Update.

[mygptrnn.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index 4d48247..de69a75 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -481,6 +481,9 @@ class Caterpillar(nn.Module):
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
+        warnings.warn("flash back", RuntimeWarning)
+        self.proba_flashback = 0.1
+
         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
         self.b_G = nn.Parameter(
             torch.full(
@@ -512,9 +515,10 @@ class Caterpillar(nn.Module):
 
         N = bs.x.size(0)
         T = bs.x.size(1)
+        H = self.w_V.size(0)
         DV = self.w_V.size(1)
         DK = self.w_K.size(1)
-        Dout = self.w_O.size(1)
+        DM = self.w_O.size(1)
         CH = self.caterpillar_height
         CL = self.caterpillar_length
 
@@ -522,6 +526,8 @@ class Caterpillar(nn.Module):
             t0 >= CL and (t1 - t0) % CL == 0
         ), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
 
+        # We cache values to deal efficiently with auto-regression
+
         if bs.init_cache:
             self.rec_V = X.new_zeros(N, CH, T, DV)
             self.rec_K = X.new_zeros(N, CH, T, DK)
@@ -530,7 +536,7 @@ class Caterpillar(nn.Module):
             self.rec_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :]
             self.rec_K[:, :, t0 - CL : t0] = self.init_K_rec[None, :, :, :]
 
-            self.cache_Y = X.new_zeros(N, T, Dout)
+            self.cache_Y = X.new_zeros(N, T, DM)
 
         ######################################################################
         # Compute the recurrent state
@@ -538,14 +544,15 @@ class Caterpillar(nn.Module):
         # This is the Gating sequence that modulates the storing of
         # the new key and value in the CH pairs of the current
         # stack. The CH gating values are independent, which means
-        # that the current K/V could be stored in all the pairs of the
+        # that the current K/V could be stored in multiple pairs of the
         # recurrent state, or not at all.
 
         G = (
             torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None]
         ).sigmoid()
 
-        G = F.dropout(G, self.attention_dropout, self.training)
+        # That bas a bad idea
+        # G = F.dropout(G, self.attention_dropout, self.training)
 
         V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
         K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
@@ -577,6 +584,41 @@ 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:
+            # 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_V = (
+                torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
+            ).long()
+            self.rec_V[:, :, t0:t1] = (
+                mask_V * V[n, src_head, src_time, dv]
+                + (1 - mask_V) * self.rec_V[:, :, t0:t1]
+            )
+
+            mask_K = (
+                torch.rand(N, CH, t1 - t0, DK, device=X.device) <= self.proba_flashback
+            ).long()
+            self.rec_K[:, :, t0:t1] = (
+                mask_K * K[n, src_head, src_time, dk]
+                + (1 - mask_K) * self.rec_K[:, :, t0:t1]
+            )
+
         ######################################################################
         # compute the readout