# Written by Francois Fleuret <francois@fleuret.org>
-import math, os, tqdm
+import math, os, tqdm, warnings
import torch, torchvision
from mygpt import BracketedSequence
-try:
- from graph import save_attention_image
-except ImportError:
- save_attention_image = None
+# from graph import save_attention_image
+save_attention_image = None
######################################################################
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)
class Task:
- def batches(self, split="train"):
+ def batches(self, split="train", nb_to_use=-1, desc=None):
pass
def vocabulary_size(self):
pass
+class TaskFromFile(Task):
+ def tensorize(self, pairs, shuffle):
+ len_max = max([len(x[0]) for x in pairs])
+
+ input = torch.cat(
+ [
+ torch.tensor(
+ [
+ [self.char2id[c] for c in s[0] + "#" * (len_max - len(s[0]))]
+ for s in pairs
+ ]
+ )
+ ],
+ 0,
+ ).to("cpu")
+
+ pred_mask = torch.cat(
+ [
+ torch.tensor(
+ [
+ [int(c) for c in s[1] + "0" * (len_max - len(s[1]))]
+ for s in pairs
+ ]
+ )
+ ],
+ 0,
+ ).to("cpu")
+
+ if shuffle:
+ i = torch.randperm(input.size(0))
+ input = input[i].contiguous()
+ pred_mask = pred_mask[i].contiguous()
+
+ return input, pred_mask
+
+ # trim all the tensors in the tuple z to remove as much token from
+ # left and right in the first tensor. If z is a tuple, all its
+ # elements are trimed according to the triming for the first
+ def trim(self, z, token="#"):
+ n = self.char2id[token]
+ if type(z) == tuple:
+ x = z[0]
+ i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+ a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+ return tuple([t[:, a:b] for t in z])
+ else:
+ i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+ a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+ return z[:, a:b]
+
+ def __init__(
+ self,
+ train_filename,
+ test_filename,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ shuffle=False,
+ device=torch.device("cpu"),
+ ):
+ self.batch_size = batch_size
+ self.device = device
+
+ def read_file(filename, nb=-1):
+ pairs = []
+ with open(filename, "r") as f:
+ while True:
+ sequence = f.readline().strip()
+ if not sequence:
+ break
+ pred_mask = f.readline().strip()
+ assert len(sequence) == len(pred_mask)
+ assert set(pred_mask).issubset({"0", "1", "2"}), f"{set(pred_mask)}"
+ pairs.append((sequence, pred_mask))
+ if len(pairs) == nb:
+ break
+
+ if nb > 0:
+ pairs = pairs[:nb]
+ assert len(pairs) == nb
+
+ return pairs
+
+ train_pairs = read_file(train_filename, nb_train_samples)
+ test_pairs = read_file(test_filename, nb_test_samples)
+
+ symbols = ["#"] + list(
+ set("".join([x[0] for x in train_pairs + test_pairs])) - set(["#"])
+ )
+ self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
+ self.id2char = dict([(n, c) for c, n in self.char2id.items()])
+
+ self.train_input, self.train_pred_masks = self.tensorize(
+ train_pairs, shuffle=shuffle
+ )
+ self.test_input, self.test_pred_masks = self.tensorize(
+ test_pairs, shuffle=shuffle
+ )
+
+ 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
+ if nb_to_use > 0:
+ input = input[:nb_to_use]
+ if desc is None:
+ desc = f"epoch-{split}"
