X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=181ac4475a2126df33206e2a4cea29a3f2d57833;hb=732349f7c16e43ff84380d28e021d671f2c56492;hp=066f1bbec05fcc0d65365823ea60931010e118cb;hpb=26ef53ee3769c3b6b92b85d15b5a43cbd18ede07;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 066f1bb..181ac44 100755
--- a/tasks.py
+++ b/tasks.py
@@ -14,10 +14,8 @@ from torch.nn import functional as F
 
 from mygpt import BracketedSequence
 
-try:
-    from graph import save_attention_image
-except ImportError:
-    save_attention_image = None
+# from graph import save_attention_image
+save_attention_image = None
 
 ######################################################################
 
@@ -72,6 +70,137 @@ class Task:
         pass
 
 
+class TaskFromFile(Task):
+    def tensorize(self, pairs):
+        len_max = max([len(x[0]) for x in pairs])
+
+        input = torch.cat(
+            [
+                torch.tensor(
+                    [
+                        [self.char2id[c] for c in s[0] + "#" * (len_max - len(s[0]))]
+                        for s in pairs
+                    ]
+                )
+            ],
+            0,
+        ).to("cpu")
+
+        pred_mask = torch.cat(
+            [
+                torch.tensor(
+                    [
+                        [int(c) for c in s[1] + "0" * (len_max - len(s[1]))]
+                        for s in pairs
+                    ]
+                )
+            ],
+            0,
+        ).to("cpu")
+
+        return input, pred_mask
+
+    # trim all the tensors in the tuple z to remove as much token from
+    # left and right in the first tensor. If z is a tuple, all its
+    # elements are trimed according to the triming for the first
+    def trim(self, z, token="#"):
+        n = self.char2id[token]
+        if type(z) == tuple:
+            x = z[0]
+            i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return tuple([t[:, a:b] for t in z])
+        else:
+            i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return z[:, a:b]
+
+    def __init__(
+        self,
+        filename,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        device=torch.device("cpu"),
+    ):
+        self.batch_size = batch_size
+        self.device = device
+
+        pairs = []
+        with open(filename, "r") as f:
+            for _ in range(nb_train_samples + nb_test_samples):
+                sequence = f.readline().strip()
+                pred_mask = f.readline().strip()
+                assert len(sequence) == len(pred_mask)
+                assert set(pred_mask) == {"0", "1", "2"}, f"{set(pred_mask)}"
+                pairs.append((sequence, pred_mask))
+
+        symbols = ["#"] + list(set("".join([x[0] for x in pairs])) - set(["#"]))
+        self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
+        self.id2char = dict([(n, c) for c, n in self.char2id.items()])
+
+        self.train_input, self.train_pred_masks = self.tensorize(
+            pairs[:nb_train_samples]
+        )
+        self.test_input, self.test_pred_masks = self.tensorize(pairs[nb_train_samples:])
+
+    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 self.trim(batch).to(self.device)
+
+    def vocabulary_size(self):
+        return len(self.char2id)
+
+    def tensor2str(self, t):
+        return ["".join([self.id2char[x.item()] for x in s]) for s in t]
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        correct = self.trim(self.test_input[:1000]).to(self.device)
+        result = correct.clone()
+        pred_mask = self.test_pred_masks[:1000, : result.size(1)].to(self.device)
+        ar_mask = (pred_mask > 0).long()
+        result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
+
+        logger(f"----------------------------------------------------------")
+
+        for e in self.tensor2str(result[:10]):
+            logger(f"test_before {e}")
+
+        masked_inplace_autoregression(
+            model,
+            self.batch_size,
+            result,
+            ar_mask,
+            deterministic_synthesis,
+            device=self.device,
+        )
+
+        logger(f"----------------------------------------------------------")
+
+        for e, c in zip(self.tensor2str(result[:10]), self.tensor2str(correct[:10])):
+            logger(f"test_after  {e}")
+            logger(f"correct     {c}")
+
+        logger(f"----------------------------------------------------------")
+
+        err_mask = (pred_mask == 2).long()
+        nb_total = err_mask.sum().item()
+        nb_correct = ((correct == result).long() * err_mask).sum().item()
+
+        logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
+        logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
+
+
 ####################
 
 import problems
@@ -111,13 +240,25 @@ class SandBox(Task):
         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)
-        )
+        assert self.nb_codes <= max_nb_codes
+        assert self.train_input.min() >= 0
+        assert self.test_input.min() >= 0
+        assert tuple(x.item() for x in self.train_ar_mask.unique()) in {
+            (0,),
+            (1,),
+            (0, 1),
+        }
+        assert tuple(x.item() for x in self.test_ar_mask.unique()) in {
+            (0,),
+            (1,),
+            (0, 1),
+        }
+
+        if logger is not None:
+            for s, a in zip(self.train_input[:100], self.train_ar_mask[:100]):
+                logger(f"train_sequences {self.problem.seq2str(s)}")
+                a = "".join(["01"[x.item()] for x in a])
+                logger(f"                {a}")
 
     def batches(self, split="train", nb_to_use=-1, desc=None):
         assert split in {"train", "test"}
@@ -151,17 +292,24 @@ class SandBox(Task):
                 device=self.device,
             )
 
