X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=332d6c5a93f4768dfd3e409389887baf3f8d2601;hb=ead4b8e4edd29578c01501d168e416b47fa4047b;hp=82d965b040becefc4f4933e055fc9e19d3a6976e;hpb=994d2408781ebaed6da16b10b2b3ebedeff82756;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 82d965b..332d6c5 100755
--- a/tasks.py
+++ b/tasks.py
@@ -20,6 +20,8 @@ def masked_inplace_autoregression(
     progress_bar_desc="autoregression",
     device=torch.device("cpu"),
 ):
+    assert input.size() == ar_mask.size()
+
     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
 
     if progress_bar_desc is not None:
@@ -27,7 +29,7 @@ def masked_inplace_autoregression(
             batches,
             dynamic_ncols=True,
             desc=progress_bar_desc,
-            total=input.size(0) // batch_size,
+            # total=input.size(0) // batch_size,
         )
 
     with torch.autograd.no_grad():
@@ -58,6 +60,236 @@ class Task:
         pass
 
 
+######################################################################
+
+
+class Problem:
+    def generate_sequences(self, nb):
+        pass
+
+    def seq2str(self, seq):
+        return "[NOT IMPLEMENTED]"
+
+
+####################
+
+
+class ProblemLevel0(Problem):
+    def __init__(self, nb_sentences=100, len_prompt=5, len_result=5):
+        self.seq = torch.randint(10, (nb_seq, 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 = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
+        return sequences, ar_mask
+
+
+class ProblemLevel1(Problem):
+    def __init__(self, nb_operators=100, len_prompt=5, len_result=8):
+        self.len_prompt = len_prompt
+        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_prompt).argmax(-1),
+            num_classes=len_prompt,
+        )
+
+    def generate_sequences(self, nb):
+        a = self.len_nb_operator
+        b = a + 1 + self.len_prompt
+        sequences = torch.empty(nb, b + 1 + self.len_result, dtype=torch.int64)
+        nb_operators = torch.randint(self.operators.size(0), (nb,))
+        sequences[:, :a] = (nb_operators[:, None] / 10 ** torch.arange(a)) % 10
+        sequences[:, a] = 10
+        sequences[:, a + 1 : b] = torch.randint(10, (nb, b - a - 1))
+        sequences[:, b] = 11
+
+        o = self.operators[nb_operators]
+        p = sequences[:, a + 1 : b]
+        print(f"{o.size()=} {p.size()=} {sequences[:,b+1:].size()=}")
+        sequences[:, b + 1 :] = o.bmm(p[:, :, None]).squeeze(-1)
+        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(self.id2char[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
+
+    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
+
+# for strain, stest in zip(train_seq, test_seq):
+# s = torch.cat((strain, stest), 0)
+
+####################
+
+
+class SandBox(Task):
+    def __init__(
+        self,
+        problem,
+        nb_train_samples,
+        nb_test_samples,
+        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 = 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
+        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 batch
+
+    def vocabulary_size(self):
+        return self.nb_codes
+
+    def produce_results(
+        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)
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                progress_bar_desc=None,
+                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()
+
+            return nb_total, nb_correct
+
+        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, 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
@@ -106,6 +338,8 @@ class PicoCLVR(Task):
         pruner_train=None,
         pruner_eval=None,
     ):
+        super().__init__()
+
         def generate_descr(nb, cache_suffix, pruner):
             return picoclvr.generate(
                 nb,
@@ -294,6 +528,8 @@ class MNIST(Task):
     def __init__(
         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
     ):
+        super().__init__()
+
         self.nb_train_samples = (nb_train_samples,)
         self.nb_test_samples = (nb_test_samples,)
         self.batch_size = batch_size
@@ -364,6 +600,8 @@ class Maze(Task):
         nb_walls,
         device=torch.device("cpu"),
     ):
+        super().__init__()
+
         self.batch_size = batch_size
         self.height = height
         self.width = width
@@ -535,6 +773,8 @@ class Snake(Task):
         prompt_length,
         device=torch.device("cpu"),
     ):
+        super().__init__()
+
         self.batch_size = batch_size
         self.height = height
         self.width = width
@@ -590,8 +830,6 @@ class Snake(Task):
             )
             result *= 1 - ar_mask
 
