# Written by Francois Fleuret <francois@fleuret.org>
-import math, os, tqdm
+import math, os, tqdm, warnings
import torch, torchvision
states, actions, rewards = escape.generate_episodes(
nb_train_samples + nb_test_samples, height, width, T, nb_walls
)
- seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
+ seq = escape.episodes2seq(states, actions, rewards)
# seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
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
+ 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 batches(self, split="train", nb_to_use=-1, desc=None):
assert split in {"train", "test"}
for batch in tqdm.tqdm(
input.split(self.batch_size), dynamic_ncols=True, desc=desc
):
+ t = torch.arange(input.size(1), device=input.device)[None, :]
+ u = torch.randint(input.size(1), (input.size(0), 1), device=input.device)
+ lr_mask = (t <= u).long() * (
+ t % self.it_len == self.index_lookahead_reward
+ ).long()
+
+ input = lr_mask * escape.lookahead_reward2code(2) + (1 - lr_mask) * 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
result = self.test_input[:250].clone()
t = torch.arange(result.size(1), device=result.device)[None, :]
- state_len = self.height * self.width
- index_action = state_len
- index_reward = state_len + 1
- index_lookahead_reward = state_len + 2
- it_len = state_len + 3 # state / action / reward / lookahead_reward
+ result[:, self.it_len :] = -1
+
+ snapshots = []
def ar(result, ar_mask, logit_biases=None):
ar_mask = ar_mask.expand_as(result)
device=self.device,
progress_bar_desc=None,
)
+ warnings.warn("keeping thinking snapshots", RuntimeWarning)
+ snapshots.append(result[:10].detach().clone())
# 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)
for u in tqdm.tqdm(
- range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+ range(self.it_len, result.size(1) - self.it_len + 1, self.it_len),
+ desc="thinking",
):
- # Generate the lookahead_reward pessimistically
- ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
- ar(result, ar_mask, logit_biases=-optimistic_bias)
+ # Generate the lookahead_reward and state
+ ar_mask = (t % self.it_len == self.index_lookahead_reward).long() * (
+ t <= u + self.index_lookahead_reward
+ ).long()
+ ar(result, ar_mask)
- # Generate the state
- ar_mask = (t >= u).long() * (t < u + state_len).long()
+ # Generate the lookahead_reward and state
+ ar_mask = (t >= u + self.index_states).long() * (
+ t < u + self.index_states + self.state_len
+ ).long()
ar(result, ar_mask)
- # Generate the lookahead_reward optimistically
- ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
+ # Re-generate the lookahead_reward
+ ar_mask = (t % self.it_len == self.index_lookahead_reward).long() * (
+ t <= u + self.index_lookahead_reward
+ ).long()
ar(result, ar_mask, logit_biases=optimistic_bias)
# Generate the action and reward
- ar_mask = (t >= u + index_action).long() * (t <= u + index_reward).long()
+ ar_mask = (t >= u + self.index_action).long() * (
+ t <= u + self.index_reward
+ ).long()
ar(result, ar_mask)
+ 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
+ )
+ 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
- )
+ lr, s, a, r = 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
+ lr, s, a, r = 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_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
+ lr, s, a, r = 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_seq_{n_epoch:04d}.txt")
with open(filename, "w") as f: