projects
/
picoclvr.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Update.
[picoclvr.git]
/
tasks.py
diff --git
a/tasks.py
b/tasks.py
index
829eb24
..
56c2b0f
100755
(executable)
--- a/
tasks.py
+++ b/
tasks.py
@@
-1874,6
+1874,7
@@
class Escape(Task):
height,
width,
T,
height,
width,
T,
+ nb_walls,
logger=None,
device=torch.device("cpu"),
):
logger=None,
device=torch.device("cpu"),
):
@@
-1885,7
+1886,7
@@
class Escape(Task):
self.width = width
states, actions, rewards = escape.generate_episodes(
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]
)
seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
# seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
@@
-1912,7
+1913,7
@@
class Escape(Task):
def thinking_autoregression(
self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
):
def thinking_autoregression(
self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
):
- result = self.test_input[:
10
0].clone()
+ result = self.test_input[:
25
0].clone()
t = torch.arange(result.size(1), device=result.device)[None, :]
state_len = self.height * self.width
t = torch.arange(result.size(1), device=result.device)[None, :]
state_len = self.height * self.width
@@
-1936,9
+1937,9
@@
class Escape(Task):
for u in tqdm.tqdm(
range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
):
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)
result[:, u - 1] = s + 1 + escape.first_lookahead_rewards_code
ar_mask = (t >= u).long() * (t < u + state_len).long()
ar(result, ar_mask)
@@
-1954,7
+1955,7
@@
class Escape(Task):
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)]
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
a = r.min(dim=1).values
b = r.max(dim=1).values
s = (a < 0).long() * a + (a >= 0).long() * b