mirror of
https://github.com/bulletphysics/bullet3
synced 2024-12-14 22:00:05 +00:00
851ca5bfb3
Use python -m pybullet_envs.examples.testEnv --env AntBulletEnv-v0 --render=1 --steps 1000 --resetbenchmark=1 Added environments: HumanoidFlagrunBulletEnv-v0, HumanoidFlagrunHarderBulletEnv-v0, StrikerBulletEnv-v0, ThrowerBulletEnv-v0, PusherBulletEnv-v0, ReacherBulletEnv-v0, CartPoleBulletEnv-v0 and register them to OpenAI Gym. Allow numpy/humanoid_running.py to use abtch or non-batch update (setJointMotorControl2/setJointMotorControlArray)
112 lines
4.2 KiB
Python
112 lines
4.2 KiB
Python
#!/usr/bin/env python
|
|
|
|
import argparse # parse input arguments
|
|
import numpy as np # arithmetic library
|
|
import time
|
|
import gym
|
|
from agents import agent_register
|
|
import pybullet as p
|
|
import kerasrl_utils
|
|
|
|
np.set_printoptions(precision=3, suppress=True, linewidth=10000)
|
|
|
|
|
|
def add_opts(parser):
|
|
parser.add_argument('--agent', type=str, default="KerasDQNAgent", help="Agent to be trained with.")
|
|
parser.add_argument('--env', type=str, default="2DDetachedCartPolev0Env", help="Environment to be trained in.")
|
|
parser.add_argument('--use-latest', action='store_true', help="Should the trainer retrain/show with the most recent save?")
|
|
parser.add_argument('--train-for', type=int, default=100, help="The number of epochs to train for.")
|
|
parser.add_argument('--test-for', type=int, default=0, help="The number of epoch to test for.")
|
|
parser.add_argument('--load-file', type=str, default=None, help="The weight file to load for training.")
|
|
parser.add_argument('--save-file', type=str, default=None, help="The weight file to save after training.")
|
|
|
|
|
|
class Trainer:
|
|
'''
|
|
The trainer class helps to easily set up a gym training session using an agent(representing the learning algorithm and the gym (being the environment)
|
|
'''
|
|
# TODO: Make training fail-safe by catching "not connected to server" and save the current state to disk (see primitive examples, they can do it)
|
|
|
|
def __init__(self):
|
|
# initialize random seed
|
|
np.random.seed(int(time.time()))
|
|
|
|
cid = p.connect(p.SHARED_MEMORY) # only show graphics if the browser is already running....
|
|
self.visualize = (cid >= 0)
|
|
if cid < 0:
|
|
cid = p.connect(p.DIRECT) # If no shared memory browser is active, we switch to headless mode (DIRECT is much faster)
|
|
|
|
def setup_exercise(self, opts):
|
|
|
|
# setup agent
|
|
agent = agent_register.make(opts.agent, opts=opts)
|
|
|
|
# setup environment
|
|
env = gym.make(opts.env)
|
|
if self.visualize:
|
|
mode = "human"
|
|
else:
|
|
mode = "none"
|
|
env.render(mode=mode)
|
|
|
|
# configurations
|
|
env.seed(int(time.time()))
|
|
#env.configureActions(agent.metadata['discrete_actions']) # configure environment to accepts discrete actions
|
|
if agent.metadata['discrete_actions']:
|
|
agent.configure(env.observation_space.shape, env.action_space.n) # configure agent to use the environment properties
|
|
else:
|
|
agent.configure(env.observation_space.shape, env.action_space.shape[0]) # configure agent to use the environment properties
|
|
|
|
if opts.use_latest:
|
|
properties = kerasrl_utils.get_latest_save("checkpoints/", opts.agent, opts.env, 0)
|
|
if properties == []:
|
|
print("No previous weight saves found for %s-%s" % (opts.agent, opts.env))
|
|
else:
|
|
opts.load_file = "checkpoints/%s-%s-%s.h5" % (properties[0], properties[1], properties[2])
|
|
print("Continue from [%s] " % opts.load_file)
|
|
|
|
if opts.load_file is not None:
|
|
print("loading weights from [%s]" % opts.load_file)
|
|
agent.load_weights(opts.load_file)
|
|
|
|
# Okay, now it's time to learn something! We visualize the training here for show, but this
|
|
# slows down training quite a lot. You can always safely abort the training prematurely using
|
|
# Ctrl + C.
|
|
agent.train(env, nb_steps=opts.train_for, visualize=self.visualize, verbosity=1)
|
|
|
|
# After training is done, we save the final weights.
|
|
if opts.save_file is not None:
|
|
print("saving weights to [%s]" % opts.save_file)
|
|
agent.save_weights(opts.save_file, overwrite=True)
|
|
|
|
# Finally, evaluate our algorithm.
|
|
agent.test(env, nb_episodes=opts.test_for, visualize=self.visualize)
|
|
|
|
if __name__ == "__main__":
|
|
"""
|
|
You can also run the trainer as a main class if you want to start your own agent/environment combination. If you know your precise arguments, just run this as your main.
|
|
"""
|
|
|
|
trainer = Trainer.Trainer()
|
|
|
|
parser = argparse.ArgumentParser()
|
|
|
|
# add all parsing options
|
|
Trainer.add_opts(parser)
|
|
|
|
opts, unknown = parser.parse_known_args() # parse agent and environment to add their opts
|
|
|
|
exec("from agents import %s" % opts.agent) # import agent type
|
|
exec("from envs import %s" % opts.env) # import env type
|
|
exec("%s.add_opts(parser)" % opts.agent)
|
|
exec("%s.add_opts(parser)" % opts.env)
|
|
|
|
# parse arguments
|
|
opts, unknown = parser.parse_known_args()
|
|
print("OPTS", opts)
|
|
print("UNKNOWN", unknown)
|
|
|
|
|
|
trainer.setup_exercise(opts)
|
|
|