Update.

[picoclvr.git] / tasks.py

diff --git a/tasks.py b/tasks.py
index 0f44760..0827a44 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1,5 +1,10 @@
 #!/usr/bin/env python
 
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
 import math, os, tqdm
 
 import torch, torchvision
@@ -29,7 +34,7 @@ def masked_inplace_autoregression(
             batches,
             dynamic_ncols=True,
             desc=progress_bar_desc,
-            # total=input.size(0) // batch_size,
+            total=(input.size(0) + batch_size - 1) // batch_size,
         )
 
     with torch.autograd.no_grad():
@@ -108,9 +113,7 @@ class ProblemLevel1(Problem):
         source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
         marker2 = torch.full((nb, 1), 11)
         result = operators.bmm(source[:, :, None]).squeeze(-1)
-        print(f"{nb_operators.dtype=} {marker1.dtype=}")
         sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
-        print(f"{sequences.size()=}")
         ar_mask = (sequences == 11).long()
         ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
         return sequences, ar_mask
@@ -1042,7 +1045,7 @@ class RPL(Task):
                 )
             ],
             0,
-        ).to(self.device)
+        )
 
     def seq2str(self, seq):
         return " ".join([self.id2token[i] for i in seq])
@@ -1067,6 +1070,7 @@ class RPL(Task):
         train_sequences = [
             rpl.generate(
                 nb_starting_values=nb_starting_values,
+                nb_result_values_max=4 * nb_starting_values,
                 max_input=max_input,
                 prog_len=prog_len,
                 nb_runs=nb_runs,
@@ -1077,6 +1081,7 @@ class RPL(Task):
         test_sequences = [
             rpl.generate(
                 nb_starting_values=nb_starting_values,
+                nb_result_values_max=4 * nb_starting_values,
                 max_input=max_input,
                 prog_len=prog_len,
                 nb_runs=nb_runs,
@@ -1091,17 +1096,20 @@ class RPL(Task):
         symbols = list(filter(lambda x: type(x) is str, symbols))
         symbols.sort()
         symbols += [str(n) for n in range(val_max + 1)]
-        print(f"{val_max=}")
         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
 
-        self.t_nul, self.t_prog = self.token2id["<nul>"], self.token2id["<prog>"]
+        self.t_nul = self.token2id["<nul>"]
+        self.t_prog = self.token2id["<prog>"]
+        self.t_input = self.token2id["<input>"]
+        self.t_output = self.token2id["<output>"]
 
         self.train_input = self.tensorize(train_sequences)
         self.test_input = self.tensorize(test_sequences)
 
         if logger is not None:
-            for x in self.train_input[:10]:
+            logger(f"value_max {val_max}")
+            for x in self.train_input[:25]:
                 end = (x != self.t_nul).nonzero().max().item() + 1
                 seq = [self.id2token[i.item()] for i in x[:end]]
                 s = " ".join(seq)
@@ -1120,7 +1128,7 @@ class RPL(Task):
             input.split(self.batch_size), dynamic_ncols=True, desc=desc
         ):
             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
-            batch = batch[:, :last]
+            batch = batch[:, :last].to(self.device)
             yield batch
 
     def vocabulary_size(self):
@@ -1129,7 +1137,8 @@ class RPL(Task):
     def produce_results(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis
     ):
-        def compute_nb_errors(input, nb_to_log=0):
+        # --------------------------------------------------------------------
+        def compute_nb_errors_prog(input, nb_to_log=0):
             result = input.clone()
             s = (result == self.t_prog).long()
             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
@@ -1169,8 +1178,52 @@ class RPL(Task):
 
             return sum_nb_total, sum_nb_errors
 
-        test_nb_total, test_nb_errors = compute_nb_errors(
-            self.test_input[:1000], nb_to_log=10
+        # --------------------------------------------------------------------
+        def compute_nb_errors_output(input, nb_to_log=0):
+            result = input.clone()
+            k = torch.arange(result.size(1), device=result.device)[None, :]
+            last_output_idx = ((result == self.t_output) * k).max(dim=1, keep_dim=True)
+            first_prog_idx = ((result == self.t_prog) * k).min(dim=1, keep_dim=True)
+            ar_mask = (k > last_output_idx).long() * (k < first_prog_idx)
+            result = (1 - ar_mask) * result + ar_mask * self.t_nul
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                device=self.device,
+            )
+
+            sum_nb_total, sum_nb_errors = 0, 0
+            for x, y in zip(input, result):
+                seq = [self.id2token[i.item()] for i in y]
+                sum_nb_total += 1
+                sum_nb_errors += 0 if (x - y).abs().max() == 0 else 1
+                if nb_to_log > 0:
+                    gt_seq = [self.id2token[i.item()] for i in x]
+                    _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
+                    gt_prog = " ".join([str(x) for x in gt_prog])
+                    prog = " ".join([str(x) for x in prog])
+                    comment = "*" if nb_errors == 0 else "-"
+                    logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
+                    for start_stack, target_stack, result_stack, correct in stacks:
+                        comment = "*" if correct else "-"
+                        start_stack = " ".join([str(x) for x in start_stack])
+                        target_stack = " ".join([str(x) for x in target_stack])
+                        result_stack = " ".join([str(x) for x in result_stack])
+                        logger(
+                            f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
+                        )
+                    nb_to_log -= 1
+
+            return sum_nb_total, sum_nb_errors
+
+        # --------------------------------------------------------------------
+
+        test_nb_total, test_nb_errors = compute_nb_errors_prog(
+            self.test_input[:1000].to(self.device), nb_to_log=10
         )
 
         logger(