MakiPose-Training

This repository contains a ready-to-use code for training a MakiFlow pose estimation model. It does not contain much code and is not complicated, so it is easy to modify to suit one's needs.

Structure

The repository consists of 3 main files:

gen_layer.py

Contains code that creates generator layer that is gonna feed the model with the data. It has several constants that you have to set, such as TFRECORDS_PATH or BATCH_SIZE.

run.py

The actual script to run.

config.json

Configuration file that contains all the info the training process: experiment folder, number of epochs, skeleton configuration, etc.

model.json

The model's architecture file. It is advisable to put the model's architecture this way, however, you are not restricted and can change the path to the architecture file in config.json.

How to use

It is assumed the data has already been prepared.

1. Put the model's architecture file and name it as model.json.
2. Set the configuration file to suit your training needs.
3. Run in console python run.py.
4. Open tensorboard. If you run the script on the local machine, open in the browser: localhost:6006. If the script is being ran on a remote machine, you can access the board through ssh. In command line enter ssh -p PORT USERNAME@REMOTE_MACHINE_ADRESS -N -f -L localhost:16006:localhost:6006 and then open localhost:16006 in the browser.