def main():
args = PARSER.parse_args()
data_path = get_path(args, "record")
model_save_path = get_path(args, "tf_vae", create=True)
ensure_validation_split(data_path)
_n_train, _avg_frames, mean, var = analyse_dataset(data_path)
if args.normalize_images:
train_data, val_data = create_tf_dataset(data_path, args.z_size, True, mean, var)
else:
train_data, val_data = create_tf_dataset(data_path, args.z_size)
shuffle_size = 5 * 1000 # Roughly 20 full episodes for shuffle windows, more increases RAM usage
train_data = train_data.shuffle(shuffle_size, reshuffle_each_iteration=True).batch(args.vae_batch_size).prefetch(2)
val_data = val_data.batch(args.vae_batch_size).prefetch(2)
current_time = datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_dir = model_save_path / "tensorboard" / current_time
vae = CVAE(args=args)
vae.compile(optimizer=vae.optimizer, loss=vae.get_loss())
vae.fit(train_data, validation_data=val_data, epochs=args.vae_num_epoch, callbacks=[
tf.keras.callbacks.TensorBoard(log_dir=str(tensorboard_dir), update_freq=50, histogram_freq=1),
LogImage(str(tensorboard_dir), val_data),
tf.keras.callbacks.ModelCheckpoint(str(model_save_path / "ckpt-e{epoch:02d}"), verbose=1),
])
vae.save(str(model_save_path))
def main():
args = PARSER.parse_args()
data_path = get_path(args, "record")
model_save_path = get_path(args, "tf_gqn", create=True)
ensure_validation_split(data_path)
train_data = load_from_tfrecord(data_path,
args.gqn_context_size,
args.gqn_batch_size,
mode='train')
test_data = load_from_tfrecord(data_path,
args.gqn_context_size,
args.gqn_batch_size,
mode='test')
current_time = datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_dir = model_save_path / "tensorboard" / current_time
# lr = tf.optimizers.schedules.ExponentialDecay(mu_i, mu_n, mu_f / mu_i, name="lr_schedule" )
lr = tf.optimizers.schedules.PolynomialDecay(mu_i,
mu_n,
mu_f,
name="lr_schedule")
sigma = tf.optimizers.schedules.PolynomialDecay(sigma_i,
sigma_n,
sigma_f,
name="sigma_schedule")
optimizer = tf.optimizers.Adam(learning_rate=lr)
model = GenerativeQueryNetwork(args.gqn_x_dim,
args.gqn_r_dim,
args.gqn_h_dim,
args.gqn_z_dim,
args.gqn_l,
name="gqn")
model.compile(optimizer, sigma, const_sigma=sigma_f)
model.fit(train_data,
validation_data=test_data,
validation_steps=5,
steps_per_epoch=S_epoch,
epochs=num_epochs,
callbacks=[
tf.keras.callbacks.TensorBoard(log_dir=str(tensorboard_dir),
update_freq=20,
histogram_freq=1),
tf.keras.callbacks.ModelCheckpoint(
str(model_save_path / "ckpt-e{epoch:02d}"),
save_freq=checkpoint_every,
verbose=1),
LogImages(tensorboard_dir, test_data),
])
def main():
args = PARSER.parse_args()
data_dir = get_path(args, "series")
train_data_path = data_dir / "series.npz"
val_data_path = data_dir / "series_validation.npz"
train_data = load_data(train_data_path)
validation_data = load_data(val_data_path)
initial_z_dir = get_path(args, "tf_initial_z", create=True)
create_initial_z(initial_z_dir, train_data)
train_dataset, validation_dataset = create_dataset(
train_data, validation_data, args.rnn_batch_size, args.rnn_max_seq_len,
args.z_size, args.rnn_input_seq_width, args.rnn_predict_done,
args.rnn_predict_reward)
train_rnn(args, train_dataset, validation_dataset)
def __init__(self, env, silent=False):
super().__init__(env)
from vae.vae import CVAE
from utils import PARSER
args = PARSER.parse_args(['--config_path', 'configs/carracing.config'])
model_path_name = "models/tf_vae"
self.vae = CVAE(args)
# self.vae.set_weights(tf.keras.models.load_model(
# model_path_name, compile=False).get_weights())
self.vae.set_weights(np.load("vae_weights.npy", allow_pickle=True))
self.observation_space = Box(low=float("-inf"),
high=float("inf"),
shape=(41, ))
self.silent = silent
s, r, done, info = env.step(a)
total_reward += r
if steps % 200 == 0 or done:
print("\naction " + str(["{:+0.2f}".format(x) for x in a]))
print("step {} total_reward {:+0.2f}".format(
steps, total_reward))
steps += 1
env.render()
if done or restart: break
env.monitor.close()
if __name__ == '__main__':
import configargparse
from utils import PARSER
args = PARSER.parse_args()
#true_env = make_env(args, dream_env=False, with_obs=True)
