def main():
########################################################################
"""
Initialize variables.
"""
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#device = 'cpu'
print("The device is " + str(device))
starting_floor = 0
total_floors = 9
max_steps = 1000
override_threshold = 1000
########################################################################
"""
Retreive the learnt policy.
"""
policy_model = unpickle_object("policy_model") #read the object from disk
########################################################################
"""
Let the agent play a game using the policy.
"""
env = create_env(starting_floor=starting_floor, total_floors=total_floors)
imitation_play(env,
policy_model,
device,
max_steps=max_steps,
override_threshold=override_threshold)
state = next_state
total_reward += reward
steps += 1
if steps >= max_steps_in_demo_episode or done:
break
#====
torch.manual_seed(hyper_params['seed'])
torch.cuda.manual_seed_all(hyper_params['seed'])
np.random.seed(hyper_params['seed'])
random.seed(hyper_params['seed'])
assert "NoFrameskip" in hyper_params[
'env'], "Require environment with no frameskip"
env = create_env(0, 1)
env.seed(hyper_params['seed'])
#env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
#env = EpisodicLifeEnv(env)
#env = FireResetEnv(env)
env = WarpFrame(env)
env = PyTorchFrame(env)
env = ClipRewardEnv(env)
env = FrameStack(env, 3)
replay_buffer = ReplayBuffer(hyper_params['replay_buffer_size'])
agent = DQNAgent(env.observation_space,
env.action_space,
replay_buffer,
#batch update parameters
max_epochs = 5e6 #number of batches to create
batch_size = 500 #number of steps in one batch
mini_batch_size = 250 #how many steps are used to update the loss
ppo_epochs = 2 #number of batches to sample in ppo update || min_batch * ppo_epoch = batch_size recommended
epochs_before_printing = 5 #number of steps at which we print update
#demo information
show_demo = False #whether or not to make a video of the current progress
max_steps_in_demo_episode = 200 #number of steps to show in demo episode
#environement parameters
starting_floor = 0
total_floors = 1
worker_id = 1
env = create_env(starting_floor, total_floors, worker_id)
policy_actions = unpickle_object(
'action_map') #map going grom actions to env actions
override_threshold = 2000 #score used to determine if agent is stuck
#deep learning setup
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = ActorCritic(len(policy_actions)).to(device)
#model = unpickle_object('policy_model')
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
########################################################################
"""
Train the model.
"""
experiment = Experiment(api_key="47QJ41M89a6zNXZgS9sY6NQfI",
project_name="unity",
import ipdb
import matplotlib.pyplot as plt
from collections import deque
import keyboard
import time
import datetime
import random
import numpy as np
from recorder import create_env, record_game, save_file_as_csv
########################################################################
"""
https://discourse.aicrowd.com/t/actions-meaning-in-retro-mode/931
6 - rotate counter-clockwise
12 - rotate clockwise
18 - move forward
3 - jump
22 - jump and move forward - may want to include this
"""
action_map = {'w': 18, 'q': 6, 'e': 12, 'a': 3, 's': 21}
max_steps_to_record = 10e100
starting_floor = 0
total_floors = 10
########################################################################
env = create_env(starting_floor=starting_floor, total_floors=total_floors)
state_hist, action_hist, reward_hist, key_hist = record_game(
env, max_steps_to_record, action_map)
save_file_as_csv(state_hist, action_hist, reward_hist, key_hist)
########################################################################