X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=c0ad5ffb08683de9027cc71e8335636f7096af64;hb=HEAD;hp=0a4dd6fa2f880e93aecbbd494621fae26b7dcdbb;hpb=a291e213a152364b74e833200191c08a36451a90;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 0a4dd6f..443419e 100755
--- a/tasks.py
+++ b/tasks.py
@@ -1,12 +1,22 @@
 #!/usr/bin/env python
 
-import math, os, tqdm
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
+import math, os, tqdm, warnings
 
 import torch, torchvision
 
 from torch import nn
 from torch.nn import functional as F
 
+from mygpt import BracketedSequence
+
+# from graph import save_attention_image
+save_attention_image = None
+
 ######################################################################
 
 
@@ -17,6 +27,7 @@ def masked_inplace_autoregression(
     ar_mask,
     deterministic_synthesis,
     forbidden_tokens=None,
+    logit_biases=None,
     progress_bar_desc="autoregression",
     device=torch.device("cpu"),
 ):
@@ -29,7 +40,7 @@ def masked_inplace_autoregression(
             batches,
             dynamic_ncols=True,
             desc=progress_bar_desc,
-            # total=input.size(0) // batch_size,
+            total=(input.size(0) + batch_size - 1) // batch_size,
         )
 
     with torch.autograd.no_grad():
@@ -38,7 +49,11 @@ def masked_inplace_autoregression(
 
         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)
@@ -48,7 +63,7 @@ def masked_inplace_autoregression(
 
 
 class Task:
-    def batches(self, split="train"):
+    def batches(self, split="train", nb_to_use=-1, desc=None):
         pass
 
     def vocabulary_size(self):
@@ -60,161 +75,166 @@ class Task:
         pass
 
 
-######################################################################
-
-
-class Problem:
-    def generate_sequences(self, nb):
-        pass
-
-    def seq2str(self, seq):
-        return "[NOT IMPLEMENTED]"
+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()
 
-class ProblemLevel0(Problem):
-    def __init__(self, nb_sentences=100, len_prompt=5, len_result=5):
-        self.seq = torch.randint(10, (nb_sentences, len_prompt + 1 + len_result))
-        self.seq[:, len_prompt] = 10
+        return input, pred_mask
 
-    def generate_sequences(self, nb):
-        sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
-        ar_mask = (sequences == 10).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_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
 
-class ProblemLevel1(Problem):
-    def __init__(self, nb_operators=100, len_source=5, len_result=8):
-        self.len_source = len_source
-        self.len_result = len_result
-        self.len_nb_operator = int(math.log(nb_operators) / math.log(10)) + 1
-        self.operators = F.one_hot(
-            torch.rand(nb_operators, len_result, len_source).argmax(-1),
-            num_classes=len_source,
+        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()])
 
-    def generate_sequences(self, nb):
-        nb_operators = torch.randint(self.operators.size(0), (nb,))
-        operators = self.operators[nb_operators]
-        nb_operators = (
-            nb_operators[:, None]
-            // 10 ** torch.arange(self.len_nb_operator - 1, -1, -1)
-        ) % 10
-        marker1 = torch.full((nb, 1), 10)
-        # source = torch.randint(10, (nb, self.len_source))
-        source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
-        marker2 = torch.full((nb, 1), 11)
-        result = operators.bmm(source[:, :, None]).squeeze(-1)
-        print(f"{nb_operators.dtype=} {marker1.dtype=}")
-        sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
-        print(f"{sequences.size()=}")
-        ar_mask = (sequences == 11).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_mask
-
-    def seq2str(self, seq):
-        return "".join("0123456789|>"[x.item()] for x in seq)
-
-
-class ProblemLevel2(Problem):
-    def __init__(self, len_source=5, len_result=8):
-        self.len_source = len_source
-        self.len_result = len_result
-
-    def generate_sequences(self, nb):
-        operators = F.one_hot(
-            torch.rand(nb, self.len_result, self.len_source).argmax(-1),
-            num_classes=self.len_source,
+        self.train_input, self.train_pred_masks = self.tensorize(
+            train_pairs, shuffle=shuffle
         )
-        source1 = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
-        # source1 = torch.randint(10, (nb, self.len_source))
-        marker1 = torch.full((nb, 1), 10)
-        result1 = operators.bmm(source1[:, :, None]).squeeze(-1)
-        marker2 = torch.full((nb, 1), 11)
-        source2 = torch.randint(10, (nb, self.len_source))
-        marker3 = torch.full((nb, 1), 12)
-        result2 = operators.bmm(source2[:, :, None]).squeeze(-1)
-
-        sequences = torch.cat(
-            (source1, marker1, result1, marker2, source2, marker3, result2), 1
+        self.test_input, self.test_pred_masks = self.tensorize(
+            test_pairs, shuffle=shuffle
         )
-        ar_mask = (sequences == 12).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_mask
 
