# Written by Francois Fleuret <francois@fleuret.org>
-import math, os, tqdm
+import math, os, tqdm, warnings
import torch, torchvision
ar_mask,
deterministic_synthesis,
forbidden_tokens=None,
+ logit_biases=None,
progress_bar_desc="autoregression",
device=torch.device("cpu"),
):
for input, ar_mask in batches:
model.masked_inplace_autoregression(
- input, ar_mask, forbidden_tokens, deterministic_synthesis
+ input,
+ ar_mask,
+ deterministic_synthesis,
+ forbidden_tokens,
+ logit_biases,
)
model.train(t)
######################################################################
-import escape
+import greed
-class Escape(Task):
+class Greed(Task):
def __init__(
self,
nb_train_samples,
self.height = height
self.width = width
- states, actions, rewards = escape.generate_episodes(
+ states, actions, rewards = greed.generate_episodes(
nb_train_samples + nb_test_samples, height, width, T, nb_walls
)
- seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
+ seq = greed.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(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
+ ).long()
+
+ batch = lr_mask * greed.lookahead_reward2code(2) + (1 - lr_mask) * batch
yield batch
def vocabulary_size(self):
- return self.nb_codes
+ return greed.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
- it_len = state_len + 3 # state / action / reward / lookahead_reward
+ snapshots = []
- def ar(result, ar_mask):
+ def ar(result, ar_mask, logit_biases=None):
ar_mask = ar_mask.expand_as(result)
result *= 1 - ar_mask
masked_inplace_autoregression(
self.batch_size,
result,
ar_mask,
- deterministic_synthesis,
+ deterministic_synthesis=deterministic_synthesis,
+ logit_biases=logit_biases,
device=self.device,
progress_bar_desc=None,
)
+ warnings.warn("keeping thinking snapshots", RuntimeWarning)
+ snapshots.append(result[:10].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
+ # Set the lookahead_reward of the firs to UNKNOWN
+ result[:, self.index_lookahead_reward] = greed.lookahead_reward2code(2)
+
+ 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.it_len),
+ desc="thinking",
):
- # Put the lookahead reward to either 0 or -1 for the
- # current iteration, with a proba that depends with the
- # sequence index, so that we have diverse examples, sample
- # the next state
- s = -(
- torch.rand(result.size(0), device=result.device)
- <= torch.linspace(0, 1, result.size(0), device=result.device)
+ # 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.index_lookahead_reward
+ ] = greed.lookahead_reward2code(2)
+ ar_mask = (t >= u + self.index_states).long() * (
+ t < u + self.index_states + self.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
).long()
- result[:, u - 1] = s + 1 + escape.first_lookahead_rewards_code
- ar_mask = (t >= u).long() * (t < u + state_len).long()
ar(result, ar_mask)
- # Put the lookahead reward to +1 for the current
- # iteration, sample the action and reward
- s = 1
- result[:, u - 1] = s + 1 + escape.first_lookahead_rewards_code
- ar_mask = (t >= u + state_len).long() * (t < u + state_len + 2).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)
- # Fix the previous lookahead rewards in a consistant state
- for v in range(0, u, it_len):
- # Extract the rewards
- r = result[:, range(v + state_len + 1 + it_len, u + it_len - 1, it_len)]
- r = r - escape.first_rewards_code - 1
- r = r.clamp(min=-1, max=1) # the reward is predicted hence can be weird
- a = r.min(dim=1).values
- b = r.max(dim=1).values
- s = (a < 0).long() * a + (a >= 0).long() * b
- result[:, v + state_len + 2] = (
- s + 1 + escape.first_lookahead_rewards_code
- )
+ 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 = greed.seq2episodes(
+ s[n : n + 1], self.height, self.width
+ )
+ str = greed.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 = greed.seq2episodes(result, self.height, self.width)
+ str = greed.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 = greed.seq2episodes(
+ result,
+ self.height,
+ self.width,
)
+ str = greed.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 = greed.seq2episodes(
+ result,
+ self.height,
+ self.width,
)
+ str = greed.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