-            # snake.solver(result,ar_mask)
-
             masked_inplace_autoregression(
                 model,
                 self.batch_size,
@@ -605,19 +843,8 @@ class Snake(Task):
 
             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 
-            # nb_total = result.size(0)
-            # nb_correct = ((result - input).abs().sum(1) == 0).sum()
-
             return nb_total, nb_correct
 
-        # train_nb_total, train_nb_correct = compute_nb_correct(
-        # self.train_input, self.train_prior_visits
-        # )
-
-        # logger(
-        # f"accuracy_train 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_nb_correct(
             self.test_input[:1000], self.test_prior_visits[:1000]
         )
@@ -646,6 +873,8 @@ class Stack(Task):
         fraction_values_for_train=None,
         device=torch.device("cpu"),
     ):
+        super().__init__()
+
         self.batch_size = batch_size
         self.nb_steps = nb_steps
         self.nb_stacks = nb_stacks
@@ -793,6 +1022,8 @@ class Expr(Task):
         batch_size,
         device=torch.device("cpu"),
     ):
+        super().__init__()
+
         self.batch_size = batch_size
         self.device = device
 
@@ -880,17 +1111,22 @@ class Expr(Task):
             values_input = expr.extract_results([self.seq2str(s) for s in input])
             values_result = expr.extract_results([self.seq2str(s) for s in result])
 
-            for i, r in zip(values_input, values_result):
-                for n, vi in i.items():
-                    vr = r.get(n)
-                    if vr is None or vr < 0:
-                        nb_missed += 1
-                    else:
-                        d = abs(vr - vi)
-                        if d >= nb_delta.size(0):
+            filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
+
+            with open(filename, "w") as f:
+                for i, r in zip(values_input, values_result):
+                    for n, vi in i.items():
+                        vr = r.get(n)
+                        f.write(f"{vi} {-1 if vr is None else vr}\n")
+
+                        if vr is None or vr < 0:
                             nb_missed += 1
                         else:
-                            nb_delta[d] += 1
+                            d = abs(vr - vi)
+                            if d >= nb_delta.size(0):
+                                nb_missed += 1
+                            else:
+                                nb_delta[d] += 1
 
             ######################################################################
 
@@ -952,3 +1188,125 @@ class Expr(Task):
 
 
 ######################################################################
+
+import world
+
+
+class World(Task):
+    def __init__(
+        self,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        vqae_nb_epochs,
+        logger=None,
+        device=torch.device("cpu"),
+        device_storage=torch.device("cpu"),
+    ):
+        super().__init__()
+
+        self.batch_size = batch_size
+        self.device = device
+
+        (
+            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,
+        )
+
+        train_frame_seq = self.frame2seq(train_frames).to(device_storage)
+        test_frame_seq = self.frame2seq(test_frames).to(device_storage)
+
+        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
+
+        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
+
+        train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
+
+        train_action_seq += nb_frame_codes
+        self.train_input = torch.cat(
+            (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
+        )
+
+        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
+        )
+
+    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 batch.to(self.device)
+
+    def vocabulary_size(self):
+        return self.nb_codes
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        k = torch.arange(
+            2 * self.len_frame_seq + self.len_action_seq, device=self.device
+        )[None, :]
+
+        input = self.test_input[:64].to(self.device)
+        result = input.clone()
+
+        ar_mask = (
+            (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
+        )
+        result *= 1 - ar_mask
+
+        masked_inplace_autoregression(
+            model,
+            self.batch_size,
+            result,
+            ar_mask,
+            deterministic_synthesis,
+            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 :]
+
+        result = torch.cat(
+            (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
+        )
+        result = result.reshape(-1, result.size(-1))
+
+        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,
+        )
+        logger(f"wrote {image_name}")
+
+
+######################################################################