-    def seq2str(self, seq):
-        return "".join("0123456789>|~"[x.item()] for x in seq)
+    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
 
-class ProblemAddition(Problem):
-    def __init__(self, nb_digits=10, zero_padded=False, inverted_result=False):
-        self.nb_digits = nb_digits
-        self.zero_padded = zero_padded
-        self.inverted_result = inverted_result
-        self.char2id = dict([(c, n) for n, c in enumerate("0123456789+=$")])
-        self.id2char = dict([(n, c) for c, n in self.char2id.items()])
+        logger(f"----------------------------------------------------------")
 
-    def tensorize(self, strings):
-        len_max = max([len(x) for x in strings])
-        return torch.cat(
-            [
-                torch.tensor(
-                    [
-                        [self.char2id[c] for c in s + "$" * (len_max - len(s))]
-                        for s in strings
-                    ]
-                )
-            ],
-            0,
+        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,
         )
 
-    def generate_sequences(self, nb):
-        sequences = []
-        for k in range(nb):
-            a, b = torch.randint(10**self.nb_digits, (2,))
-            c = a + b
-            a, b, c = str(a.item()), str(b.item()), str(c.item())
-            if self.zero_padded:
-                a = "0" * (self.nb_digits - len(a)) + a
-                b = "0" * (self.nb_digits - len(b)) + b
-                c = "0" * (self.nb_digits + 1 - len(c)) + c
-            if self.inverted_result:
-                c = c[::-1]
-            sequences.append(f"{a}+{b}={c}$")
-
-        sequences = self.tensorize(sequences)
-        ar_mask = (sequences == self.char2id["="]).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_mask
+        logger(f"----------------------------------------------------------")
 
-    def seq2str(self, seq):
-        return "".join(self.id2char[x.item()] for x in seq)
+        for e, c in zip(self.tensor2str(result[:50]), self.tensor2str(correct[:50])):
+            logger(f"test_after  {e}")
+            logger(f"correct     {c}")
 
+        logger(f"----------------------------------------------------------")
 
-# class ProblemUnion(Problem):
-# problems = [ProblemByheart()]
-# nb_common_codes = 100
+        err_mask = (pred_mask == 2).long()
+        nb_total = err_mask.sum().item()
+        nb_correct = ((correct == result).long() * err_mask).sum().item()
 
-# def generate_sequences(nb_samples):
-# problem_indexes = torch.randint(len(problems), (nb_samples,))
-# nb_samples_per_problem = torch.one_hot(problem_indexes).sum(0)
-# print(f"{nb_samples_per_problem}")
-# all_seq = []
-# for nb, p in zip(nb_samples_per_problem, problems):
-# all_seq.append(p.generate_sequences(nb_samples_per_problem[nb]))
-# return all_seq
+        logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
+        logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
 
-# for strain, stest in zip(train_seq, test_seq):
-# s = torch.cat((strain, stest), 0)
 
 ####################
 
+import problems
+
 
 class SandBox(Task):
     def __init__(
@@ -250,13 +270,25 @@ class SandBox(Task):
         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"}
@@ -290,17 +322,24 @@ class SandBox(Task):
                 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, 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()
+            # nb_total = ar_mask.sum().item()
+            # nb_correct = ((result == input).long() * ar_mask).sum().item()
 
             return nb_total, nb_correct
 
@@ -320,6 +359,41 @@ class SandBox(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
+        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(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}")
+
 
