def fn():
import pickle
fn = 'env_data.npy'
l = []
with open(fn, 'rb') as f:
l = pickle.load(f)
env = PalleteWorld(env_id=env_id, datasets=l)
env.seed(seed + rank)
return env
def __init__(self, seed=None, id=0):
self.id = id
# import pickle
# l = []
# with open('env_data.npy', 'rb') as f:
# l = pickle.load(f)
# self.env = PalleteWorld(datasets=l)
self.env = PalleteWorld(n_random_fixed=1, env_id=id)
self.episode_step = 0
if seed is not None:
self.env.seed(seed)
print('env {} is created with seed {}'.format(self.id, seed))
def __init__(self, num_env):
self.envs = []
for _ in range(num_env):
self.envs.append(PalleteWorld(n_random_fixed=10))
from environment import PalleteWorld
import pickle
from sample_agent.reinforce_agg import *
fn = 'env_data.npy'
# env = PalleteWorld(env_id=0, n_random_fixed=5)
# env.seed(1)
# env.save_dataset(fn)
#
fn = 'env_data.npy'
l = []
with open(fn, 'rb') as f:
l = pickle.load(f)
for i, items in enumerate(l):
env = PalleteWorld(datasets=[items], env_id=i)
main(env)
def fn():
env = PalleteWorld(env_id=env_id, n_random_fixed=10)
env.seed(seed + rank)
return env
if env_cfg.ENV.RENDER:
env.render()
clock.tick(env_cfg.ENV.TICK_INTERVAL)
policy.rewards.append(reward)
ep_reward += reward
if done:
# print('episode end')
# print('len of action : {}, episode actions : {}'.format(len(env.previous_actions), env.previous_actions))
ep_rewards.append(ep_reward)
n_placed_items = len(info['placed_items'])
break
finish_episode()
if i_episode % 10 == 0:
print('Episode {}\tLast reward: {:.2f}'.format(i_episode, ep_reward))
writer.add_scalar('data/final_reward', ep_reward, i_episode)
writer.add_scalar('data/n_placed_items', n_placed_items, i_episode)
def checkEqual(iterator):
return len(set(iterator)) <= 1
if checkEqual(ep_rewards[-20:]) and len(ep_rewards) > 100:
print("Solved! last reward is {}".format(ep_rewards[-1]))
return ep_rewards[-1]
if __name__ == '__main__':
env = PalleteWorld(n_random_fixed=1)
main()
class Game:
"""
Game wrapper.
"""
def __init__(self, seed=None, id=0):
self.id = id
# import pickle
# l = []
# with open('env_data.npy', 'rb') as f:
# l = pickle.load(f)
# self.env = PalleteWorld(datasets=l)
self.env = PalleteWorld(n_random_fixed=1, env_id=id)
self.episode_step = 0
if seed is not None:
self.env.seed(seed)
print('env {} is created with seed {}'.format(self.id, seed))
def step(self, action):
"""
Apply action to the game.
Args:
action : action of the action_space to take.
Returns:
The new observation, the reward and a boolean if the game has ended.
"""
self.episode_step = self.episode_step + 1
a = env_cfg.Action(bin_index=action, priority=1, rotate=1)
o, reward, done, _ = self.env.step(a)
# this is for visual state representation.
import torch
boxes = torch.zeros(10, 10,
env_cfg.ENV.BIN_MAX_COUNT) # x, y, box count
for i, box in enumerate(o[0]): # box = x,y
boxes[0:box[1], 0:box[0], i] = 1
o = np.concatenate((o[1], boxes), axis=-1)
print('A agent is taking action {} in {} step in game {}.'.format(
action, self.episode_step, self.id))
# this is for visual state representation.
# import torch
# embedding = torch.nn.Embedding(num_embeddings=env_cfg.ENV.BIN_MAX_X_SIZE*env_cfg.ENV.BIN_MAX_X_SIZE+env_cfg.ENV.BIN_MAX_Y_SIZE+1, embedding_dim=1)
# i = [b[0] * env_cfg.ENV.BIN_MAX_X_SIZE + b[1] for b in o[0]]
# boxlist = embedding(torch.Tensor(i).long()).squeeze(1).detach().numpy()
# o = np.concatenate((o[1][:,:,0].flatten(), boxlist))
return o, reward, done
def to_play(self):
"""
Return the current player.
Returns:
The current player, it should be an element of the players list in the config.
"""
return 0
def reset(self):
"""
Reset the game for a new game.
Returns:
Initial observation of the game.
"""
o = self.env.reset()
print('game {} is reset.'.format(self.id))
self.episode_step = 0
# import torch
# embedding = torch.nn.Embedding(num_embeddings=env_cfg.ENV.BIN_MAX_X_SIZE*env_cfg.ENV.BIN_MAX_X_SIZE+env_cfg.ENV.BIN_MAX_Y_SIZE+1, embedding_dim=1)
# i = [b[0] * env_cfg.ENV.BIN_MAX_X_SIZE + b[1] for b in o[0]]
# boxlist = embedding(torch.Tensor(i).long()).squeeze(1).detach().numpy()
# o = np.concatenate((o[1][:,:,0].flatten(), boxlist))
# this is for visual state representation.
import torch
boxes = torch.zeros(10, 10,
env_cfg.ENV.BIN_MAX_COUNT) # x, y, box count
for i, box in enumerate(o[0]): # box = x,y
boxes[0:box[1], 0:box[0], i] = 1
o = np.concatenate((o[1], boxes), axis=-1)
return o
def close(self):
"""
Properly close the game.
"""
self.env.close()
def render(self):
"""
Display the game observation.
"""
self.env.render()
input("Press enter to take a step ")
from env_config import *
import logging
import pygame
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
from tensorboardX import SummaryWriter
LOGLEVEL = os.environ.get('LOGLEVEL', 'INFO').upper()
logging.basicConfig(level=LOGLEVEL)
writer = SummaryWriter()
if __name__ == '__main__':
env = PalleteWorld(n_random_fixed=1)
agent = BDQNAgent()
if ENV.RENDER:
clock = pygame.time.Clock()
total_step = 0
for e in range(ENV.N_EPISODES):
state = env.reset()
d = False
step = 0
while True:
while True:
action = agent.get_action(state)