# Written by Francois Fleuret <francois@fleuret.org>
-import math, os, tqdm
+import math, os, tqdm, warnings
import torch, torchvision
class Task:
- def batches(self, split="train"):
+ def batches(self, split="train", nb_to_use=-1, desc=None):
pass
def vocabulary_size(self):
self.train_input = self.tensorize(self.train_descr)
self.test_input = self.tensorize(self.test_descr)
- def batches(self, split="train"):
+ 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
for batch in tqdm.tqdm(
self.t_nul = self.token2id["#"]
self.t_true = self.token2id["true"]
self.t_false = self.token2id["false"]
- self.t_pipe = self.token2id["|"]
+ # self.t_pipe = self.token2id["|"]
# Tokenize the train and test sets
self.train_input = self.str2tensor(self.train_descr)
None if len(self.play_descr) == 0 else self.str2tensor(self.play_descr)
)
- def batches(self, split="train"):
+ 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
for batch in tqdm.tqdm(
self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
- def batches(self, split="train"):
+ 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
for batch in tqdm.tqdm(
######################################################################
-import escape
+import greed
-class Escape(Task):
+class Greed(Task):
def __init__(
self,
nb_train_samples,
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 = escape.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 = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
- # 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)
+ 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.world.it_len == self.world.index_lookahead_reward
+ ).long()
+
+ 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"}
input = self.train_input if split == "train" else self.test_input
for batch in tqdm.tqdm(
input.split(self.batch_size), dynamic_ncols=True, desc=desc
):
- yield batch
+ yield self.wipe_lookahead_rewards(batch)
def vocabulary_size(self):
- return escape.nb_codes
+ return self.world.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[:, 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[:100].detach().clone())
# Generate iteration after iteration
- 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)
+ result = self.test_input[:250].clone()
+ # Erase all the content but that of the first iteration
+ result[:, self.world.it_len :] = -1
+ # Set the lookahead_reward of the firs to UNKNOWN
+ result[:, self.world.index_lookahead_reward] = self.world.lookahead_reward2code(
+ greed.REWARD_UNKNOWN
+ )
- snapshots = []
+ t = torch.arange(result.size(1), device=result.device)[None, :]
for u in tqdm.tqdm(
- range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+ range(0, result.size(1), self.world.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())
-
- # Generate the state
- ar_mask = (t >= u).long() * (t < u + state_len).long()
+ # Generate the next state but keep the initial one, the
+ # lookahead_reward of previous iterations are set to
+ # UNKNOWN
+ if u > 0:
+ result[
+ :, 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.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)
- 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()
- 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())
+ # Set the lookahead_reward to UNKNOWN for the next iterations
+ 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:
- s, a, r, lr = escape.seq2episodes(
- s[n : n + 1], self.height, self.width, lookahead=True
+ lr, s, a, r = self.world.seq2episodes(
+ s[n : n + 1],
)
- str = escape.episodes2str(
- s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True
+ str = self.world.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 = 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:
def produce_results(
self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
):
- result = self.test_input[:250].clone()
+ result = self.wipe_lookahead_rewards(self.test_input[:250].clone())
# 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 = 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_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
- 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 = 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_seq_{n_epoch:04d}.txt")
with open(filename, "w") as f:
projects / picoclvr.git / blobdiff