 ######################################################################
 
@@ -415,7 +489,7 @@ class PicoCLVR(Task):
         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(
@@ -475,6 +549,10 @@ class PicoCLVR(Task):
             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
         )
 
+        logger(
+            f"main_test_accuracy {n_epoch} {1-nb_missing_properties/nb_requested_properties}"
+        )
+
     ######################################################################
 
     def produce_results(
@@ -676,15 +754,17 @@ class Maze(Task):
     def compute_error(
         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
     ):
+        model_device = next(model.parameters()).device
         nb_total, nb_correct = 0, 0
         count = torch.zeros(
             self.width * self.height,
             self.width * self.height,
-            device=self.device,
+            device=model_device,
             dtype=torch.int64,
         )
 
         for input in self.batches(split, nb_to_use):
+            input = input.to(model_device)
             result = input.clone()
             ar_mask = result.new_zeros(result.size())
             ar_mask[:, self.height * self.width :] = 1
@@ -745,6 +825,8 @@ class Maze(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         if count is not None:
             proportion_optimal = count.diagonal().sum().float() / count.sum()
             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
@@ -756,7 +838,7 @@ class Maze(Task):
                         eol = " " if j < count.size(1) - 1 else "\n"
                         f.write(f"{count[i,j]}{eol}")
 
-        input = self.test_input[:48]
+        input = self.test_input[:48].to(next(model.parameters()).device)
         result = input.clone()
         ar_mask = result.new_zeros(result.size())
         ar_mask[:, self.height * self.width :] = 1
@@ -884,6 +966,8 @@ class Snake(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
 
 ######################################################################
 
@@ -993,6 +1077,8 @@ class Stack(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         ##############################################################
         # Log a few generated sequences
         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
@@ -1014,6 +1100,34 @@ class Stack(Task):
             device=self.device,
         )
 
+        #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+        for label, input in [
+            ("train", self.train_input[:32]),
+            ("test", self.test_input[:32]),
+        ]:
+            output = model(BracketedSequence(input)).x
+            output = output.log_softmax(dim=-1)
+            filename = os.path.join(
+                result_dir, f"stack_with_crossentropy_{n_epoch:04d}_{label}.txt"
+            )
+            with open(filename, "w") as f:
+                for n in range(input.size(0)):
+                    s = stack.seq_to_str(
+                        input[n], nb_stacks=self.nb_stacks, nb_digits=self.nb_digits
+                    )
+                    for t, k, w in zip(range(input[n].size(0)), input[n], s.split(" ")):
+                        u = (
+                            " " * (10 - len(w))
+                            + w
+                            + " "
+                            + str(output[n][t][k].exp().item())
+                            + "\n"
+                        )
+                        f.write(u)
+                    f.write("\n")
+            logger(f"wrote {filename}")
+        #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+
         for n in range(result.size(0)):
             logger(
                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
@@ -1056,6 +1170,7 @@ class RPL(Task):
         max_input=9,
         prog_len=6,
         nb_runs=5,
+        no_prog=False,
         logger=None,
         device=torch.device("cpu"),
     ):
@@ -1063,10 +1178,12 @@ class RPL(Task):
 
         self.batch_size = batch_size
         self.device = device
+        self.no_prog = no_prog
 
         train_sequences = [
             rpl.generate(
                 nb_starting_values=nb_starting_values,
+                nb_result_values_max=4 * nb_starting_values,
                 max_input=max_input,
                 prog_len=prog_len,
                 nb_runs=nb_runs,
@@ -1077,6 +1194,7 @@ class RPL(Task):
         test_sequences = [
             rpl.generate(
                 nb_starting_values=nb_starting_values,
+                nb_result_values_max=4 * nb_starting_values,
                 max_input=max_input,
                 prog_len=prog_len,
                 nb_runs=nb_runs,
@@ -1091,16 +1209,49 @@ class RPL(Task):
         symbols = list(filter(lambda x: type(x) is str, symbols))
         symbols.sort()
         symbols += [str(n) for n in range(val_max + 1)]
-        print(f"{val_max=}")
         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
 
