self.train_input = seq[:nb_train_samples].to(self.device)
self.test_input = seq[nb_train_samples:].to(self.device)
- self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
-
def batches(self, split="train", nb_to_use=-1, desc=None):
assert split in {"train", "test"}
input = self.train_input if split == "train" else self.test_input
yield batch
def vocabulary_size(self):
- return self.nb_codes
+ return escape.nb_codes
def thinking_autoregression(
self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
index_lookahead_reward = state_len + 2
it_len = state_len + 3 # state / action / reward / lookahead_reward
+ result[:, it_len:] = -1
+
def ar(result, ar_mask, logit_biases=None):
ar_mask = ar_mask.expand_as(result)
result *= 1 - ar_mask
# Generate iteration after iteration
- optimistic_bias = result.new_zeros(self.nb_codes, device=result.device)
- optimistic_bias[(-1) + escape.first_lookahead_rewards_code + 1] = math.log(1e-1)
- optimistic_bias[(1) + escape.first_lookahead_rewards_code + 1] = math.log(1e1)
+ optimistic_bias = result.new_zeros(escape.nb_codes, device=result.device)
+ optimistic_bias[escape.lookahead_reward2code(-1)] = -math.log(1e1)
+ optimistic_bias[escape.lookahead_reward2code(1)] = math.log(1e1)
+
+ snapshots = []
for u in tqdm.tqdm(
range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
):
- # Generate the lookahead_reward pessimistically
+ # Re-generate the lookahead_reward pessimistically in the
+ # previous iterations
ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
ar(result, ar_mask, logit_biases=-optimistic_bias)
+ snapshots.append(result[:10].detach().clone())
# Generate the state
ar_mask = (t >= u).long() * (t < u + state_len).long()
ar(result, ar_mask)
+ snapshots.append(result[:10].detach().clone())
- # Generate the lookahead_reward optimistically
+ # Re-generate the lookahead_reward optimistically in the
+ # previous iterations
ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
ar(result, ar_mask, logit_biases=optimistic_bias)
+ snapshots.append(result[:10].detach().clone())
# Generate the action and reward
ar_mask = (t >= u + index_action).long() * (t <= u + index_reward).long()
ar(result, ar_mask)
+ snapshots.append(result[:10].detach().clone())
+
+ 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 s in snapshots:
+ lr, s, a, r = escape.seq2episodes(
+ s[n : n + 1], self.height, self.width, lookahead=True
+ )
+ str = escape.episodes2str(
+ lr, s, a, r, unicode=True, ansi_colors=True
+ )
+ f.write(str)
+ f.write("\n\n")
# Saving the generated sequences
- s, a, r, lr = escape.seq2episodes(
- result, self.height, self.width, lookahead=True
- )
- str = escape.episodes2str(
- s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True
- )
+ s, a, r, lr = escape.seq2episodes(result, self.height, self.width)
+ str = escape.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
s, a, r, lr = escape.seq2episodes(
- result, self.height, self.width, lookahead=True
- )
- str = escape.episodes2str(
- s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True
+ result,
+ self.height,
+ self.width,
)
+ str = escape.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:
# Saving the generated sequences
s, a, r, lr = escape.seq2episodes(
- result, self.height, self.width, lookahead=True
- )
- str = escape.episodes2str(
- s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True
+ result,
+ self.height,
+ self.width,
)
+ str = escape.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: