Tensorflow Project Template

This repo is forked from Tensorflow-Project-Template. Based on their setting, I add the following features:

Features

Base Config

I use built-in Configuration rather than JSON file to record variables. For example, the exp_name could be used in constructing model_dir, save_path, model_name or anything. It is inconvenient and obvious to define these in the code. But, JSON file can not support this kind of reference.

Note that we could also use shell to achieve this.

Base Dataset

I write a container for small dataset that can fit in memory. Simply, the dataset instance is built with X,y. Then, we can access its data directly or call next batch function.

Remember to add the dataset in data_loader.py.

Base Model

I already built saver, global step and other repetitous ops used during training. Mostly, we need only define X and Y and overwrite build_model function

Base Runner

I did the things about sess and variables initialization. We only need to specific the training process, including logging and so on.

Run

Train a demo mnist example.

python -m mains.train -it -c configs.example -m models.example -d datasets.mnist

Test

python -m mains.test -c configs.example -m models.example -d datasets.mnist

Args:

• -it: is training
• -c: config file
• -m: model file
• -d: data file

TODO

• Base config
• Base dataset
• Base model
• Base runner
• A toy example
• Doc
• Early Stopping Mechanism
• Large Dataset Loader