height,
width,
T,
+ nb_walls,
logger=None,
device=torch.device("cpu"),
):
self.width = width
states, actions, rewards = escape.generate_episodes(
- nb_train_samples + nb_test_samples, height, width, T
+ nb_train_samples + nb_test_samples, height, width, T, nb_walls
)
seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
# seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
def thinking_autoregression(
self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
):
- result = self.test_input[:100].clone()
+ result = self.test_input[:250].clone()
t = torch.arange(result.size(1), device=result.device)[None, :]
state_len = self.height * self.width
for u in tqdm.tqdm(
range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
):
- # Put the lookahead reward to -1 for the current iteration,
- # sample the next state
- s = -1
+ # Put the lookahead reward to either 0 or -1 for the
+ # current iteration, sample the next state
+ s = -1 # (torch.rand(result.size(0), device = result.device) < 0.2).long()
result[:, u - 1] = s + 1 + escape.first_lookahead_rewards_code
ar_mask = (t >= u).long() * (t < u + state_len).long()
ar(result, ar_mask)
for v in range(0, u, it_len):
# Extract the rewards
r = result[:, range(v + state_len + 1 + it_len, u + it_len - 1, it_len)]
- r = r - escape.first_lookahead_rewards_code - 1
+ r = r - escape.first_rewards_code - 1
a = r.min(dim=1).values
b = r.max(dim=1).values
s = (a < 0).long() * a + (a >= 0).long() * b