+ for batch in tqdm.tqdm(
+ input.split(self.batch_size), dynamic_ncols=True, desc=desc
+ ):
+ yield self.trim(batch).to(self.device)
+
+ def vocabulary_size(self):
+ return len(self.char2id)
+
+ def tensor2str(self, t):
+ return ["".join([self.id2char[x.item()] for x in s]) for s in t]
+
+ def produce_results(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis
+ ):
+ correct = self.trim(self.test_input[:1000]).to(self.device)
+ result = correct.clone()
+ pred_mask = self.test_pred_masks[:1000, : result.size(1)].to(self.device)
+ ar_mask = (pred_mask > 0).long()
+ result *= 1 - ar_mask # paraaaaanoiaaaaaaa
+
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:50]):
+ logger(f"test_before {e}")
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ logger(f"----------------------------------------------------------")
+
+ for e, c in zip(self.tensor2str(result[:50]), self.tensor2str(correct[:50])):
+ logger(f"test_after {e}")
+ logger(f"correct {c}")
+
+ logger(f"----------------------------------------------------------")
+
+ err_mask = (pred_mask == 2).long()
+ nb_total = err_mask.sum().item()
+ nb_correct = ((correct == result).long() * err_mask).sum().item()
+
+ logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
+ logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
+
+
####################
import problems
self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
# A bit of paranoia never hurts
- assert (
- self.nb_codes <= max_nb_codes
- and self.train_input.min() >= 0
- and self.test_input.min() >= 0
- and tuple(self.train_ar_mask.unique()) == (0, 1)
- and tuple(self.test_ar_mask.unique()) == (0, 1)
- )
+ assert self.nb_codes <= max_nb_codes
+ assert self.train_input.min() >= 0
+ assert self.test_input.min() >= 0
+ assert tuple(x.item() for x in self.train_ar_mask.unique()) in {
+ (0,),
+ (1,),
+ (0, 1),
+ }
+ assert tuple(x.item() for x in self.test_ar_mask.unique()) in {
+ (0,),
+ (1,),
+ (0, 1),
+ }
+
+ if logger is not None:
+ for s, a in zip(self.train_input[:100], self.train_ar_mask[:100]):
+ logger(f"train_sequences {self.problem.seq2str(s)}")
+ a = "".join(["01"[x.item()] for x in a])
+ logger(f" {a}")
def batches(self, split="train", nb_to_use=-1, desc=None):
assert split in {"train", "test"}
device=self.device,
)
+ log_ground_truth = ar_mask.min() == 0
+
if logger is not None:
for sp, st in zip(result[:10], input[:10]):
logger(
f"test_sequences {n_epoch} prediction {self.problem.seq2str(sp)}"
)
- logger(
- f" {n_epoch} ground truth {self.problem.seq2str(st)}"
- )
+ if log_ground_truth:
+ logger(
+ f" {n_epoch} ground truth {self.problem.seq2str(st)}"
+ )
- nb_total = ar_mask.sum().item()
- nb_correct = ((result == input).long() * ar_mask).sum().item()
+ nb_total, nb_correct = self.problem.compute_nb_correct(
+ input, ar_mask, result
+ )
+
+ # nb_total = ar_mask.sum().item()
+ # nb_correct = ((result == input).long() * ar_mask).sum().item()
return nb_total, nb_correct
logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
- if save_attention_image is None:
- logger("no save_attention_image (is pycairo installed?)")