+            log_ground_truth = ar_mask.min() == 0
+
             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)}"
-                    )
+                    if log_ground_truth:
+                        logger(
+                            f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
+                        )
+
+            nb_total, nb_correct = self.problem.compute_nb_correct(
+                input, ar_mask, result
+            )
 
-            nb_total = ar_mask.sum().item()
-            nb_correct = ((result == input).long() * ar_mask).sum().item()
+            # nb_total = ar_mask.sum().item()
+            # nb_correct = ((result == input).long() * ar_mask).sum().item()
 
             return nb_total, nb_correct
 
@@ -183,9 +331,7 @@ class SandBox(Task):
 
         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 
-        if save_attention_image is None:
-            logger("no save_attention_image (is pycairo installed?)")
-        else:
+        if save_attention_image is not None:
             for k in range(10):
                 ns = torch.randint(self.test_input.size(0), (1,)).item()
                 input = self.test_input[ns : ns + 1].clone()
@@ -193,30 +339,30 @@ class SandBox(Task):
                 with torch.autograd.no_grad():
                     t = model.training
                     model.eval()
-                    model.record_attention(True)
+                    # model.record_attention(True)
                     model(BracketedSequence(input))
                     model.train(t)
-                    ram = model.retrieve_attention()
-                    model.record_attention(False)
-
-                tokens_output = [c for c in self.problem.seq2str(input[0])]
-                tokens_input = ["n/a"] + tokens_output[:-1]
-                for n_head in range(ram[0].size(1)):
-                    filename = os.path.join(
-                        result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
-                    )
-                    attention_matrices = [m[0, n_head] for m in ram]
-                    save_attention_image(
-                        filename,
-                        tokens_input,
-                        tokens_output,
-                        attention_matrices,
-                        k_top=10,
-                        # min_total_attention=0.9,
-                        token_gap=12,
-                        layer_gap=50,
-                    )
-                    logger(f"wrote {filename}")
+                    # ram = model.retrieve_attention()
+                    # model.record_attention(False)
+
+                # tokens_output = [c for c in self.problem.seq2str(input[0])]
+                # tokens_input = ["n/a"] + tokens_output[:-1]
+                # for n_head in range(ram[0].size(1)):
+                # filename = os.path.join(
+                # result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
+                # )
+                # attention_matrices = [m[0, n_head] for m in ram]
+                # save_attention_image(
+                # filename,
+                # tokens_input,
+                # tokens_output,
+                # attention_matrices,
+                # k_top=10,
+                ##min_total_attention=0.9,
+                # token_gap=12,
+                # layer_gap=50,
+                # )
+                # logger(f"wrote {filename}")
 
 
 ######################################################################
@@ -1460,6 +1606,7 @@ class Grid(Task):
         nb_test_samples,
         batch_size,
         size,
+        fraction_play=0.0,
         logger=None,
         device=torch.device("cpu"),
     ):
@@ -1468,6 +1615,7 @@ class Grid(Task):
         self.device = device
         self.batch_size = batch_size
         self.grid_factory = grid.GridFactory(size=size)
+        self.fraction_play = fraction_play
 
         if logger is not None:
             logger(
@@ -1475,15 +1623,25 @@ class Grid(Task):
             )
 
         self.train_descr = self.grid_factory.generate_samples(
-            nb_train_samples, lambda r: tqdm.tqdm(r)
+            nb=nb_train_samples,
+            fraction_play=fraction_play,
+            progress_bar=lambda r: tqdm.tqdm(r),
         )
+
         self.test_descr = self.grid_factory.generate_samples(
-            nb_test_samples, lambda r: tqdm.tqdm(r)
+            nb=nb_test_samples, fraction_play=0.0, progress_bar=lambda r: tqdm.tqdm(r)
         )
 
+        if fraction_play > 0:
+            self.play_descr = self.grid_factory.generate_samples(
+                nb=25, fraction_play=1.0, progress_bar=lambda r: tqdm.tqdm(r)
+            )
+        else:
+            self.play_descr = []
+
         # Build the tokenizer
         tokens = set()
-        for d in [self.train_descr, self.test_descr]:
+        for d in [self.train_descr, self.test_descr, self.play_descr]:
             for s in d:
                 for t in s.strip().split(" "):
                     tokens.add(t)
@@ -1497,10 +1655,14 @@ class Grid(Task):
         self.t_nul = self.token2id["#"]
         self.t_true = self.token2id["true"]
         self.t_false = self.token2id["false"]
+        self.t_pipe = self.token2id["|"]
 
