# 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)
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
+ index_lookahead_reward = 0
+ index_states = 1
+ index_action = state_len + 1
+ index_reward = state_len + 2
+ it_len = state_len + 3 # lookahead_reward / state / action / reward
result[:, it_len:] = -1
+ snapshots = []
+
def ar(result, ar_mask, logit_biases=None):
ar_mask = ar_mask.expand_as(result)
result *= 1 - ar_mask
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[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"
):
- # 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())
+ lr, _, _, _ = escape.seq2episodes(result[:, :u], self.height, self.width)
- # Generate the state
- ar_mask = (t >= u).long() * (t < u + state_len).long()
+ # Generate the lookahead_reward and state
+ ar_mask = (t % it_len == index_lookahead_reward).long() * (
+ t <= u + index_lookahead_reward
+ ).long()
+ ar(result, ar_mask)
+
+ # Generate the lookahead_reward and state
+ ar_mask = (t >= u + index_states).long() * (
+ t < u + index_states + state_len
+ ).long()
ar(result, ar_mask)
- snapshots.append(result[:10].detach().clone())
- # Re-generate the lookahead_reward optimistically in the
- # previous iterations
- ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
+ # Re-generate the lookahead_reward
+ ar_mask = (t % it_len == index_lookahead_reward).long() * (
+ t <= u + 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:
- s, a, r, lr = escape.seq2episodes(
- s[n : n + 1], self.height, self.width, lookahead=True
+ lr, s, a, r = escape.seq2episodes(
+ s[n : n + 1], self.height, self.width
)
str = escape.episodes2str(
- s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True
+ 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: