# Initialize policy
expert = DDPG(state_dim, action_dim, max_action, net_type="dense")
expert.load("model-here",
directory="../duckietown_rl/pytorch_models",
for_inference=True)
# Initialize the environment
env.reset()
# Get features(state representation) for RL agent
obs = env.get_features()
EPISODES, STEPS = 20, 1000
DEBUG = False
# please notice
logger = Logger(env, log_file=f'train-{int(EPISODES*STEPS/1000)}k.log')
start_time = time.time()
print(
f"[INFO]Starting to get logs for {EPISODES} episodes each {STEPS} steps..")
with torch.no_grad():
# let's collect our samples
for episode in range(0, EPISODES):
for steps in range(0, STEPS):
# we use our 'expert' to predict the next action.
action = expert.predict(np.array(obs))
# Apply the action
observation, reward, done, info = env.step(action)
# Get features(state representation) for RL agent
obs = env.get_features()
DEBUG = True
#for map_name in ["loop_obstacles"]:
for map_name in ["udem1"]:
env = DuckietownEnv(
map_name=
map_name, # check the Duckietown Gym documentation, there are many maps of different complexity
max_steps=EPISODES * STEPS,
distortion=True,
domain_rand=False)
# this is an imperfect demonstrator... I'm sure you can construct a better one.
expert = PurePursuitExpert(env=env)
# please notice
logger = Logger(env, log_file='train.log')
# let's collect our samples
for episode in range(0, EPISODES):
for steps in range(0, STEPS):
if USER_INPUT:
key = get_user_direction()
try:
# we use our 'expert' to predict the next action.
action = expert.predict(None, user_input=key)
observation_big, reward, done, info = env.step(action)
# we can resize the image here
observation = cv2.resize(observation_big, (80, 60))
# NOTICE: OpenCV changes the order of the channels !!!
import numpy as np
import pandas as pd
import os
import collections
import rosbag
import cv_bridge
from copy import copy
from extract_data_functions import image_preprocessing, synchronize_data
from _loggers import Logger
import cv2
# A collection of ros messages coming from a single topic.
MessageCollection = collections.namedtuple("MessageCollection",
["topic", "type", "messages"])
frank_logger = Logger(log_file='training_data.log')
def extract_messages(path, requested_topics):
# check if path is string and requested_topics a list
assert isinstance(path, str)
assert isinstance(requested_topics, list)
bag = rosbag.Bag(path)
_, available_topics = bag.get_type_and_topic_info()
#print(available_topics)
# check if the requested topics exist in bag's topics and if yes extract the messages only for them