X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=d6845e8fb3a0f4c123181ea57047ffe10260acfa;hb=a1ae050705970007f965d2586c53e9bd262e46aa;hp=4d5077ae1b46b7398aed253c14129d5f6b451879;hpb=a3c32b845b6903fd290f2b09d5c53203ff112b79;p=mygptrnn.git

diff --git a/main.py b/main.py
index 4d5077a..d6845e8 100755
--- a/main.py
+++ b/main.py
@@ -87,7 +87,9 @@ parser.add_argument("--model", type=str, default=None)
 
 parser.add_argument("--attention", type=str, default=None)
 
-parser.add_argument("--proportion_memex", type=float, default=0)
+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)
 
@@ -738,7 +740,7 @@ log_string(f"device {device}")
 
 vocabulary_size = task.vocabulary_size()
 
-if args.proportion_memex > 0:
+if args.memex_proba > 0:
     vocabulary_size += 1
 
 log_string(f"vocabulary_size {vocabulary_size}")
@@ -902,22 +904,27 @@ 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
 
-    def add_memex(batches, proportion_memex):
+    def add_memex(batches, memex_proba):
         for input in batches:
-            if torch.rand(1).item() < proportion_memex:
+            if torch.rand(1).item() < memex_proba:
+                sep = torch.full(
+                    (input.size(0), 1), vocabulary_size - 1, device=input.device
+                )
+
                 yield torch.cat(
                     [
                         input,
-                        torch.full(
-                            (input.size(0), 1), vocabulary_size - 1, device=input.device
-                        ),
+                        sep,
                         input,
                     ],
                     dim=1,
                 )
             yield input
 
-    train_batches = add_memex(task.batches(split="train"), args.proportion_memex)
+    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()