"Visualy test Agent, after reset top frame should be a reduced image of size {} by {}"
.format(width, height))
agent._reset()
io.imshow(agent.state[-1])
plt.show()
input("Press any key to continue")
plt.close()
print(
"resetting experience buffer, then observe state and invoke agent play method 100 times ..."
)
experience_buffer.buffer.clear()
agent._reset()
for i in range(100):
play_random = True
agent.play_step(play_random)
print("number of experiences in the buffer: {}. Buffer capacity {}".format(
len(experience_buffer), capacity))
input(
"Press any key to see a random initial state image from the experience buffer"
)
states, actions, rewards, dones, next_states = experience_buffer.sample(1)
io.imshow(states[0][3])
plt.show()
input("Press any key to see the next state image for the same experience")
plt.close()
io.imshow(next_states[0][3])
plt.show()
input("Press any key to continue")
# to track the mean across epochs
total_mean = 0
total_count = 0
dt = datetime.datetime.now()
date_time_string = dt.strftime("%Y%m%d-%H-%M-%S")
model_path = "models/" + date_time_string
os.makedirs(model_path, mode=0o755, exist_ok=True)
steps = 0
last_reward_step = 0
if args.initialise_buffer == "init":
print("Initialising and saving buffer with %d transitions" %
args.replay_buffer_size)
for step in trange(args.replay_buffer_size, leave=False):
agent.play_step(True, device, write_to_buffer=False)
with open("experience_buffer.pickle", "wb") as f:
pickle.dump(experience_buffer, f)
sys.exit()
elif args.initialise_buffer == "load":
print("loading buffer ...")
with open("experience_buffer.pickle", "rb") as f:
experience_buffer = pickle.load(f)
print("buffer size: {}".format(len(experience_buffer)))
else:
print("Initialising buffer with %d transitions" % initialisation_steps)
for step in trange(initialisation_steps, leave=False):
agent.play_step(True, device)