- else:
+ if save_attention_image is not None:
for k in range(10):
ns = torch.randint(self.test_input.size(0), (1,)).item()
input = self.test_input[ns : ns + 1].clone()
with torch.autograd.no_grad():
t = model.training
model.eval()
- model.record_attention(True)
+ # model.record_attention(True)
model(BracketedSequence(input))
model.train(t)
- ram = model.retrieve_attention()
- model.record_attention(False)
-
- tokens_output = [c for c in self.problem.seq2str(input[0])]
- tokens_input = ["n/a"] + tokens_output[:-1]
- for n_head in range(ram[0].size(1)):
- filename = os.path.join(
- result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
- )
- attention_matrices = [m[0, n_head] for m in ram]
- save_attention_image(
- filename,
- tokens_input,
- tokens_output,
- attention_matrices,
- k_top=10,
- # min_total_attention=0.9,
- token_gap=12,
- layer_gap=50,
- )
- logger(f"wrote {filename}")
+ # ram = model.retrieve_attention()
+ # model.record_attention(False)
+
+ # tokens_output = [c for c in self.problem.seq2str(input[0])]
+ # tokens_input = ["n/a"] + tokens_output[:-1]
+ # for n_head in range(ram[0].size(1)):
+ # filename = os.path.join(
+ # result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
+ # )
+ # attention_matrices = [m[0, n_head] for m in ram]
+ # save_attention_image(
+ # filename,
+ # tokens_input,
+ # tokens_output,
+ # attention_matrices,
+ # k_top=10,
+ ##min_total_attention=0.9,
+ # token_gap=12,
+ # layer_gap=50,
+ # )
+ # logger(f"wrote {filename}")
######################################################################
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(
class Grid(Task):
# Make a tensor from a list of strings
- def tensorize(self, descr):
+ def str2tensor(self, descr):
token_descr = [s.strip().split(" ") for s in descr]
l = max([len(s) for s in token_descr])
token_descr = [s + ["#"] * (l - len(s)) for s in token_descr]
return torch.tensor(id_descr, device=self.device)
# Make a list of strings from a tensor
- def detensorize(self, x):
+ def tensor2str(self, x):
return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
# trim all the tensors in the tuple z to remove as much token from
nb_test_samples,
batch_size,
size,
+ fraction_play=0.0,
logger=None,
device=torch.device("cpu"),
):
self.device = device
self.batch_size = batch_size
self.grid_factory = grid.GridFactory(size=size)
+ self.fraction_play = fraction_play
if logger is not None:
logger(
)
self.train_descr = self.grid_factory.generate_samples(
- nb_train_samples, lambda r: tqdm.tqdm(r)
+ nb=nb_train_samples,
+ fraction_play=fraction_play,
+ progress_bar=lambda r: tqdm.tqdm(r),
)
+
self.test_descr = self.grid_factory.generate_samples(
- nb_test_samples, lambda r: tqdm.tqdm(r)
+ nb=nb_test_samples, fraction_play=0.0, progress_bar=lambda r: tqdm.tqdm(r)
)
+ if fraction_play > 0:
+ self.play_descr = self.grid_factory.generate_samples(
+ nb=25, fraction_play=1.0, progress_bar=lambda r: tqdm.tqdm(r)
+ )
+ else:
+ self.play_descr = []
+
# Build the tokenizer
tokens = set()
- for d in [self.train_descr, self.test_descr]:
+ for d in [self.train_descr, self.test_descr, self.play_descr]:
for s in d:
for t in s.strip().split(" "):
tokens.add(t)
self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
self.t_nul = self.token2id["#"]
- self.t_true = self.token2id["<true>"]
- self.t_false = self.token2id["<false>"]
+ self.t_true = self.token2id["true"]
+ self.t_false = self.token2id["false"]
+ # self.t_pipe = self.token2id["|"]
# Tokenize the train and test sets
- self.train_input = self.tensorize(self.train_descr)
- self.test_input = self.tensorize(self.test_descr)
+ self.train_input = self.str2tensor(self.train_descr)
+ self.test_input = self.str2tensor(self.test_descr)
+ self.play_input = (
+ 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(
correct = self.test_input[:1000]
result = correct.clone()
ar_mask = torch.logical_or(result == self.t_true, result == self.t_false).long()
- result *= 1 - ar_mask
+ result *= 1 - ar_mask # paraaaaanoiaaaaaaa
- for e in self.detensorize(result[:10]):
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
logger(f"test_before {e}")
masked_inplace_autoregression(
