X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=mygpt.py;h=676b92181bd4bcc8eb5e2d22e58cdfa672e8e947;hb=ffe183868ac8563fd82fc8312fda90f6f8a95833;hp=17f2f6d721a6c3848df0e65908589947eb6f3fd9;hpb=de0831357e74b1d1a61b1a41890e20ed1a2c9b96;p=mygptrnn.git

diff --git a/mygpt.py b/mygpt.py
index 17f2f6d..676b921 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -483,7 +483,6 @@ class Caterpillar(nn.Module):
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
-        self.proba_flashback = 0.0
         self.proba_gate_dropout = 0.0
 
         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
@@ -540,6 +539,9 @@ class Caterpillar(nn.Module):
 
             self.cache_Y = X.new_zeros(N, T, DM)
 
+        V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
+        K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
+
         ######################################################################
         # Compute the recurrent state
 
@@ -558,24 +560,26 @@ class Caterpillar(nn.Module):
 
         G = G / G.sum(1, keepdim=True).clamp(min=1)
 
-        if self.training and self.proba_gate_dropout > 0.0:
-            warnings.warn("gate dropout", RuntimeWarning)
-            epsilon = 0.5
-
-        V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
-        K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
-
         # We prepare the arguments for the parallel scan
 
         A = 1 - G.sum(1)
         gated_V = torch.einsum("nhet,nhtd->netd", G, V)
         gated_K = torch.einsum("nhet,nhtd->netd", G, K)
 
-        # Initial recurrent state
+        # We start from cached values, which matters in inference
 
         init_rec_V = self.rec_V[:, :, t0 - CL : t0]
         init_rec_K = self.rec_K[:, :, t0 - CL : t0]
 
+        ######################################################################
+
+        if self.training and self.proba_gate_dropout > 0.0:
+            # This is a better implementation of "flashbacks".  A is
+            # NxExT where e is the caterpillar's row.
+
+            warnings.warn("gate dropout", RuntimeWarning)
+            epsilon = 0.5
+
         #################################################################
         # Associative scan
 
@@ -595,42 +599,6 @@ class Caterpillar(nn.Module):
         self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 
-        #################################################################
-
-        if self.training and self.proba_flashback > 0.0:
-            warnings.warn("flash back", RuntimeWarning)
-            # This piece of code makes the assumption that there is
-            # nothing informative before t0, otherwise we'd have to
-            # implement a cache for V and K too. This should not be
-            # too much of a problem since this is used only during
-            # train, where full sequence are available
-
-            n = torch.arange(N, device=X.device)[:, None, None, None]
-            t = torch.arange(t0, t1, device=X.device)[None, None, :, None]
-            dv = torch.arange(DV, device=X.device)[None, None, None, :]
-            dk = torch.arange(DK, device=X.device)[None, None, None, :]
-
-            u = (
-                torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
-            ) * CL
-
-            src_time = t - u - t0
-            src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
-
-            mask = (
-                torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
-            ).long()
-
-            self.rec_V[:, :, t0:t1] = (
-                mask * V[n, src_head, src_time, dv]
-                + (1 - mask) * self.rec_V[:, :, t0:t1]
-            )
-
-            self.rec_K[:, :, t0:t1] = (
-                mask * K[n, src_head, src_time, dk]
-                + (1 - mask) * self.rec_K[:, :, t0:t1]
-            )
-
         ######################################################################
         # compute the readout