Part 7: Summary

Now you know how to prepare training data, create a model, train that model, and load it back in for testing. To get set up as quickly as possible on new projects, I’ve created a full project template in TensorPlates.

Creating a New Project

To start a new project simply do the following:

git clone git@github.com:crosleythomas/tensorplates.git
cp -r tensorplates/project <your project directory>

Tip: After creating a new project you can make your life easier by creating an alias inside your .bash_profile to activate the associate virtual environment and cd to that directory. For example, say you have a project called flowers and have created a virtual environment called env inside the project directory.

##################################
###   Inside ~/.bash_profile   ###
##################################
alias flowers='cd ~/Documents/projects/flowers && source env/bin/activate'

Project Implementation Checklist

Now, you can step through completing your project with this checklist:

• Setup
  Copy project structure
  Gather new dataset
  Create or activate virtual environment
  
• Data Preparation
  Load data
  Define output file(s) structure
  Define parse_fn
  
• Data Loading
  Get paths to TFRecords
  Define Dataset’s map parse function
  
• Model
  Define any additional config parameters
  Make sure runtime parameters make sense for your current system (i.e. num_parallel_calls in dataset.map, buffer_size in dataset.shuffle )
  Implement model_fn
  Define params
  
• Training
  
  • Training/Evaluation loop is already defined is the template
    You should only need to define params.train_epochs
    
• Evaluation
  Mount model_dir if running remotely
  Open TensorBoard and Watch
  
• Deploy
  This one is completely up to you!
  

Continue Reading

Continue to Part 8 Part 8 cites the tutorials I found most helpful and other resources you may want to use.