X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=a3d47f54451464470bd9d37710bda4f42c1ab4ce;hb=c9dbc3abf436df8af1379d04ab51159e821496f1;hp=eef84af68d4393b2a0b9adb1efb00743fb0318ee;hpb=5366dfd7bd57ec3298d1030f7d5327ff26bc5aad;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index eef84af..a3d47f5 100755
--- a/tasks.py
+++ b/tasks.py
@@ -67,36 +67,154 @@ class Problem:
     def generate_sequences(self, nb):
         pass
 
-    def log_performance(self, sequences, logger):
-        pass
+    def seq2str(self, seq):
+        return "[NOT IMPLEMENTED]"
+
+
+####################
 
 
-class ProblemByheart(Problem):
-    def __init__(self):
-        nb_seq, len_prompt, len_result = 100, 5, 5
-        self.seq = torch.randint(10, (nb_seq, len_prompt + 1 + len_result))
+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
 
     def generate_sequences(self, nb):
         sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
-        ar_mask = (sequences==10).long()
+        ar_mask = (sequences == 10).long()
+        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
+        return sequences, ar_mask
+
+
+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 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,
+        )
+        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
+        )
+        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)
+
+
+####################
+
+
+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+=$")])
+        self.id2char = dict([(n, c) for c, n in self.char2id.items()])
+
+    def tensorize(self, strings):
+        len_max = max([len(x) for x in strings])
+        return torch.cat(
+            [
+                torch.tensor(
+                    [
+                        [self.char2id[c] for c in s + "$" * (len_max - len(s))]
+                        for s in strings
+                    ]
+                )
+            ],
+            0,
+        )
+
+    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
 
-        # problems = [ProblemByheart()]
-        # nb_common_codes = 100
+    def seq2str(self, seq):
+        return "".join(self.id2char[x.item()] for x in seq)
+
+
+# class ProblemUnion(Problem):
+# problems = [ProblemByheart()]
+# nb_common_codes = 100
+
+# 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
 
-        # 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
+# for strain, stest in zip(train_seq, test_seq):
+# s = torch.cat((strain, stest), 0)
+
+####################
 
-        # for strain, stest in zip(train_seq, test_seq):
-            # s = torch.cat((strain, stest), 0)
 
 class SandBox(Task):
     def __init__(
@@ -107,17 +225,39 @@ class SandBox(Task):
         batch_size,
         logger=None,
         device=torch.device("cpu"),
+        max_nb_codes=1024,
     ):
         super().__init__()
 
         self.batch_size = batch_size
         self.device = device
+        self.problem = problem
+
+        self.train_input, self.train_ar_mask = self.problem.generate_sequences(
+            nb_train_samples
+        )
+        self.test_input, self.test_ar_mask = self.problem.generate_sequences(
+            nb_test_samples
+        )
 
-        self.train_input, self.train_ar_mask = problem.generate_sequences(nb_train_samples)
-        self.test_input, self.test_ar_mask = problem.generate_sequences(nb_test_samples)
+        self.train_input, self.train_ar_mask = self.train_input.to(
+            device
+        ), self.train_ar_mask.to(device)
+        self.test_input, self.test_ar_mask = self.test_input.to(
+            device
+        ), self.test_ar_mask.to(device)
 
         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)
+        )
+
     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
@@ -134,11 +274,12 @@ class SandBox(Task):
         return self.nb_codes
 
     def produce_results(
-        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
     ):
+        def compute_accuracy(input, ar_mask, logger=None):
+            input, ar_mask = input[:nmax], ar_mask[:nmax]
+            result = input.clone() * (1 - ar_mask)
 
-        def compute_accuracy(input, ar_mask):
-            result = input.clone() * (1-ar_mask)
             masked_inplace_autoregression(
                 model,
                 self.batch_size,
@@ -149,23 +290,37 @@ class SandBox(Task):
                 device=self.device,
             )
 
+            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)}"
+                    )
+
             nb_total = ar_mask.sum().item()
-            nb_correct = ((result==input).long() * ar_mask).sum().item()
+            nb_correct = ((result == input).long() * ar_mask).sum().item()
 
             return nb_total, nb_correct
 
-        train_nb_total, train_nb_correct = compute_accuracy(self.train_input, self.train_ar_mask)
+        train_nb_total, train_nb_correct = compute_accuracy(
+            self.train_input, self.train_ar_mask
+        )
 
         logger(
             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
         )
 
-        test_nb_total, test_nb_correct = compute_accuracy(self.test_input, self.test_ar_mask)
+        test_nb_total, test_nb_correct = compute_accuracy(
+            self.test_input, self.test_ar_mask, logger
+        )
 
         logger(
             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}%"
         )
 
+
 ######################################################################
 
 import picoclvr
@@ -866,6 +1021,124 @@ 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 __init__(
+        self,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        device=torch.device("cpu"),
+    ):
+        super().__init__()
+
+        self.batch_size = batch_size
+        self.device = device
+
+        train_sequences = [
+            rpl.generate()
+            for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
+        ]
+        test_sequences = [
+            rpl.generate() 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)
+
+        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,
+            )
+
+            if nb_to_log > 0:
+                for x in result[:nb_to_log]:
+                    s = " ".join([self.id2token[i.item()] for i in x])
+                    logger(f"check {n_epoch} {s}")
+                nb_to_log -= min(nb_to_log, result.size(0))
+
+            sum_nb_total, sum_nb_errors = 0, 0
+            for x in result:
+                seq = [self.id2token[i.item()] for i in x]
+                nb_total, nb_errors = rpl.check(seq)
+                sum_nb_total += nb_total
+                sum_nb_errors += nb_errors
+
+            return sum_nb_total, sum_nb_errors
+
+        test_nb_total, test_nb_errors = compute_nb_errors(self.test_input, 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}%"
+        )
+
+
 ######################################################################
 
 
@@ -1102,8 +1375,6 @@ class World(Task):
             device_storage=device_storage,
         )
 
-        print(f"{train_action_seq.size()=}")
-
         train_frame_seq = self.frame2seq(train_frames).to(device_storage)
         test_frame_seq = self.frame2seq(test_frames).to(device_storage)
 
@@ -1115,7 +1386,7 @@ class World(Task):
         self.nb_codes = nb_frame_codes + nb_action_codes
 
         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
-        print(f"{train_action_seq.device=} {nb_frame_codes.device=}")
+
         train_action_seq += nb_frame_codes
         self.train_input = torch.cat(
             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
@@ -1174,7 +1445,6 @@ class World(Task):
             (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
         )
         result = result.reshape(-1, result.size(-1))
-        print(f"{result.size()=}")
 
         frames = self.seq2frame(result)
         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")