Training using a customized TF Estimator

This is a project using a customized TF Estimator to train and test some TFRecords data. The model models and training pipeline for click-through rate prediction are customized.

Dataset

The data comes from the text files of the standardized data format of iPinYou RTB dataset. Using this project, text files are transformed into TFRecords file.

Data information:

feature = click,          shape = (10000, 1),  Unique count = 2,   min = 0,   max = 1
feature = weekday,        shape = (10000, 1),  Unique count = 1,   min = 4,   max = 4
feature = region,         shape = (10000, 1),  Unique count = 35,  min = 0,   max = 395
feature = city,           shape = (10000, 1),  Unique count = 359, min = 0,   max = 399
feature = adexchange,     shape = (10000, 1),  Unique count = 3,   min = 1,   max = 3
feature = slotformat,     shape = (10000, 1),  Unique count = 2,   min = 0,   max = 1
feature = hour,           shape = (10000, 1),  Unique count = 1,   min = 0,   max = 0
feature = slotwidth,      shape = (10000, 1),  Unique count = 6,   min = 160, max = 1000
feature = slotheight,     shape = (10000, 1),  Unique count = 4,   min = 90,  max = 600
feature = slotvisibility, shape = (10000, 1),  Unique count = 4,   min = 0,   max = 255
feature = slotprice,      shape = (10000, 1),  Unique count = 46,  min = 0,   max = 280
feature = usertag,        shape = (10000, 39), Unique count = 45,  min = -1,  max = 16706

How to create the environment

1. Using pipenv

• To create or activate a virtual env: pipenv shell

  • Install all required packages:
    • install packages exactly as specified in Pipfile.lock: pipenv sync
    • install using the Pipfile, including the dev packages: pipenv install --dev

2. Using docker

• Build the image: make build
• Create a container: docker run -it train_custom_tfestimator:master bash

How to run the program

• Run the main program: make train
• Tuning the model:
  • Go inside the folder cd tune/.
  • Edit the configurations in search_space.json and tune_config.yml.
  • Run the tuner: nnictl create --config tune_config.yml --port <PORT_NUMBER>