torchlure 0.2407.1

No project description provided

Torch Lure

Depndencies

pip install git+https://github.com/Farama-Foundation/Minari.git@19565bd8cd33f2e4a3a9a8e4db372044b01ea8d3

pip install torchlure

Usage

import torchlure as lure

# Optimizers
lure.SophiaG(lr=1e-3, weight_decay=0.2)

# Functions
lure.tanh_exp(x)
lure.TanhExp()

lure.quantile_loss(y_pred, y_target, quantile=0.5)
lure.QuantileLoss(quantile=0.5)

lure.RMSNrom(dim=256, eps=1e-6)

# Noise Scheduler
lure.LinearNoiseScheduler(beta=1e-4, beta_end=0.02, num_timesteps=1000)
lure.CosineNoiseScheduler(max_beta=0.999, s=0.008, num_timesteps=1000):

Dataset

import gymnasium as gym
import numpy as np
import torch
from torchlure.datasets import MinariEpisodeDataset, MinariTrajectoryDataset

env = gym.make("Hopper-v4")
minari_dataset = MinariEpisodeDataset("Hopper-v4.2407")
minari_dataset.create(env, n_episodes=100)
minari_dataset.info()
# Observation space: Box(-inf, inf, (11,), float64)
# Action space: Box(-1.0, 1.0, (3,), float32)
# Total episodes: 100
# Total steps: 2,182

traj_dataset = MinariTrajectoryDataset(minari_dataset, traj_len=20)

ep = traj_dataset[2]
ep = traj_dataset[[3, 8, 15]]
ep = traj_dataset[np.arange(16)]
ep = traj_dataset[torch.arange(16)]
ep = traj_dataset[-16:]

ep["observations"].shape, ep["actions"].shape, ep["rewards"].shape, ep[
    "terminations"
].shape, ep["truncate"].shape
# (torch.Size([16, 20, 4, 4, 16]),
#  torch.Size([16, 20]),
#  torch.Size([16, 20]),
#  torch.Size([16, 20]),
#  torch.Size([16, 20]))