device=self.device,
)
- for e in self.detensorize(result[:10]):
- logger(f"test_after {e}")
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
+ logger(f"test_after {e}")
+
+ logger(f"----------------------------------------------------------")
nb_total = ar_mask.sum().item()
nb_correct = ((correct == result).long() * ar_mask).sum().item()
- logger(f"test_performance {nb_total=} {nb_correct=}")
- logger(f"main_test_accuracy {nb_correct / nb_total}")
+ logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
+ logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
+
+ if self.play_input is not None:
+ result = self.play_input.clone()
+ ar_mask = (result == self.t_pipe).long().cumsum(dim=1).clamp(max=1)
+ result *= 1 - ar_mask # paraaaaanoiaaaaaaa
+
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
+ logger(f"play_before {e}")
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
+ logger(f"play_after {e}")
+
+ logger(f"----------------------------------------------------------")
######################################################################
-import world
+import qmlp
-class World(Task):
+class QMLP(Task):
+ ######################
+
def __init__(
self,
nb_train_samples,
nb_test_samples,
batch_size,
- vqae_nb_epochs,
+ result_dir,
logger=None,
device=torch.device("cpu"),
- device_storage=torch.device("cpu"),
):
super().__init__()
- self.batch_size = batch_size
self.device = device
+ self.batch_size = batch_size
+ self.nb_samples_per_mlp = 256
- (
- train_frames,
- train_action_seq,
- test_frames,
- test_action_seq,
- self.frame2seq,
- self.seq2frame,
- ) = world.create_data_and_processors(
- nb_train_samples,
- nb_test_samples,
- mode="first_last",
- nb_steps=30,
- nb_epochs=vqae_nb_epochs,
- logger=logger,
- device=device,
- device_storage=device_storage,
+ if logger is not None:
+ logger(
+ f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
+ )
+
+ seq, q_test_set, test_error = qmlp.generate_sequence_and_test_set(
+ nb_mlps=nb_train_samples + nb_test_samples,
+ nb_samples=self.nb_samples_per_mlp,
+ device=self.device,
+ batch_size=64,
+ nb_epochs=250,
+ nb_mlps_per_batch=1024,
)
- train_frame_seq = self.frame2seq(train_frames).to(device_storage)
- test_frame_seq = self.frame2seq(test_frames).to(device_storage)
+ self.train_input = seq[:nb_train_samples]
+ self.train_q_test_set = q_test_set[:nb_train_samples]
+ self.train_ref_test_errors = test_error[:nb_train_samples]
+ self.test_input = seq[nb_train_samples:]
+ self.test_q_test_set = q_test_set[nb_train_samples:]
+ self.test_ref_test_errors = test_error[nb_train_samples:]
+
+ filename = os.path.join(result_dir, f"train_errors_ref.dat")
+ with open(filename, "w") as f:
+ for e in self.train_ref_test_errors:
+ f.write(f"{e}\n")
- nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
- nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
+ filename = os.path.join(result_dir, f"test_errors_ref.dat")
+ with open(filename, "w") as f:
+ for e in self.test_ref_test_errors:
+ f.write(f"{e}\n")
- self.len_frame_seq = train_frame_seq.size(1)
- self.len_action_seq = train_action_seq.size(1)
- self.nb_codes = nb_frame_codes + nb_action_codes
+ self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
- train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -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
+ for batch in tqdm.tqdm(
+ input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
+ ):
+ yield batch
+
+ def vocabulary_size(self):
+ return self.nb_codes
+
+ def produce_results(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis
+ ):
+ correct = self.test_input[:1000]
+ result = correct.clone()
+ ar_mask = (
+ torch.arange(result.size(1), device=result.device)
+ > self.nb_samples_per_mlp * 3 + 1
+ ).long()[None, :]
+ ar_mask = ar_mask.expand_as(result)
+ result *= 1 - ar_mask # paraaaaanoiaaaaaaa
- train_action_seq += nb_frame_codes
- self.train_input = torch.cat(
- (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
)
- test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
- test_action_seq += nb_frame_codes
- self.test_input = torch.cat(
- (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
