X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=d6845e8fb3a0f4c123181ea57047ffe10260acfa;hb=a1ae050705970007f965d2586c53e9bd262e46aa;hp=3aa696b38784c8270c5a840c3e4d5be61dacad2f;hpb=c45d89eb5383eedf60466678eae623582bd5781c;p=mygptrnn.git

diff --git a/main.py b/main.py
index 3aa696b..d6845e8 100755
--- a/main.py
+++ b/main.py
@@ -87,6 +87,10 @@ parser.add_argument("--model", type=str, default=None)
 
 parser.add_argument("--attention", type=str, default=None)
 
+parser.add_argument("--memex_proba", type=float, default=0)
+
+parser.add_argument("--memex_nb_epochs", type=float, default=1)
+
 parser.add_argument("--dim_model", type=int, default=None)
 
 parser.add_argument("--dim_keys", type=int, default=None)
@@ -101,7 +105,9 @@ parser.add_argument("--caterpillar_height", type=int, default=None)
 
 parser.add_argument("--gate_dropout_proba", type=float, default=0.0)
 
-parser.add_argument("--gate_dropout_sync", type=bool, default=False)
+parser.add_argument("--gate_dropout_sync", type=str2bool, default=False)
+
+parser.add_argument("--gate_dropout_replace", type=str2bool, default=False)
 
 parser.add_argument("--rho_inner_loss", type=float, default=0.0)
 
@@ -734,6 +740,9 @@ log_string(f"device {device}")
 
 vocabulary_size = task.vocabulary_size()
 
+if args.memex_proba > 0:
+    vocabulary_size += 1
+
 log_string(f"vocabulary_size {vocabulary_size}")
 
 ##############################
@@ -895,7 +904,29 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
 
     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 
-    for input in task.batches(split="train"):
+    def add_memex(batches, memex_proba):
+        for input in batches:
+            if torch.rand(1).item() < memex_proba:
+                sep = torch.full(
+                    (input.size(0), 1), vocabulary_size - 1, device=input.device
+                )
+
+                yield torch.cat(
+                    [
+                        input,
+                        sep,
+                        input,
+                    ],
+                    dim=1,
+                )
+            yield input
+
+    train_batches = add_memex(
+        task.batches(split="train"),
+        args.memex_proba if n_epoch < args.memex_nb_epochs else 0.0,
+    )
+
+    for input in train_batches:
         model.reset_inner_loss()
         input = input.to(device)