You can easily add a new policy model for a specific morphology to visualize its behaviour inside the demo.
1. Train a policy
You first need a saved policy model corresponding to one of the available morphologies. If you want to train a new policy on your own, follow the installation and launching steps of the TeachMyAgent repository.
This policy model must be in the TensorFlow SavedModel format and organized as follows:
📂 policy_folder
┣ 📂 tf1_save
┃ ┣ 📂 variables
┃ ┃ ┣ 📜 variables.data-00000-of-00001
┃ ┃ ┗ 📜 variables.index
┃ ┣ 📜 model_info.plk
┃ ┗ 📜 saved_model.pb
┗ 📜 name.txt -- contains the name and age of the policy
The name and age of the policy in name.txt must be separated by a / and the age must be "adult", "teenager" or "baby".
2. Add the policy to the demo
Then you just need to add this policy_folder to the corresponding morphology folder in policy_models among the following:
📂 policy_models
┣ 📂 climber
┃ ┗ 📂 chimpanzee
┣ 📂 swimmer
┃ ┗ 📂 fish
┣ 📂 walker
┃ ┣ 📂 bipedal
┃ ┗ 📂 spider
Your policy model will now automatically appear inside the demo, in the list of agents available for the corresponding morpholgy.
3. [Optional] Set up for local launch
3.1. Convert your policy model to a web-friendly format
ls -d policy_models/<type>/<morphology>/<policy_folder>/ | xargs -I"{}" tensorflowjs_converter --input_format=tf_saved_model --saved_model_tags=serve --skip_op_check {}tf1_save web_demo/{}
3.2. Generate the list of policy models
python3 policies_to_json.py