def test_pong_num_actions(self):
env = gym.make('PongDeterministic-v4')
env = wrap_deepmind(env, frame_stack=True)
env = wrap_pytorch(env)
env.seed(42)
observation = env.reset()
self.assertEqual(env.action_space.n, 6)
def test_frame_shape(self):
env = gym.make('BreakoutDeterministic-v4')
env = wrap_deepmind(env, frame_stack=True)
env = wrap_pytorch(env)
env.seed(42)
observation = env.reset()
self.assertEqual(observation.shape, (4, 84, 84))
def test_env_input(self):
env = gym.make('BreakoutDeterministic-v4')
env = wrap_deepmind(env, frame_stack=True)
env = wrap_pytorch(env)
env.seed(42)
state = env.reset()
# Need to normalize inputs to range of 0-1
state = torch.tensor(state, dtype=torch.float32) / 255.0
state = state.unsqueeze(dim=0)
DQNModel = AtariCNN(env.observation_space.shape, num_actions=4)
out = DQNModel(state)
self.assertEqual(out.size(), (1, 4))
def test_breakout_rewards(self):
env = gym.make('BreakoutDeterministic-v4')
env = wrap_deepmind(env, frame_stack=True)
env = wrap_pytorch(env)
env.seed(42)
state = env.reset()
done = False
tot_reward = 0.0
while not done:
state, reward, done, _ = env.step(env.action_space.sample())
tot_reward += reward
self.assertGreaterEqual(tot_reward, 0)
def make_env(env_id):
env = gym.make(env_id)
env = wrap_deepmind(env, frame_stack=True)
env = wrap_pytorch(env)
env.seed(42)
return env
def make_env(env_id):
env = gym.make(env_id)
env = wrap_deepmind(env, frame_stack=True)
env = wrap_pytorch(env)
env.seed(42)
return env
def make_net(inp_shape, num_actions):
PolicyNet = AtariCNN(inp_shape, num_actions)
TargetNet = AtariCNN(inp_shape, num_actions)
return PolicyNet, TargetNet
random.seed(42)
env = gym.make('PongDeterministic-v4')
env = wrap_deepmind(env, frame_stack=True, clip_rewards=True)
env = wrap_pytorch(env)
env.seed(42)
# lst = []
# for i in range(1):
# state = env.reset()
# done = False
# tot_reward = 0.0
# while not done:
# state, reward, done, _ = env.step(env.action_space.sample())
# env.render()
# time.sleep(0.01)
# tot_reward += reward
# print(i, tot_reward, state.shape)
# lst.append(tot_reward)