X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=eef84af68d4393b2a0b9adb1efb00743fb0318ee;hb=5366dfd7bd57ec3298d1030f7d5327ff26bc5aad;hp=15d97b84ceedbc884fd64ab01288e796ce88dfd9;hpb=bf48dc69f7f57ad391481c8917570e35f661cc4a;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 15d97b8..eef84af 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():
@@ -60,6 +62,112 @@ class Task:
 
 ######################################################################
 
+
+class Problem:
+    def generate_sequences(self, nb):
+        pass
+
+    def log_performance(self, sequences, logger):
+        pass
+
+
+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))
+        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
+
+        # 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"),
+    ):
+        super().__init__()
+
+        self.batch_size = batch_size
+        self.device = device
+
+        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.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
+    ):
+
+        def compute_accuracy(input, ar_mask):
+            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 = 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(
+            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 +214,8 @@ class PicoCLVR(Task):
         pruner_train=None,
         pruner_eval=None,
     ):
+        super().__init__()
+
         def generate_descr(nb, cache_suffix, pruner):
             return picoclvr.generate(
                 nb,
@@ -294,6 +404,8 @@ 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 +476,8 @@ class Maze(Task):
         nb_walls,
         device=torch.device("cpu"),
     ):
+        super().__init__()
+
         self.batch_size = batch_size
         self.height = height
         self.width = width
@@ -535,6 +649,8 @@ class Snake(Task):
         prompt_length,
         device=torch.device("cpu"),
     ):
+        super().__init__()
+
         self.batch_size = batch_size
         self.height = height
         self.width = width
@@ -633,6 +749,8 @@ 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
@@ -780,6 +898,8 @@ class Expr(Task):
         batch_size,
         device=torch.device("cpu"),
     ):
+        super().__init__()
+
         self.batch_size = batch_size
         self.device = device
 
@@ -944,6 +1064,7 @@ class Expr(Task):
 
 
 ######################################################################
+
 import world
 
 
@@ -953,16 +1074,21 @@ class World(Task):
         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
 
         (
-            self.train_input,
-            self.train_actions,
-            self.test_input,
-            self.test_actions,
+            train_frames,
+            train_action_seq,
+            test_frames,
+            test_action_seq,
             self.frame2seq,
             self.seq2frame,
         ) = world.create_data_and_processors(
@@ -970,10 +1096,36 @@ class World(Task):
             nb_test_samples,
             mode="first_last",
             nb_steps=30,
-            nb_epochs=2,
+            nb_epochs=vqae_nb_epochs,
+            logger=logger,
+            device=device,
+            device_storage=device_storage,
         )
 
-        self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+        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)
+
+        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)
+        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
+        )
+
+        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"}
@@ -985,7 +1137,7 @@ class World(Task):
         for batch in tqdm.tqdm(
             input.split(self.batch_size), dynamic_ncols=True, desc=desc
         ):
-            yield batch
+            yield batch.to(self.device)
 
     def vocabulary_size(self):
         return self.nb_codes
@@ -993,7 +1145,47 @@ class World(Task):
     def produce_results(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis
     ):
-        pass
+        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))
+        print(f"{result.size()=}")
+
+        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}")
 
 
 ######################################################################