X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=35bf02c0123815b31d479df63f935839c7523b33;hb=02b0a7bb770f07f2e91f1c77b899815516087b6a;hp=324aeba663a5b4c7453ce03ff5deb7062e1da7e3;hpb=fd2166de6350fc3f2b3fdb90849115574e3ae843;p=picoclvr.git

diff --git a/main.py b/main.py
index 324aeba..35bf02c 100755
--- a/main.py
+++ b/main.py
@@ -120,6 +120,13 @@ parser.add_argument("--stack_nb_digits", type=int, default=3)
 
 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 
+##############################
+# Expr options
+
+parser.add_argument("--expr_nb_variables", type=int, default=5)
+
+parser.add_argument("--expr_sequence_length", type=int, default=30)
+
 ######################################################################
 
 args = parser.parse_args()
@@ -163,10 +170,10 @@ default_args = {
         "nb_test_samples": 1000,
     },
     "expr": {
-        "nb_epochs": 5,
+        "nb_epochs": 50,
         "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 1000,
+        "nb_train_samples": 250000,
+        "nb_test_samples": 10000,
     },
 }
 
@@ -1012,22 +1019,38 @@ class TaskExpr(Task):
         self,
         nb_train_samples,
         nb_test_samples,
+        nb_variables,
+        sequence_length,
         batch_size,
         device=torch.device("cpu"),
     ):
         self.batch_size = batch_size
         self.device = device
 
-        train_sequences = expr.generate_sequences(nb_train_samples)
-        test_sequences = expr.generate_sequences(nb_test_samples)
+        train_sequences = expr.generate_sequences(
+            nb_train_samples,
+            nb_variables=nb_variables,
+            length=2 * sequence_length,
+            randomize_length=True,
+        )
+        test_sequences = expr.generate_sequences(
+            nb_test_samples,
+            nb_variables=nb_variables,
+            length=sequence_length,
+        )
         self.char2id = dict(
             [
                 (c, n)
-                for n, c in enumerate(set("#"+"".join(train_sequences + test_sequences)))
+                for n, c in enumerate(
+                    set("#" + "".join(train_sequences + test_sequences))
+                )
             ]
         )
         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
-        len_max = max([len(x) for x in train_sequences + test_sequences])
+
+        self.filler, self.space = self.char2id["#"], self.char2id[" "]
+
+        len_max = max([len(x) for x in train_sequences])
         self.train_input = torch.cat(
             [
                 torch.tensor(
@@ -1039,6 +1062,8 @@ class TaskExpr(Task):
             ],
             0,
         ).to(device)
+
+        len_max = max([len(x) for x in test_sequences])
         self.test_input = torch.cat(
             [
                 torch.tensor(
@@ -1050,6 +1075,7 @@ class TaskExpr(Task):
             ],
             0,
         ).to(device)
+
         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 
     def batches(self, split="train", nb_to_use=-1, desc=None):
@@ -1062,11 +1088,17 @@ class TaskExpr(Task):
         for batch in tqdm.tqdm(
             input.split(self.batch_size), dynamic_ncols=True, desc=desc
         ):
+            if split == "train":
+                last = (batch != self.filler).max(0).values.nonzero().max() + 1
+                batch = batch[:, :last]
             yield batch
 
     def vocabulary_size(self):
         return self.nb_codes
 
+    def seq2str(self, s):
+        return "".join([self.id2char[k.item()] for k in s])
+
     def produce_results(self, n_epoch, model):
         with torch.autograd.no_grad():
             t = model.training
@@ -1074,15 +1106,14 @@ class TaskExpr(Task):
 
             def compute_nb_correct(input):
                 result = input.clone()
-                space = self.char2id["#"]
-                ar_mask = (result == space).long().cumsum(dim=1).clamp(max=1)
-                result = (1 - ar_mask) * result + space * ar_mask
+                ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
+                result = (1 - ar_mask) * result + ar_mask * self.filler
                 masked_inplace_autoregression(
                     model, self.batch_size, result, ar_mask, device=self.device
                 )
 
-                nb_total = ar_mask.sum()
-                nb_correct = ((input == result).long() * ar_mask).sum()
+                nb_total = input.size(0)
+                nb_correct = (input == result).long().min(1).values.sum()
 
                 return nb_total, nb_correct
 
@@ -1096,18 +1127,18 @@ class TaskExpr(Task):
             # Log a few generated sequences
             input = self.test_input[:10]
             result = input.clone()
-            space = self.char2id["#"]
-            ar_mask = (result == space).long().cumsum(dim=1).clamp(max=1)
-            result = (1 - ar_mask) * result + space * ar_mask
+            ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
+            result = (1 - ar_mask) * result + ar_mask * self.filler
             for n in range(result.size(0)):
-                s = "".join([self.id2char[k.item()] for k in result[n]])
-                log_string(f"test_before {s}")
+                log_string(f"test_before {self.seq2str(result[n])}")
             masked_inplace_autoregression(
                 model, self.batch_size, result, ar_mask, device=self.device
             )
+            correct = (1 - ar_mask) * self.space + ar_mask * input
             for n in range(result.size(0)):
-                s = "".join([self.id2char[k.item()] for k in result[n]])
-                log_string(f"test_after  {s}")
+                comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
+                log_string(f"test_after  {self.seq2str(result[n])} {comment}")
+                log_string(f"correct     {self.seq2str(correct[n])}")
             ##############################################################
 
             model.train(t)
@@ -1193,6 +1224,8 @@ elif args.task == "expr":
     task = TaskExpr(
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
+        nb_variables=args.expr_nb_variables,
+        sequence_length=args.expr_sequence_length,
         batch_size=args.batch_size,
         device=device,
     )