X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=b137cdbca1cd085ee8dec0185c514118e928b5d3;hb=9112db2ed7d8c262c4ef8298cf6637515675f967;hp=0414bb615435d610d2d7c1c0123763081462f330;hpb=ec5bad2e7911bdf9b7851342a3ee007d41b80963;p=mygptrnn.git

diff --git a/mygpt.py b/mygpt.py
index 0414bb6..b137cdb 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -504,6 +504,7 @@ class Caterpillar(nn.Module):
 
         self.gate_dropout_proba = args.gate_dropout_proba
         self.gate_dropout_sync = args.gate_dropout_sync
+        self.gate_dropout_replace = args.gate_dropout_replace
 
         ######################################################################
 
@@ -527,14 +528,14 @@ class Caterpillar(nn.Module):
             dim_v,
         )
 
-    def reset_inner_loss(self):
-        self.acc_attention = 0
-        self.acc_nb = 0
+    # def reset_inner_loss(self):
+    # self.acc_attention = 0
+    # self.acc_nb = 0
 
-    def get_inner_loss(self):
-        # warnings.warn("l2 regularization", RuntimeWarning)
-        # return (self.acc_attention / self.acc_nb).pow(2).sum()
-        return torch.tensor([0], device=self.w_Q.device)
+    # def get_inner_loss(self):
+    # warnings.warn("l2 regularization", RuntimeWarning)
+    # return (self.acc_attention / self.acc_nb).pow(2).sum()
+    # return torch.tensor([0], device=self.w_Q.device)
 
     def forward(self, bs):
         # Dimensions to make the source a bit clearer, that's needed
@@ -612,11 +613,8 @@ class Caterpillar(nn.Module):
             gated_V = gated_V.unflatten(2, (-1, L))
             gated_K = gated_K.unflatten(2, (-1, L))
 
-            next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
-            next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
-
-            next_V = next_V.flatten(2, 3)
-            next_K = next_K.flatten(2, 3)
+            next_V = pscan_dim(A, gated_V, init_rec_V, dim=2).flatten(2, 3)
+            next_K = pscan_dim(A, gated_K, init_rec_K, dim=2).flatten(2, 3)
 
             return next_V, next_K
 
@@ -629,15 +627,21 @@ class Caterpillar(nn.Module):
 
             warnings.warn("gate dropout", RuntimeWarning)
 
+            if self.gate_dropout_sync:
+                shape_kill = (N, 1, 1)
+            else:
+                shape_kill = (N, H, R)
+
             # Pick a point in each of the NxHxR timeline and set this
             # entry and the following to 1
             kill = (
-                torch.rand(N, H, R, t1 - t0, device=G.device).sort(dim=3).indices == 0
+                torch.rand(*shape_kill, t1 - t0, device=G.device).sort(dim=3).indices
+                == 0
             ).cumsum(dim=3)
 
             # Keep these mask for only some of the NxHxR
             kill = kill * (
-                torch.rand(N, H, R, 1, device=G.device) <= self.gate_dropout_proba
+                torch.rand(*shape_kill, 1, device=G.device) <= self.gate_dropout_proba
             )
 
             # The coefficient to keep are the complementary
@@ -645,10 +649,14 @@ class Caterpillar(nn.Module):
 
             masked_next_V, masked_next_K = recurrence(G * mask, V, K)
 
-            next_V = next_V.detach() + (masked_next_V - masked_next_V.detach()) / (
+            if self.gate_dropout_replace:
+                next_V = next_V.detach()
+                next_K = next_K.detach()
+
+            next_V = next_V + (masked_next_V - masked_next_V.detach()) / (
                 1 - self.gate_dropout_proba
             )
-            next_K = next_K.detach() + (masked_next_K - masked_next_K.detach()) / (
+            next_K = next_K + (masked_next_K - masked_next_K.detach()) / (
                 1 - self.gate_dropout_proba
             )