X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=da39a830e3b4d899e2f3521f73444fa9cdd8c36b;hb=0c6d29f73e35adbbaab1263de439f73efa98d99e;hp=0fac0a70df09a32bbe2e688358350bc235283d81;hpb=f38523dc7bcd791867b45423babb8ddb3358b31e;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 0fac0a7..da39a83 100755
--- a/tasks.py
+++ b/tasks.py
@@ -34,7 +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():
@@ -1070,6 +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,
@@ -1080,6 +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,
@@ -1179,10 +1181,14 @@ class RPL(Task):
         # --------------------------------------------------------------------
         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)
+            k = torch.arange(result.size(1), device=result.device)[None, :]
+            last_output_idx = (
+                ((result == self.t_output) * k).max(dim=1, keepdim=True).values
+            )
+            first_prog_idx = (
+                ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
+            )
+            ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 
             masked_inplace_autoregression(
@@ -1195,25 +1201,20 @@ class RPL(Task):
             )
 
             sum_nb_total, sum_nb_errors = 0, 0
-            for x, y in zip(input, result):
+            for x, y, i, j in zip(input, result, last_output_idx, first_prog_idx):
                 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
+                correct = (x - y).abs().max() == 0
+                sum_nb_errors += 0 if correct 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}]"
-                        )
+                    result_stack = [self.id2token[i.item()] for i in y[i : j + 1]]
+                    target_stack = [self.id2token[i.item()] for i in x[i : j + 1]]
+                    comment = "*" if correct else "-"
+                    result_stack = " ".join([str(x) for x in result_stack])
+                    target_stack = " ".join([str(x) for x in target_stack])
+                    logger(
+                        f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
+                    )
                     nb_to_log -= 1
 
             return sum_nb_total, sum_nb_errors
@@ -1225,7 +1226,15 @@ class RPL(Task):
         )
 
         logger(
-            f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
+            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}%"
+        )
+
+        test_nb_total, test_nb_errors = compute_nb_errors_output(
+            self.test_input[:1000].to(self.device), nb_to_log=10
+        )
+
+        logger(
+            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}%"
         )