X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=0ab18233a72e27b5b0b02e96a21abd5009d37ad6;hb=ce969e8372fb161d86be29042a20b044ee6efe2a;hp=b2f7d7dc5f750610333d03b7da6c183d83ff7a7a;hpb=16cb07f99cf770fb4e97824f874a68cbddd4c1cf;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index b2f7d7d..0ab1823 100755
--- a/tasks.py
+++ b/tasks.py
@@ -181,6 +181,43 @@ class SandBox(Task):
             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}")
+
+        if save_attention_image is None:
+            logger("no save_attention_image (is pycairo installed?)")
+        else:
+            for k in range(10):
+                ns = torch.randint(self.test_input.size(0), (1,)).item()
+                input = self.test_input[ns : ns + 1].clone()
+
+                with torch.autograd.no_grad():
+                    t = model.training
+                    model.eval()
+                    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}")
+
 
 ######################################################################
 
@@ -336,6 +373,10 @@ class PicoCLVR(Task):
             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
         )
 
+        logger(
+            f"main_test_accuracy {n_epoch} {1-nb_missing_properties/nb_requested_properties}"
+        )
+
     ######################################################################
 
     def produce_results(
@@ -606,6 +647,8 @@ class Maze(Task):
             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}")
+
         if count is not None:
             proportion_optimal = count.diagonal().sum().float() / count.sum()
             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
@@ -745,6 +788,8 @@ class Snake(Task):
             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}")
+
 
 ######################################################################
 
@@ -854,6 +899,8 @@ class Stack(Task):
             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}")
+
         ##############################################################
         # Log a few generated sequences
         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
@@ -1126,6 +1173,8 @@ class RPL(Task):
                 f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
             )
 
+            logger(f"main_test_accuracy {n_epoch} {1-test_nb_errors/test_nb_total}")
+
         test_nb_total, test_nb_errors = compute_nb_errors_output(
             self.test_input[:1000].to(self.device), nb_to_log=10
         )
@@ -1134,7 +1183,9 @@ class RPL(Task):
             f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
         )
 
-        if save_attention_image is not None:
+        if save_attention_image is None:
+            logger("no save_attention_image (is pycairo installed?)")
+        else:
             ns = torch.randint(self.test_input.size(0), (1,)).item()
             input = self.test_input[ns : ns + 1].clone()
             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
@@ -1322,6 +1373,8 @@ class Expr(Task):
             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}")
+
         nb_total = test_nb_delta.sum() + test_nb_missed
         for d in range(test_nb_delta.size(0)):
             logger(
@@ -1366,6 +1419,131 @@ class Expr(Task):
         ##############################################################
 
 
+######################################################################
+
+import grid
+
+
+class Grid(Task):
+    # Make a tensor from a list of strings
+    def tensorize(self, descr):
+        token_descr = [s.strip().split(" ") for s in descr]
+        l = max([len(s) for s in token_descr])
+        token_descr = [s + ["#"] * (l - len(s)) for s in token_descr]
+        id_descr = [[self.token2id[u] for u in s] for s in token_descr]
+        return torch.tensor(id_descr, device=self.device)
+
+    # Make a list of strings from a tensor
+    def detensorize(self, x):
+        return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
+
+    # 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.token2id[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,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        height,
+        width,
+        logger=None,
+        device=torch.device("cpu"),
+    ):
+        super().__init__()
+
+        self.device = device
+        self.batch_size = batch_size
+        self.grid_factory = grid.GridFactory(height=height, width=width)
+
+        if logger is not None:
+            logger(
+                f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
+            )
+
+        self.train_descr = self.grid_factory.generate_samples(
+            nb_train_samples, lambda r: tqdm.tqdm(r)
+        )
+        self.test_descr = self.grid_factory.generate_samples(
+            nb_test_samples, lambda r: tqdm.tqdm(r)
+        )
+
+        # Build the tokenizer
+        tokens = set()
+        for d in [self.train_descr, self.test_descr]:
+            for s in d:
+                for t in s.strip().split(" "):
+                    tokens.add(t)
+        # make this set a sorted list to get the same tensors given
+        # the same descr
+        tokens = list(tokens)
+        tokens.sort()
+        tokens = ["#"] + tokens
+        self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
+        self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
+        self.t_nul = self.token2id["#"]
+        self.t_true = self.token2id["<true>"]
+        self.t_false = self.token2id["<false>"]
+
+        # Tokenize the train and test sets
+        self.train_input = self.tensorize(self.train_descr)
+        self.test_input = self.tensorize(self.test_descr)
+
+    def batches(self, split="train"):
+        assert split in {"train", "test"}
+        input = self.train_input if split == "train" else self.test_input
+        for batch in tqdm.tqdm(
+            input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
+        ):
+            yield self.trim(batch)
+
+    def vocabulary_size(self):
+        return len(self.token2id)
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        correct = self.test_input[:1000]
+        result = correct.clone()
+        ar_mask = torch.logical_or(result == self.t_true, result == self.t_false).long()
+        result *= 1 - ar_mask
+
+        for e in self.detensorize(result[:10]):
+            logger(f"test_before {e}")
+
+        masked_inplace_autoregression(
+            model,
+            self.batch_size,
+            result,
+            ar_mask,
+            deterministic_synthesis,
+            device=self.device,
+        )
+
+        for e in self.detensorize(result[:10]):
+            logger(f"test_after {e}")
+
+        nb_total = ar_mask.sum().item()
+        nb_correct = ((correct == result).long() * ar_mask).sum().item()
+
+        logger(f"test_performance {nb_total=} {nb_correct=}")
+        logger(f"main_test_accuracy {nb_correct / nb_total}")
+
+
 ######################################################################
 
 import world