def validation(self):
self.is_valid = True
self.is_test = True
saver = ReplaySaver()
state = self.env.reset(resize=RESIZE, size=RESIZE_SIZE)
state = torch.from_numpy(state).permute(2, 0, 1).unsqueeze(0).type(
torch.FloatTensor).to(self.device)
done = False
total_reward = 0
if VALID_INVINCIBLE == True:
self.invincible = True
else:
self.invincible = False
self.valid_hit = 0
vaild_t = 0
if VALID_INVINCIBLE == False:
while not done:
action = self.select_action(state)
state, reward, done, end = self.step(action)
total_reward += reward
print('\r Now rewards: {}'.format(total_reward), end='\r')
saver.get_current_frame()
else:
while (not done and vaild_t <= VALID_MAX_T):
vaild_t += 1
action = self.select_action(state)
state, reward, done, end = self.step(action)
total_reward += reward
print('\r Now rewards: {}'.format(total_reward), end='\r')
saver.get_current_frame()
self.valid_hit += self.env.is_collision()
self.valid_hit_list.append(self.valid_hit)
np.save('./checkpoints/valid_hits', np.array(self.valid_hit_list))
self.valid_reward.append(total_reward)
print('Valid reward: {} | Hit get: {}'.format(total_reward,
self.valid_hit))
np.save('./checkpoints/valid_reward', np.array(self.valid_reward))
saver.save_best()
print('Making video...')
saver.make_video('./checkpoints/checkpoint_{}.mp4'.format(
self.checkpoint_n))
print('Validation video made!!')
self.is_test = False
self.is_valid = False
if INVINCIBLE == True:
self.invincible = True
else:
self.invincible = False
def test(self,
episodes=10,
saving_path='./test.mp4',
size=(750, 750),
fps=60):
saver = ReplaySaver()
rewards = []
best_reward = -np.inf
for i in range(episodes):
done = False
state = self.env.reset(resize=is_resize, size=resize_size)
saver.get_current_frame()
episode_reward = 0.0
# play one game
while (not done):
state = torch.from_numpy(state).permute(
2, 0, 1).unsqueeze(0).type(torch.cuda.FloatTensor).to(
self.device)
action = self.make_action(state, test=True)
state, reward, done, end = self.env.step(action,
resize=is_resize,
size=resize_size)
saver.get_current_frame()
episode_reward += reward
print('\r episode: {} | Now reward: {} '.format(
i + 1, episode_reward),
end='\r')
while (end):
_, _, _, end = self.env.step(action)
saver.get_current_frame()
if episode_reward <= best_reward:
saver.reset()
else:
best_reward = episode_reward
saver.save_best()
saver.reset()
rewards.append(episode_reward)
print('\nRun %d episodes' % (episodes))
print('Mean:', np.mean(rewards))
print('Median:', np.median(rewards))
print('Saving best reward video')
saver.make_video(path=saving_path, size=size, fps=fps)
print('Video saved :)')
def test_random(self,
episodes=10,
saving_path='./test_random.mp4',
size=(750, 750),
fps=30,
invincible_mode=False,
test_max_t=500):
saver = ReplaySaver()
rewards = []
best_reward = -np.inf
self.is_test = True
if invincible_mode == True:
self.invincible = True
else:
self.invincible = False
for i in range(episodes):
done = False
state = self.env.reset(resize=RESIZE, size=RESIZE_SIZE)
saver.get_current_frame()
episode_reward = 0.0
# play one game
if self.invincible == False:
while (not done):
action = np.random.choice(5, 1)[0]
state, reward, done, end = self.env.step(
action,
resize=RESIZE,
size=RESIZE_SIZE,
invincible=self.invincible)
saver.get_current_frame()
episode_reward += reward
print('\r episode: {} | Now reward: {} '.format(
i + 1, episode_reward),
end='\r')
else:
t = 0
while (t <= test_max_t):
t += 1
action = np.random.choice(5, 1)[0]
state, reward, done, end = self.env.step(
action,
resize=RESIZE,
size=RESIZE_SIZE,
invincible=self.invincible)
saver.get_current_frame()
episode_reward += reward
print('\r episode: {} | Now reward: {} '.format(
i + 1, episode_reward),
end='\r')
if episode_reward <= best_reward:
saver.reset()
else:
best_reward = episode_reward
saver.save_best()
saver.reset()
rewards.append(episode_reward)
print('Run %d episodes' % (episodes))
print('Mean:', np.mean(rewards))
print('Median:', np.median(rewards))
print('Saving best reward video')
saver.make_video(path=saving_path, size=size, fps=fps)
np.save('./checkpoints/test_random_reward', rewards)
print('Video saved :)')
def test(self,
episodes=10,
saving_path='./test.mp4',
size=(750, 750),
fps=30,
invincible_mode=False,
test_max_t=500):
saver = ReplaySaver()
rewards = []
best_reward = -np.inf
self.is_test = True
if invincible_mode == True:
self.invincible = True
else:
self.invincible = False
for i in range(episodes):
done = False
state = self.env.reset(resize=RESIZE, size=RESIZE_SIZE)
state = torch.from_numpy(state).permute(2, 0, 1).unsqueeze(0).type(
torch.cuda.FloatTensor).to(DEVICE)
saver.get_current_frame()
episode_reward = 0.0
# play one game
if self.invincible == False:
while (not done):
action = self.select_action(state)
state, reward, done, end = self.step(action)
saver.get_current_frame()
episode_reward += reward
print('\r episode: {} | Now reward: {} '.format(
i + 1, episode_reward),
end='\r')
else:
self.t = 0
while (self.t <= test_max_t):
self.t += 1
action = self.select_action(state)
state, reward, done, end = self.step(action)
saver.get_current_frame()
episode_reward += reward
print('\r episode: {} | Now reward: {} '.format(
i + 1, episode_reward),
end='\r')
if episode_reward <= best_reward:
saver.reset()
else:
best_reward = episode_reward
saver.save_best()
saver.reset()
rewards.append(episode_reward)
print('Run %d episodes' % (episodes))
print('Mean:', np.mean(rewards))
print('Median:', np.median(rewards))
print('Saving best reward video')
saver.make_video(path=saving_path, size=size, fps=fps)
np.save('./checkpoints/test_reward', rewards)
print('Video saved :)')