Visualize Training Results

1

Is there any way to visualize results from training without using Tensorboard? I’m working on a VM and it doesn’t allow us to use a browser. I know we can just save the results as a dataframe and plot those, but wouldn’t that be an arduous process (plotting 50,000 episodes’ worth of results). Or would that be my only other option?

2

Hey @Yared_Kokeb I’m assuming you’re using Ray Tune or RLlib, correct? If so, I would highly recommend using Weights & Biases. It’s free to use, provides excellent visualizations, and you can log all your results and metrics, and then access it on your account from any device. The Ray libraries also integrate with it really well: Using Weights & Biases with Tune — Ray v1.5.2.

Let me know if this is what you’re looking for!

1 Like
4

Is there a way to integrate wandb to training classes? I only see integration for training functions, and when I add the ‘wandb’ configuration to my tune.run it just gives me an error.

5

For reference, here is my code (copied the cartpole class from github):

ray.init()
env_config = {
    "config_file": trainer_config_file,
    "mpc_dict": mpc_dict
}
config ={
     "env": CartPoleEnv,
     # "env_config": env_config,
     "num_gpus": 0,
     "num_workers": 1,
     "lr": 1e-3,
     "framework": "torch",
     "wandb": {"project": "Optimization_Project"}
}
alg = input("Please enter the model used for this training: ")
date = datetime.now().strftime("%d-%m-%Y_%H:%M:%S")
results = tune.run(
    alg, 
    stop={"training_iteration": 10}, 
    config=config, 
    local_dir="results/test_viz/{}".format(date),
    loggers = DEFAULT_LOGGERS + (WandbLogger,)
)
ray.shutdown()
6

Nevermind, resolved!