width,
T,
nb_walls,
+ nb_coins,
logger=None,
device=torch.device("cpu"),
):
self.batch_size = batch_size
self.device = device
- self.height = height
- self.width = width
- states, actions, rewards = greed.generate_episodes(
- nb_train_samples + nb_test_samples, height, width, T, nb_walls
+ self.world = greed.GreedWorld(height, width, T, nb_walls, nb_coins)
+
+ states, actions, rewards = self.world.generate_episodes(
+ nb_train_samples + nb_test_samples
)
- seq = greed.episodes2seq(states, actions, rewards)
- # seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
+ seq = self.world.episodes2seq(states, actions, rewards)
self.train_input = seq[:nb_train_samples].to(self.device)
self.test_input = seq[nb_train_samples:].to(self.device)
- self.state_len = self.height * self.width
- self.index_lookahead_reward = 0
- self.index_states = 1
- self.index_action = self.state_len + 1
- self.index_reward = self.state_len + 2
- self.it_len = self.state_len + 3 # lookahead_reward / state / action / reward
-
def wipe_lookahead_rewards(self, batch):
t = torch.arange(batch.size(1), device=batch.device)[None, :]
u = torch.randint(batch.size(1), (batch.size(0), 1), device=batch.device)
lr_mask = (t <= u).long() * (
- t % self.it_len == self.index_lookahead_reward
+ t % self.world.it_len == self.world.index_lookahead_reward
).long()
- return lr_mask * greed.lookahead_reward2code(2) + (1 - lr_mask) * batch
+ return (
+ lr_mask * self.world.lookahead_reward2code(greed.REWARD_UNKNOWN)
+ + (1 - lr_mask) * batch
+ )
def batches(self, split="train", nb_to_use=-1, desc=None):
assert split in {"train", "test"}
yield self.wipe_lookahead_rewards(batch)
def vocabulary_size(self):
- return greed.nb_codes
+ return self.world.nb_codes
def thinking_autoregression(
self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
progress_bar_desc=None,
)
warnings.warn("keeping thinking snapshots", RuntimeWarning)
- snapshots.append(result[:10].detach().clone())
+ snapshots.append(result[:100].detach().clone())
# Generate iteration after iteration
result = self.test_input[:250].clone()
# Erase all the content but that of the first iteration
- result[:, self.it_len :] = -1
+ result[:, self.world.it_len :] = -1
# Set the lookahead_reward of the firs to UNKNOWN
- result[:, self.index_lookahead_reward] = greed.lookahead_reward2code(2)
+ result[:, self.world.index_lookahead_reward] = self.world.lookahead_reward2code(
+ greed.REWARD_UNKNOWN
+ )
t = torch.arange(result.size(1), device=result.device)[None, :]
for u in tqdm.tqdm(
- range(0, result.size(1), self.it_len),
+ range(0, result.size(1), self.world.it_len),
desc="thinking",
):
# Generate the next state but keep the initial one, the
# UNKNOWN
if u > 0:
result[
- :, u + self.index_lookahead_reward
- ] = greed.lookahead_reward2code(2)
- ar_mask = (t >= u + self.index_states).long() * (
- t < u + self.index_states + self.state_len
+ :, u + self.world.index_lookahead_reward
+ ] = self.world.lookahead_reward2code(greed.REWARD_UNKNOWN)
+ ar_mask = (t >= u + self.world.index_states).long() * (
+ t < u + self.world.index_states + self.world.state_len
).long()
ar(result, ar_mask)
# Generate the action and reward with lookahead_reward to +1
- result[:, u + self.index_lookahead_reward] = greed.lookahead_reward2code(1)
- ar_mask = (t >= u + self.index_action).long() * (
- t <= u + self.index_reward
+ result[
+ :, u + self.world.index_lookahead_reward
+ ] = self.world.lookahead_reward2code(greed.REWARD_PLUS)
+ ar_mask = (t >= u + self.world.index_reward).long() * (
+ t <= u + self.world.index_action
).long()
ar(result, ar_mask)
# Set the lookahead_reward to UNKNOWN for the next iterations
- result[:, u + self.index_lookahead_reward] = greed.lookahead_reward2code(2)
+ result[
+ :, u + self.world.index_lookahead_reward
+ ] = self.world.lookahead_reward2code(greed.REWARD_UNKNOWN)
filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")
with open(filename, "w") as f:
- for n in range(10):
+ for n in range(snapshots[0].size(0)):
for s in snapshots:
- lr, s, a, r = greed.seq2episodes(
- s[n : n + 1], self.height, self.width
+ lr, s, a, r = self.world.seq2episodes(
+ s[n : n + 1],
)
- str = greed.episodes2str(
+ str = self.world.episodes2str(
lr, s, a, r, unicode=True, ansi_colors=True
)
f.write(str)
# Saving the generated sequences
- lr, s, a, r = greed.seq2episodes(result, self.height, self.width)
- str = greed.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
+ lr, s, a, r = self.world.seq2episodes(result)
+ str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
filename = os.path.join(result_dir, f"test_thinking_seq_{n_epoch:04d}.txt")
with open(filename, "w") as f:
# Saving the ground truth
- lr, s, a, r = greed.seq2episodes(
+ lr, s, a, r = self.world.seq2episodes(
result,
- self.height,
- self.width,
)
- str = greed.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
+ str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
filename = os.path.join(result_dir, f"test_true_seq_{n_epoch:04d}.txt")
with open(filename, "w") as f:
# Re-generating from the first frame
ar_mask = (
- torch.arange(result.size(1), device=result.device)
- >= self.height * self.width + 3
+ torch.arange(result.size(1), device=result.device) >= self.world.it_len
).long()[None, :]
ar_mask = ar_mask.expand_as(result)
result *= 1 - ar_mask # paraaaaanoiaaaaaaa
# Saving the generated sequences
- lr, s, a, r = greed.seq2episodes(
+ lr, s, a, r = self.world.seq2episodes(
result,
- self.height,
- self.width,
)
- str = greed.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
+ str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
filename = os.path.join(result_dir, f"test_seq_{n_epoch:04d}.txt")
with open(filename, "w") as f: