X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=0f44760093b1dbe9ccb9835663b4e1772d9a9b4b;hb=439c597d409c344283f8996f042daf79d3f24de2;hp=c5418b4bb616386d8f34e038ff96460dfafff585;hpb=e781d77071fa26f393f50451f91c70f4a0850ca5;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index c5418b4..0f44760 100755
--- a/tasks.py
+++ b/tasks.py
@@ -1021,6 +1021,163 @@ class Stack(Task):
         ##############################################################
 
 
+######################################################################
+
+import rpl
+
+
+class RPL(Task):
+    def tensorize(self, sequences):
+        len_max = max([len(x) for x in sequences])
+        return torch.cat(
+            [
+                torch.tensor(
+                    [
+                        [
+                            self.token2id[str(c)]
+                            for c in s + ["<nul>"] * (len_max - len(s))
+                        ]
+                        for s in sequences
+                    ]
+                )
+            ],
+            0,
+        ).to(self.device)
+
+    def seq2str(self, seq):
+        return " ".join([self.id2token[i] for i in seq])
+
+    def __init__(
+        self,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        nb_starting_values=3,
+        max_input=9,
+        prog_len=6,
+        nb_runs=5,
+        logger=None,
+        device=torch.device("cpu"),
+    ):
+        super().__init__()
+
+        self.batch_size = batch_size
+        self.device = device
+
+        train_sequences = [
+            rpl.generate(
+                nb_starting_values=nb_starting_values,
+                max_input=max_input,
+                prog_len=prog_len,
+                nb_runs=nb_runs,
+            )
+            for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
+        ]
+
+        test_sequences = [
+            rpl.generate(
+                nb_starting_values=nb_starting_values,
+                max_input=max_input,
+                prog_len=prog_len,
+                nb_runs=nb_runs,
+            )
+            for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
+        ]
+
+        symbols = list(
+            set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
+        )
+        val_max = max([x if type(x) is int else 0 for x in symbols])
+        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.train_input = self.tensorize(train_sequences)
+        self.test_input = self.tensorize(test_sequences)
+
+        if logger is not None:
+            for x in self.train_input[:10]:
+                end = (x != self.t_nul).nonzero().max().item() + 1
+                seq = [self.id2token[i.item()] for i in x[:end]]
+                s = " ".join(seq)
+                logger(f"example_seq {s}")
+
+        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
+        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
+        ):
+            last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
+            batch = batch[:, :last]
+            yield batch
+
+    def vocabulary_size(self):
+        return self.nb_codes
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        def compute_nb_errors(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)
+            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 x, y in zip(input, result):
+                seq = [self.id2token[i.item()] for i in y]
+                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_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])
+                    comment = "*" if nb_errors == 0 else "-"
+                    logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
+                    for start_stack, target_stack, result_stack, correct in stacks:
+                        comment = "*" if correct else "-"
+                        start_stack = " ".join([str(x) for x in start_stack])
+                        target_stack = " ".join([str(x) for x in target_stack])
+                        result_stack = " ".join([str(x) for x in result_stack])
+                        logger(
+                            f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
+                        )
+                    nb_to_log -= 1
+
+            return sum_nb_total, sum_nb_errors
+
+        test_nb_total, test_nb_errors = compute_nb_errors(
+            self.test_input[:1000], 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}%"
+        )
+
+
 ######################################################################