######################################################################
nb_state_codes = 4
-nb_rewards_codes = 3
nb_actions_codes = 5
+nb_rewards_codes = 3
+nb_lookahead_rewards_codes = 3
first_state_code = 0
-first_rewards_code = first_state_code + nb_state_codes
-first_actions_code = first_rewards_code + nb_rewards_codes
-nb_codes = first_actions_code + nb_actions_codes
+first_actions_code = first_state_code + nb_state_codes
+first_rewards_code = first_actions_code + nb_actions_codes
+first_lookahead_rewards_code = first_rewards_code + nb_rewards_codes
+nb_codes = first_lookahead_rewards_code + nb_lookahead_rewards_codes
######################################################################
-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]
######################################################################
-def episodes2seq(states, actions, rewards):
+def episodes2seq(states, actions, rewards, lookahead_delta=None):
states = states.flatten(2) + first_state_code
actions = actions[:, :, None] + first_actions_code
- rewards = (rewards[:, :, None] + 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
- )
-
- return torch.cat([states, actions, rewards], dim=2).flatten(1)
-
-
-def seq2episodes(seq, height, width):
- seq = seq.reshape(seq.size(0), -1, height * width + 2)
+
+ if lookahead_delta is not None:
+ # r = rewards
+ # u = F.pad(r, (0, lookahead_delta - 1)).as_strided(
+ # (r.size(0), r.size(1), lookahead_delta),
+ # (r.size(1) + lookahead_delta - 1, 1, 1),
+ # )
+ # a = u[:, :, 1:].min(dim=-1).values
+ # b = u[:, :, 1:].max(dim=-1).values
+ # s = (a < 0).long() * a + (a >= 0).long() * b
+ # lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code
+
+ # a[n,t]=min_s>t r[n,s]
+ a = rewards.new_zeros(rewards.size())
+ b = rewards.new_zeros(rewards.size())
+ for t in range(a.size(1) - 1):
+ a[:, t] = rewards[:, t + 1 :].min(dim=-1).values
+ b[:, t] = rewards[:, t + 1 :].max(dim=-1).values
+ s = (a < 0).long() * a + (a >= 0).long() * b
+ lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code
+
+ 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
+ # )
+
+ 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
+ # )
+ return torch.cat([states, actions, rewards, lookahead_rewards], dim=2).flatten(
+ 1
+ )
+
+
+def seq2episodes(seq, height, width, lookahead=False):
+ seq = seq.reshape(seq.size(0), -1, height * width + (3 if lookahead else 2))
states = seq[:, :, : height * width] - first_state_code
states = states.reshape(states.size(0), states.size(1), height, width)
actions = seq[:, :, height * width] - first_actions_code
rewards = seq[:, :, height * width + 1] - first_rewards_code - 1
- return states, actions, rewards
+
+ if lookahead:
+ lookahead_rewards = (
+ seq[:, :, height * width + 2] - first_lookahead_rewards_code - 1
+ )
+ return states, actions, rewards, lookahead_rewards
+ else:
+ return states, actions, rewards
+
+
+def seq2str(seq):
+ def token2str(t):
+ if t >= first_state_code and t < first_state_code + nb_state_codes:
+ return " #@$"[t - first_state_code]
+ elif t >= first_actions_code and t < first_actions_code + nb_actions_codes:
+ return "ISNEW"[t - first_actions_code]
+ elif t >= first_rewards_code and t < first_rewards_code + nb_rewards_codes:
+ return "-0+"[t - first_rewards_code]
+ elif (
+ t >= first_lookahead_rewards_code
+ and t < first_lookahead_rewards_code + nb_lookahead_rewards_codes
+ ):
+ return "n.p"[t - first_lookahead_rewards_code]
+ else:
+ return "?"
+
+ return ["".join([token2str(x.item()) for x in row]) for row in seq]
######################################################################
-def episodes2str(states, actions, rewards, unicode=False, ansi_colors=False):
+def episodes2str(
+ states, actions, rewards, lookahead_rewards=None, unicode=False, ansi_colors=False
+):
if unicode:
- symbols = " █@$"
+ symbols = "·█@$"
# vert, hori, cross, thin_hori = "║", "═", "╬", "─"
- vert, hori, cross, thin_hori = "┃", "━", "╋", "─"
+ vert, hori, cross, thin_vert, thin_hori = "┃", "━", "╋", "│", "─"
else:
symbols = " #@$"
- vert, hori, cross, thin_hori = "|", "-", "+", "-"
+ vert, hori, cross, thin_vert, thin_hori = "|", "-", "+", "|", "-"
hline = (cross + hori * states.size(-1)) * states.size(1) + cross + "\n"
result = hline
for n in range(states.size(0)):
+
+ def state_symbol(v):
+ v = v.item()
+ return "?" if v < 0 or v >= len(symbols) else symbols[v]
+
for i in range(states.size(2)):
result += (
vert
+ vert.join(
- [
- "".join([symbols[v.item()] for v in row])
- for row in states[n, :, i]
- ]
+ ["".join([state_symbol(v) for v in row]) for row in states[n, :, i]]
)
+ vert
+ "\n"
)
- result += (vert + thin_hori * states.size(-1)) * states.size(1) + vert + "\n"
-
- def status_bar(a, r):
- a = "ISNEW"[a.item()]
- r = "" if r == 0 else f"{r.item()}"
- return a + " " * (states.size(-1) - len(a) - len(r)) + r
+ # 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()
+ sb_a = "ISNEW"[a] if a >= 0 and a < 5 else "?"
+ sb_r = "- +"[r + 1] if r in {-1, 0, 1} else "?"
+ if lr is None:
+ sb_lr = ""
+ else:
+ lr = lr.item()
+ sb_lr = "n p"[lr + 1] if lr in {-1, 0, 1} else "?"
+ return (
+ sb_a
+ + "/"
+ + sb_r
+ + " " * (states.size(-1) - 1 - len(sb_a + sb_r + sb_lr))
+ + sb_lr
+ )
- result += (
- vert
- + vert.join([status_bar(a, r) for a, r in zip(actions[n], rewards[n])])
- + vert
- + "\n"
- )
+ if lookahead_rewards is None:
+ result += (
+ vert
+ + vert.join([status_bar(a, r) for a, r in zip(actions[n], rewards[n])])
+ + vert
+ + "\n"
+ )
+ else:
+ result += (
+ vert
+ + vert.join(
+ [
+ status_bar(a, r, lr)
+ for a, r, lr in zip(
+ actions[n], rewards[n], lookahead_rewards[n]
+ )
+ ]
+ )
+ + vert
+ + "\n"
+ )
result += hline
######################################################################
if __name__ == "__main__":
- nb, height, width, T = 8, 4, 6, 20
- states, actions, rewards = generate_episodes(nb, height, width, T)
- seq = episodes2seq(states, actions, rewards)
- s, a, r = seq2episodes(seq, height, width)
- print(episodes2str(s, a, r, unicode=True, ansi_colors=True))
+ 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))
+ # print()
+ # for s in seq2str(seq):
+ # print(s)