def run(episodes, discount, useStreak):
start_time = time.time()
print('run with: ' + str(discount) + str(useStreak))
batch_size = 32
hidden_dim = 300
experienceSize = 20000
epsilon_min_after = 1500
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
writer = SummaryWriter('runs/run_discount_' + str(discount) + '_streak_' +
str(useStreak))
trainer = Trainer(hidden_dim, experienceSize, discount, batch_size, device,
writer)
for e in range(episodes):
observation = trainer.reset()
done = False
batchReward = 0
steps = 0
threes = 0
while not done:
action = trainer.policyAction(observation['board'], e,
epsilon_min_after)
old_obs = observation
observation, reward, done, _ = trainer.step(int(action))
reshaped = trainer.reshape(torch.tensor(observation['board']))
threes += trainer.streakReward(trainer.player, reshaped,
int(action))
if useStreak:
reward = trainer.change_reward_streak(reward, done, reshaped,
int(action), useStreak)
else:
reward = trainer.change_reward(reward, done)
next_state = observation['board']
exp = Experience(old_obs['board'], action, reward, next_state,
int(done))
trainer.addExperience(exp)
batchReward += reward
loss = trainer.train()
steps += 1
threes /= 3
if loss != None:
writer.add_scalar('trainLoss', loss, e)
writer.add_scalar('batchReward', batchReward, e)
writer.add_scalar('steps', steps, e)
writer.add_scalar('threes', threes, e)
if (e % 1000 == 0) and (e > 0):
trainer.switchPosition()
if e % 50 == 0:
trainer.synchronize()
firstStep = str(
trainer.policy(
trainer.reshape(torch.tensor(
trainer.reset()['board'])))[0])
writer.add_text('first_qs', firstStep, e)
if e % 50 == 0:
trainer.save("model_state_discount_" + str(discount) +
'_useStreak_' + str(useStreak))
# print(e)
# print("episode: " + str(e) + " meanReward generateEpisodes: " + str(batchReward) + " meanLoss: " + str(loss))
# print("steps: " + str(steps))
# firstStep = str(trainer.policy(trainer.reshape(torch.tensor(trainer.reset()['board'])))[0])
# with torch.no_grad():
# print(firstStep)
evaluator = Evaluator(100, trainer)
evaluator.winPercentage(e)
# if e % 25000 == 0:
# trainer.switch()
# trainer.save("model_state_"+str(e))
print("--- %s seconds ---" % (time.time() - start_time))
