######################################################################
-def generate_episodes(nb, height=6, width=6, T=10):
+def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
rnd = torch.rand(nb, height, width)
rnd[:, 0, :] = 0
rnd[:, -1, :] = 0
rnd[:, :, -1] = 0
wall = 0
- for k in range(3):
+ for k in range(nb_walls):
wall = wall + (
rnd.flatten(1).argmax(dim=1)[:, None]
== torch.arange(rnd.flatten(1).size(1))[None, :]
).long().reshape(rnd.size())
+
rnd = rnd * (1 - wall.clamp(max=1))
states = wall[:, None, :, :].expand(-1, T, -1, -1).clone()
)
hit = (hit > 0).long()
- assert hit.min() == 0 and hit.max() <= 1
+ # assert hit.min() == 0 and hit.max() <= 1
rewards[:, t + 1] = -hit + (1 - hit) * agent[:, t + 1, -1, -1]
r = rewards[:, :, None]
rewards = (r + 1) + first_rewards_code
- assert (
- states.min() >= first_state_code
- and states.max() < first_state_code + nb_state_codes
- )
- assert (
- actions.min() >= first_actions_code
- and actions.max() < first_actions_code + nb_actions_codes
- )
- assert (
- rewards.min() >= first_rewards_code
- and rewards.max() < first_rewards_code + nb_rewards_codes
- )
+ # assert (
+ # states.min() >= first_state_code
+ # and states.max() < first_state_code + nb_state_codes
+ # )
+ # assert (
+ # actions.min() >= first_actions_code
+ # and actions.max() < first_actions_code + nb_actions_codes
+ # )
+ # assert (
+ # rewards.min() >= first_rewards_code
+ # and rewards.max() < first_rewards_code + nb_rewards_codes
+ # )
if lookahead_delta is None:
return torch.cat([states, actions, rewards], dim=2).flatten(1)
else:
- assert (
- lookahead_rewards.min() >= first_lookahead_rewards_code
- and lookahead_rewards.max()
- < first_lookahead_rewards_code + nb_lookahead_rewards_codes
- )
+ # assert (
+ # lookahead_rewards.min() >= first_lookahead_rewards_code
+ # and lookahead_rewards.max()
+ # < first_lookahead_rewards_code + nb_lookahead_rewards_codes
+ # )
return torch.cat([states, actions, rewards, lookahead_rewards], dim=2).flatten(
1
)
states, actions, rewards, lookahead_rewards=None, unicode=False, ansi_colors=False
):
if unicode:
- symbols = " █@$"
+ symbols = "·█@$"
# vert, hori, cross, thin_hori = "║", "═", "╬", "─"
vert, hori, cross, thin_vert, thin_hori = "┃", "━", "╋", "│", "─"
else:
+ "\n"
)
- result += (vert + thin_hori * states.size(-1)) * states.size(1) + vert + "\n"
+ # result += (vert + thin_hori * states.size(-1)) * states.size(1) + vert + "\n"
def status_bar(a, r, lr=None):
a, r = a.item(), r.item()
######################################################################
if __name__ == "__main__":
- nb, height, width, T = 1000, 4, 6, 20
- states, actions, rewards = generate_episodes(nb, height, width, T)
+ nb, height, width, T, nb_walls = 25, 5, 7, 25, 5
+ states, actions, rewards = generate_episodes(nb, height, width, T, nb_walls)
seq = episodes2seq(states, actions, rewards, lookahead_delta=T)
s, a, r, lr = seq2episodes(seq, height, width, lookahead=True)
print(episodes2str(s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True))