         # Tokenize the train and test sets
         self.train_input = self.str2tensor(self.train_descr)
         self.test_input = self.str2tensor(self.test_descr)
+        self.play_input = (
+            None if len(self.play_descr) == 0 else self.str2tensor(self.play_descr)
+        )
 
     def batches(self, split="train"):
         assert split in {"train", "test"}
@@ -1548,6 +1710,32 @@ class Grid(Task):
         logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
         logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
 
+        if self.play_input is not None:
+            result = self.play_input.clone()
+            ar_mask = (result == self.t_pipe).long().cumsum(dim=1).clamp(max=1)
+            result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
+
+            logger(f"----------------------------------------------------------")
+
+            for e in self.tensor2str(result[:10]):
+                logger(f"play_before {e}")
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                device=self.device,
+            )
+
+            logger(f"----------------------------------------------------------")
+
+            for e in self.tensor2str(result[:10]):
+                logger(f"play_after  {e}")
+
+            logger(f"----------------------------------------------------------")
+
 
 ######################################################################
 
@@ -1555,7 +1743,6 @@ import qmlp
 
 
 class QMLP(Task):
-
     ######################
 
     def __init__(
@@ -1563,6 +1750,7 @@ class QMLP(Task):
         nb_train_samples,
         nb_test_samples,
         batch_size,
+        result_dir,
         logger=None,
         device=torch.device("cpu"),
     ):
@@ -1577,19 +1765,31 @@ class QMLP(Task):
                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
             )
 
-        seq, q_test_set = generate_sequence_and_test_set(
-            nb_mlps=nb_train_samples+nb_test_samples,
+        seq, q_test_set, test_error = qmlp.generate_sequence_and_test_set(
+            nb_mlps=nb_train_samples + nb_test_samples,
             nb_samples=self.nb_samples_per_mlp,
             device=self.device,
             batch_size=64,
             nb_epochs=250,
-            nb_mlps_per_batch=1024
+            nb_mlps_per_batch=1024,
         )
 
         self.train_input = seq[:nb_train_samples]
         self.train_q_test_set = q_test_set[:nb_train_samples]
+        self.train_ref_test_errors = test_error[:nb_train_samples]
         self.test_input = seq[nb_train_samples:]
         self.test_q_test_set = q_test_set[nb_train_samples:]
+        self.test_ref_test_errors = test_error[nb_train_samples:]
+
+        filename = os.path.join(result_dir, f"train_errors_ref.dat")
+        with open(filename, "w") as f:
+            for e in self.train_ref_test_errors:
+                f.write(f"{e}\n")
+
+        filename = os.path.join(result_dir, f"test_errors_ref.dat")
+        with open(filename, "w") as f:
+            for e in self.test_ref_test_errors:
+                f.write(f"{e}\n")
 
         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 
@@ -1599,7 +1799,7 @@ class QMLP(Task):
         for batch in tqdm.tqdm(
             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
         ):
-            yield self.trim(batch)
+            yield batch
 
     def vocabulary_size(self):
         return self.nb_codes
@@ -1609,14 +1809,13 @@ class QMLP(Task):
     ):
         correct = self.test_input[:1000]
         result = correct.clone()
-        ar_mask = torch.arange(result.size(1)) > self.nb_samples_per_mlp * 3 + 1
+        ar_mask = (
+            torch.arange(result.size(1), device=result.device)
+            > self.nb_samples_per_mlp * 3 + 1
+        ).long()[None, :]
+        ar_mask = ar_mask.expand_as(result)
         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
 
-        logger(f"----------------------------------------------------------")
-
-        for e in self.tensor2str(result[:10]):
-            logger(f"test_before {e}")
-
         masked_inplace_autoregression(
             model,
             self.batch_size,
@@ -1626,18 +1825,14 @@ class QMLP(Task):
             device=self.device,
         )
 
-        logger(f"----------------------------------------------------------")
-
-        for e in self.tensor2str(result[:10]):
-            logger(f"test_after  {e}")
-
-        logger(f"----------------------------------------------------------")
-
-        q_train_set = result[:, : nb_samples * 3]
-        q_params = result[:, nb_samples * 3 + 1 :]
-        error_test = evaluate_q_params(q_params, q_test_set, nb_mlps_per_batch=17)
+        q_train_set = result[:, : self.nb_samples_per_mlp * 3]
+        q_params = result[:, self.nb_samples_per_mlp * 3 + 1 :]
+        error_test = qmlp.evaluate_q_params(q_params, self.test_q_test_set)
 
-        logger(f"{error_test=}")
+        filename = os.path.join(result_dir, f"test_errors_{n_epoch:04d}.dat")
+        with open(filename, "w") as f:
+            for e in error_test:
+                f.write(f"{e}\n")
 
 
 ######################################################################