+ q_train_set = result[:, : self.nb_samples_per_mlp * 3]
+ q_params = result[:, self.nb_samples_per_mlp * 3 + 1 :]
+ error_test = qmlp.evaluate_q_params(q_params, self.test_q_test_set)
+
+ filename = os.path.join(result_dir, f"test_errors_{n_epoch:04d}.dat")
+ with open(filename, "w") as f:
+ for e in error_test:
+ f.write(f"{e}\n")
+
+
+######################################################################
+
+import greed
+
+
+class Greed(Task):
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ height,
+ width,
+ T,
+ nb_walls,
+ nb_coins,
+ logger=None,
+ device=torch.device("cpu"),
+ ):
+ super().__init__()
+
+ self.batch_size = batch_size
+ self.device = device
+
+ 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 = 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):
for batch in tqdm.tqdm(
input.split(self.batch_size), dynamic_ncols=True, desc=desc
):
- yield batch.to(self.device)
+ yield self.wipe_lookahead_rewards(batch)
def vocabulary_size(self):
- return self.nb_codes
+ return self.world.nb_codes
+
+ def thinking_autoregression(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ snapshots = []
+
+ def ar(result, ar_mask, logit_biases=None):
+ ar_mask = ar_mask.expand_as(result)
+ result *= 1 - ar_mask
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis=deterministic_synthesis,
+ logit_biases=logit_biases,
+ device=self.device,
+ progress_bar_desc=None,
+ )
+ warnings.warn("keeping thinking snapshots", RuntimeWarning)
+ snapshots.append(result[:100].detach().clone())
+
+ # Generate iteration after iteration
+
+ 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
+ )
+
+ t = torch.arange(result.size(1), device=result.device)[None, :]
+
+ for u in tqdm.tqdm(
+ range(0, result.size(1), self.world.it_len),
+ desc="thinking",
+ ):
+ # 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)
+
+ # 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(snapshots[0].size(0)):
+ for s in snapshots:
+ lr, s, a, r = self.world.seq2episodes(
+ s[n : n + 1],
+ )
+ str = self.world.episodes2str(
+ lr, s, a, r, unicode=True, ansi_colors=True
+ )
+ f.write(str)
+ f.write("\n\n")
+
+ # Saving the generated sequences
+
+ 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:
+ f.write(str)
+ logger(f"wrote {filename}")
def produce_results(
- self, n_epoch, model, result_dir, logger, deterministic_synthesis
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
):
- k = torch.arange(
- 2 * self.len_frame_seq + self.len_action_seq, device=self.device
- )[None, :]
+ result = self.wipe_lookahead_rewards(self.test_input[:250].clone())
- input = self.test_input[:64].to(self.device)
- result = input.clone()
+ # Saving the ground truth
- ar_mask = (
- (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
+ lr, s, a, r = self.world.seq2episodes(
+ result,
)
- result *= 1 - ar_mask
+ 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:
+ f.write(str)
+ logger(f"wrote {filename}")
+
+ # Re-generating from the first frame
+
+ ar_mask = (
+ 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
masked_inplace_autoregression(
model,
device=self.device,
)
- seq_start = input[:, : self.len_frame_seq]
- seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
- seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
+ # Saving the generated sequences
- result = torch.cat(
- (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
+ lr, s, a, r = self.world.seq2episodes(
+ result,
)
- result = result.reshape(-1, result.size(-1))
+ str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
- frames = self.seq2frame(result)
- image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
- torchvision.utils.save_image(
- frames.float() / (world.Box.nb_rgb_levels - 1),
- image_name,
- nrow=12,
- padding=1,
- pad_value=0.0,
+ filename = os.path.join(result_dir, f"test_seq_{n_epoch:04d}.txt")
+ with open(filename, "w") as f:
+ f.write(str)
+ logger(f"wrote {filename}")
+
+ self.thinking_autoregression(
+ n_epoch, model, result_dir, logger, deterministic_synthesis, nmax
)
- logger(f"wrote {image_name}")
######################################################################
projects / picoclvr.git / blobdiff