Update.

diff --git a/main.py b/main.py
index 18b19db..35f02a3 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -853,7 +853,7 @@ def one_epoch(model, task, learning_rate):
 
     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 
-    log_string(f"train)perplexity {n_epoch} {train_perplexity}")
+    log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 
 
 ######################################################################
@@ -887,7 +887,7 @@ def run_tests(model, task, deterministic_synthesis):
         )
 
         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
-        log_string(f"test)perplexity {n_epoch} {test_perplexity}")
+        log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 
 
 ######################################################################
@@ -897,7 +897,29 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 
     one_epoch(model, task, learning_rate)
 
-    run_tests(model, task, deterministic_synthesis=True)
+    run_tests(model, task, deterministic_synthesis=False)
+
+    # --------------------------------------------
+
+    if n_epoch >= 3:
+        nb_required = 1000
+        kept = []
+
+        while sum([x.size(0) for x in kept]) < nb_required:
+            new_problems, nb_correct = task.create_new_problems(
+                n_epoch=n_epoch,
+                result_dir=args.result_dir,
+                logger=log_string,
+                nb=nb_required,
+                model=model,
+                nb_runs=10,
+            )
+
+            to_keep = new_problems[torch.logical_and(nb_correct >= 8, nb_correct < 10)]
+            log_string(f"keep {to_keep.size(0)} problems")
+            kept.append(to_keep)
+
+        new_problems = torch.cat(kept, dim=0)[:nb_required]
 
     # --------------------------------------------
 

diff --git a/tasks.py b/tasks.py
index b4e6f67..1b28108 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -395,133 +395,6 @@ class SandBox(Task):
                 # logger(f"wrote {filename}")
 
 
-######################################################################
-
-import world
-
-
-class World(Task):
-    def __init__(
-        self,
-        nb_train_samples,
-        nb_test_samples,
-        batch_size,
-        logger=None,
-        device=torch.device("cpu"),
-    ):
-        super().__init__()
-
-        self.batch_size = batch_size
-        self.device = device
-        self.height = 6
-        self.width = 8
-
-        self.train_input = world.generate(
-            nb_train_samples, height=self.height, width=self.width
-        )
-        self.train_ar_mask = (
-            (torch.arange(self.train_input.size(1)) > self.train_input.size(1) // 2)
-            .long()[None, :]
-            .expand_as(self.train_input)
-        )
-
-        self.test_input = world.generate(
-            nb_test_samples, height=self.height, width=self.width
-        )
-        self.test_ar_mask = (
-            (torch.arange(self.test_input.size(1)) > self.test_input.size(1) // 2)
-            .long()[None, :]
-            .expand_as(self.test_input)
-        )
-
-        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
-
-    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,
-            )
-
-            nb_total, nb_correct = (
-                input.size(0),
-                (input == result).long().min(dim=1).values.sum(),
-            )
-
-            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}%"
-        )
-
-        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
-
-        ##############################
-
-        input, ar_mask = self.test_input[:64], self.test_ar_mask[:64]
-        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,
-        )
-
-        img = world.sample2img(result.to("cpu"), self.height, self.width)
-
-        image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
-        torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=8, padding=2)
-        logger(f"wrote {image_name}")
-
-
 ######################################################################
 
 import picoclvr
@@ -2220,3 +2093,176 @@ class Greed(Task):
 
 
 ######################################################################
+######################################################################
+
+import world
+
+
+class World(Task):
+    def __init__(
+        self,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        logger=None,
+        device=torch.device("cpu"),
+    ):
+        super().__init__()
+
+        self.batch_size = batch_size
+        self.device = device
+        self.height = 6
+        self.width = 8
+
+        self.train_input = world.generate(
+            nb_train_samples, height=self.height, width=self.width
+        )
+        self.train_ar_mask = (
+            (torch.arange(self.train_input.size(1)) > self.train_input.size(1) // 2)
+            .long()[None, :]
+            .expand_as(self.train_input)
+        )
+
+        self.test_input = world.generate(
+            nb_test_samples, height=self.height, width=self.width
+        )
+        self.test_ar_mask = (
+            (torch.arange(self.test_input.size(1)) > self.test_input.size(1) // 2)
+            .long()[None, :]
+            .expand_as(self.test_input)
+        )
+
+        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
+
+    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,
+            )
+
+            nb_total, nb_correct = (
+                input.size(0),
+                (input == result).long().min(dim=1).values.sum(),
+            )
+
+            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}%"
+        )
+
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
+        ##############################
+
+        input, ar_mask = self.test_input[:64], self.test_ar_mask[:64]
+        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,
+        )
+
+        img = world.sample2img(result.to("cpu"), self.height, self.width)
+        image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
+        torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=8, padding=2)
+        logger(f"wrote {image_name}")
+
+    def create_new_problems(self, n_epoch, result_dir, logger, nb, model, nb_runs):
+        new_problems = torch.empty(
+            nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
+        )
+        ar_mask = torch.full(new_problems.size(), 1, device=self.device)
+
+        masked_inplace_autoregression(
+            model,
+            self.batch_size,
+            new_problems,
+            ar_mask,
+            deterministic_synthesis=False,
+            progress_bar_desc="new problems",
+            device=self.device,
+        )
+
+        img = world.sample2img(new_problems[:64].to("cpu"), self.height, self.width)
+        image_name = os.path.join(result_dir, f"world_new_{n_epoch:04d}.png")
+        torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=8, padding=2)
+        logger(f"wrote {image_name}")
+
+        nb_correct = torch.empty(nb, device=self.device, dtype=torch.int64)
+
+        for n in tqdm.tqdm(
+            range(new_problems.size(0)), dynamic_ncols=True, desc="checking problems"
+        ):
+            result = new_problems[n][None, :].expand(nb_runs, -1).clone()
+            ar_mask = (
+                (torch.arange(result.size(1), device=self.device) > result.size(1) // 2)
+                .long()[None, :]
+                .expand_as(result)
+            )
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis=False,
+                progress_bar_desc=None,
+                device=self.device,
+            )
+
+            nb_correct[n] = (
+                (new_problems[n][None, :] == result).long().min(dim=1).values.sum()
+            )
+
+        return new_problems, nb_correct