate_tfrecord.scala)

• Step 2: training and evaluation

python main.py

you need confirm tfrecord dataset path and param "data_gen_method"("spark" or "python")

• you can change Two-Tower model to DIN model in main.py's model_fn

Reference:

https://github.com/zhougr1993/DeepInterestNetwork.git