-        self.t_nul, self.t_prog = self.token2id["<nul>"], self.token2id["<prog>"]
+        self.t_nul = self.token2id["<nul>"]
+        self.t_input = self.token2id["<in>"]
+        self.t_output = self.token2id["<out>"]
+        self.t_prog = self.token2id["<prg>"]
+        self.t_end = self.token2id["<end>"]
 
         self.train_input = self.tensorize(train_sequences)
         self.test_input = self.tensorize(test_sequences)
 
+        if no_prog:
+            # Excise the program from every train and test example
+            k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
+                None, :
+            ]
+            p = (
+                ((self.train_input == self.t_prog).long() * k)
+                .max(1, keepdim=True)
+                .values
+            )
+            self.train_input = (
+                self.train_input * (k <= p).long()
+                + self.t_end * (k == p + 1).long()
+                + self.t_nul * (k > p + 1).long()
+            )
+            k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
+                None, :
+            ]
+            p = (
+                ((self.test_input == self.t_prog).long() * k)
+                .max(1, keepdim=True)
+                .values
+            )
+            self.test_input = (
+                self.test_input * (k <= p).long()
+                + self.t_end * (k == p + 1).long()
+                + self.t_nul * (k > p + 1).long()
+            )
+
         if logger is not None:
+            logger(f"value_max {val_max}")
             for x in self.train_input[:25]:
                 end = (x != self.t_nul).nonzero().max().item() + 1
                 seq = [self.id2token[i.item()] for i in x[:end]]
@@ -1130,7 +1281,7 @@ class RPL(Task):
         self, n_epoch, model, result_dir, logger, deterministic_synthesis
     ):
         # --------------------------------------------------------------------
-        def compute_nb_errors(input, nb_to_log=0):
+        def compute_nb_errors_prog(input, nb_to_log=0):
             result = input.clone()
             s = (result == self.t_prog).long()
             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
@@ -1146,13 +1297,13 @@ class RPL(Task):
             )
 
             sum_nb_total, sum_nb_errors = 0, 0
-            for x, y in zip(input, result):
-                seq = [self.id2token[i.item()] for i in y]
+            for one_input, one_result in zip(input, result):
+                seq = [self.id2token[i.item()] for i in one_result]
                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
                 sum_nb_total += 1
                 sum_nb_errors += 0 if nb_errors == 0 else 1
                 if nb_to_log > 0:
-                    gt_seq = [self.id2token[i.item()] for i in x]
+                    gt_seq = [self.id2token[i.item()] for i in one_input]
                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
                     gt_prog = " ".join([str(x) for x in gt_prog])
                     prog = " ".join([str(x) for x in prog])
@@ -1171,15 +1322,109 @@ class RPL(Task):
             return sum_nb_total, sum_nb_errors
 
         # --------------------------------------------------------------------
+        def compute_nb_errors_output(input, nb_to_log=0):
+            result = input.clone()
+            k = torch.arange(result.size(1), device=result.device)[None, :]
+            last_output_idx = (
+                ((result == self.t_output) * k).max(dim=1, keepdim=True).values
+            )
+            first_prog_idx = (
+                ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
+            )
+            ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
+            result = (1 - ar_mask) * result + ar_mask * self.t_nul
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                device=self.device,
+            )
+
+            sum_nb_total, sum_nb_errors = 0, 0
+            for one_input, one_result, i, j in zip(
+                input, result, last_output_idx, first_prog_idx
+            ):
+                seq = [self.id2token[i.item()] for i in one_result]
+                sum_nb_total += 1
+                correct = (one_input - one_result).abs().max() == 0
+                sum_nb_errors += 0 if correct else 1
+                if nb_to_log > 0:
+                    result_stack = [
+                        self.id2token[i.item()] for i in one_result[i : j + 1]
+                    ]
+                    target_stack = [
+                        self.id2token[i.item()] for i in one_input[i : j + 1]
+                    ]
+                    comment = "*" if correct else "-"
+                    result_stack = " ".join([str(x) for x in result_stack])
+                    target_stack = " ".join([str(x) for x in target_stack])
+                    logger(
+                        f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
+                    )
+                    nb_to_log -= 1
+
+            return sum_nb_total, sum_nb_errors
+
+        # --------------------------------------------------------------------
+
+        if not self.no_prog:
+            test_nb_total, test_nb_errors = compute_nb_errors_prog(
+                self.test_input[:1000].to(self.device), nb_to_log=10
+            )
+
+            logger(
+                f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
+            )
+
+            logger(f"main_test_accuracy {n_epoch} {1-test_nb_errors/test_nb_total}")
 