run_caracing_by_hunman()
# from ppaquette_gym_doom.doom_take_cover import DoomTakeCoverEnv
# from gym.utils import seeding
# class DoomTakeCoverMDNRNN(DoomTakeCoverEnv):
# def __init__(self, args, render_mode=False, load_model=True, with_obs=False):
# super(DoomTakeCoverMDNRNN, self).__init__()
# self.with_obs = with_obs
# self.no_render = True
# if render_mode:
# self.no_render = False
import matplotlib.animation
import matplotlib.pyplot as plt
import wrappers
import gym
import car_racing_environment
import numpy as np
import os
from PIL import Image
import json
import tensorflow as tf
import random
from vae.vae import CVAE
# from env import make_env
from utils import PARSER
args = PARSER.parse_args(['--config_path', 'configs/carracing.config'])
import pygame
pygame.init()
screen = pygame.display.set_mode((600, 300))
frame_skip = 3
seed = 2
env = wrappers.EvaluationWrapper(wrappers.VaeCarWrapper(
gym.make("CarRacingSoftFS{}-v0".format(frame_skip))),
seed,
evaluate_for=15,
report_each=1)
DATA_DIR = "export"
model_path_name = "models/tf_vae".format(args.exp_name, args.env_name)
vae = CVAE(args)
def main():
print("Setting niceness to 19")
if "nice" in os.__dict__:
os.nice(19)
args = PARSER.parse_args()
def make_env_with_args():
return make_env(args=args, keep_image=True, wrap_rnn=False)
dir_name = get_path(args, "record", create=True)
controller = None
if args.extract_use_controller:
controller = make_controller(args=args)
env = make_env_with_args()
has_camera_data = isinstance(
env.observation_space,
gym.spaces.Dict) and "camera" in env.observation_space.spaces
format_str = "[{success:s}] {done:s} after {frames:4d} frames, reward {reward:6.1f} " \
"(Total: {total_frames:7d} frames, {successful_trials:3d}/{total_trials:3d} successful trials)"
total_frames = 0
successful_trials = 0
for trial in range(args.max_trials):
try:
seed = random.randint(0, 2**31 - 1)
filename = dir_name / (str(seed) + ".npz")
np.random.seed(seed)
env.seed(seed)
recording_image = []
recording_camera = []
recording_action = []
recording_reward = []
recording_done = []
# random policy
if args.extract_use_controller:
controller.init_random_model_params(stddev=np.random.rand() *
0.01)
repeat_action = np.random.randint(1, 11)
action = [0] * args.a_width
total_reward = 0
obs = env.reset()
frame = 0
ended_early = False
for frame in range(args.max_frames):
# Save current observation
recording_image.append(obs["image"])
if has_camera_data:
recording_camera.append(obs["camera"])
# Get next action (random)
if not args.extract_repeat_actions or frame % repeat_action == 0:
if args.extract_use_controller:
action = controller.get_action(obs["features"])
else:
action = np.random.rand(args.a_width) * 2.0 - 1.0
if args.extract_repeat_actions:
repeat_action = np.random.randint(1, 11)
# Save action
recording_action.append(action)
# Perform action
obs, reward, done, _info = env.step(action)
total_reward += reward
# Save reward and done flag
recording_reward.append(reward)
recording_done.append(done)
# Stop when done
if done:
ended_early = True
break
total_frames += (frame + 1)
enough_frames = len(recording_image) >= args.min_frames
# Save episode to disk (if it has required minimum length)
if enough_frames:
successful_trials += 1
recording_image = np.array(recording_image, dtype=np.uint8)
recording_camera = np.array(recording_camera, dtype=np.float16)
recording_action = np.array(recording_action, dtype=np.float16)
recording_reward = np.array(recording_reward, dtype=np.float16)
recording_done = np.array(recording_done, dtype=np.bool)
data = {
"image": recording_image,
"action": recording_action,
"reward": recording_reward,
"done": recording_done
}
if has_camera_data:
data["camera"] = recording_camera
np.savez_compressed(str(filename), **data)
print(
format_str.format(success="O" if enough_frames else " ",
done="Done" if ended_early else "Stop",
frames=frame + 1,
reward=total_reward,
total_frames=total_frames,
successful_trials=successful_trials,
total_trials=trial + 1))
except gym.error.Error as e:
print("Gym raised an error: " + str(e))
env.close()
env = make_env_with_args()
env.close()