-        test_nb_total, test_nb_errors = compute_nb_errors(
+        test_nb_total, test_nb_errors = compute_nb_errors_output(
             self.test_input[:1000].to(self.device), nb_to_log=10
         )
 
         logger(
-            f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
+            f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
         )
 
+        if save_attention_image is None:
+            logger("no save_attention_image (is pycairo installed?)")
+        else:
+            ns = torch.randint(self.test_input.size(0), (1,)).item()
+            input = self.test_input[ns : ns + 1].clone()
+            last = (input != self.t_nul).max(0).values.nonzero().max() + 3
+            input = input[:, :last].to(self.device)
+
+            with torch.autograd.no_grad():
+                t = model.training
+                model.eval()
+                model.record_attention(True)
+                model(BracketedSequence(input))
+                model.train(t)
+                ram = model.retrieve_attention()
+                model.record_attention(False)
+
+            tokens_output = [self.id2token[i.item()] for i in 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"rpl_attention_{n_epoch}_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}")
+
 
 ######################################################################
 
@@ -1334,6 +1579,8 @@ class Expr(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         nb_total = test_nb_delta.sum() + test_nb_missed
         for d in range(test_nb_delta.size(0)):
             logger(
@@ -1380,64 +1627,317 @@ class Expr(Task):
 
 ######################################################################
 
-import world
+import grid
 
 
-class World(Task):
+class Grid(Task):
+    # Make a tensor from a list of strings
+    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]
+        id_descr = [[self.token2id[u] for u in s] for s in token_descr]
+        return torch.tensor(id_descr, device=self.device)
+
+    # Make a list of strings from a tensor
+    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
+    # 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.token2id[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,
         nb_train_samples,
         nb_test_samples,
         batch_size,
-        vqae_nb_epochs,
+        size,
+        fraction_play=0.0,
         logger=None,
         device=torch.device("cpu"),
-        device_storage=torch.device("cpu"),
     ):
         super().__init__()
 
+        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(
+                f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
+            )
+
+        self.train_descr = self.grid_factory.generate_samples(
+            nb=nb_train_samples,
+            fraction_play=fraction_play,
+            progress_bar=lambda r: tqdm.tqdm(r),
+        )
+
+        self.test_descr = self.grid_factory.generate_samples(
+            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, self.play_descr]:
+            for s in d:
+                for t in s.strip().split(" "):
+                    tokens.add(t)
+        # make this set a sorted list to get the same tensors given
+        # the same descr
+        tokens = list(tokens)
+        tokens.sort()
+        tokens = ["#"] + tokens
+        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_pipe = self.token2id["|"]
+
+        # Tokenize the train and test sets
+        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", 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 self.trim(batch)
+
+    def vocabulary_size(self):
+        return len(self.token2id)
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        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  # paraaaaanoiaaaaaaa
+
+        logger(f"----------------------------------------------------------")
+
+        for e in self.tensor2str(result[:10]):
+            logger(f"test_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"test_after  {e}")
+
+        logger(f"----------------------------------------------------------")
+
+        nb_total = ar_mask.sum().item()
+        nb_correct = ((correct == result).long() * ar_mask).sum().item()
+
+        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 qmlp
+
+
+class QMLP(Task):
+    ######################
+
+    def __init__(
+        self,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        result_dir,
+        logger=None,
+        device=torch.device("cpu"),
+    ):
+        super().__init__()
+
         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.nb_codes = max(self.train_input.max(), self.test_input.max()) + 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
 
-        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
+    def vocabulary_size(self):
+        return self.nb_codes
 
-        train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
+    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):
@@ -1450,25 +1950,121 @@ class World(Task):
         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,
@@ -1479,25 +2075,21 @@ class World(Task):
             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}")
 
 
 ######################################################################