Update.

[mygptrnn.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index 5754d55..5ea927e 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -509,32 +509,22 @@ def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
 
     fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device)
 
-    fbt_V = fbt[:, :, :, None].expand_as(rec_V[:, :, t0:t1])
-    fbh_V = fbh[:, :, :, None].expand_as(rec_V[:, :, t0:t1])
+    fbt_V = fbt[:, :, :, None]
+    fbh_V = fbh[:, :, :, None]
     t = fbt_V.clamp(min=0)
-    n = torch.arange(V.size(0), device=V.device)[:, None, None, None].expand_as(
-        rec_V[:, :, t0:t1]
-    )
-    d = torch.arange(V.size(3), device=V.device)[None, None, None, :].expand_as(
-        rec_V[:, :, t0:t1]
-    )
+    n = torch.arange(V.size(0), device=V.device)[:, None, None, None]
+    d = torch.arange(V.size(3), device=V.device)[None, None, None, :]
     q = V[:, :, t0:t1][n, fbh_V, t, d]
     rec_V[:, :, t0:t1] = q * (fbt_V >= 0) + rec_V[:, :, t0:t1] * (fbt_V < 0)
 
-    fbt_K = fbt[:, :, :, None].expand_as(rec_K[:, :, t0:t1])
-    fbh_K = fbh[:, :, :, None].expand_as(rec_K[:, :, t0:t1])
+    fbt_K = fbt[:, :, :, None]
+    fbh_K = fbh[:, :, :, None]
     t = fbt_K.clamp(min=0)
-    n = torch.arange(K.size(0), device=K.device)[:, None, None, None].expand_as(
-        rec_K[:, :, t0:t1]
-    )
-    d = torch.arange(K.size(3), device=K.device)[None, None, None, :].expand_as(
-        rec_K[:, :, t0:t1]
-    )
+    n = torch.arange(K.size(0), device=K.device)[:, None, None, None]
+    d = torch.arange(K.size(3), device=K.device)[None, None, None, :]
     q = K[:, :, t0:t1][n, fbh_K, t, d]
     rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0)
 
-    # print("SANITY", (fbt_K >=0).float().sum()/fbt_K.numel())
-
 
 ######################################################################
 
@@ -666,36 +656,41 @@ class Caterpillar(nn.Module):
         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 
         if self.training and self.proba_flashback:
-            insert_flash_back(
-                self.rec_V,
-                V,
-                self.rec_K,
-                K,
-                t0,
-                t1,
-                CL,
-                proba=self.proba_flashback / CL,
-            )
+            # insert_flash_back(self.rec_V,V,self.rec_K,K,t0,t1,CL,proba=self.proba_flashback / CL,)
 
-            # 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)[None, None, None, :]
-            # dk = torch.arange(DK)[None, None, None, :]
+            # 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
 
-            # u = (
-            # torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
-            # ) * CL
+            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, :]
 
-            # src_time = t - u - t0
-            # src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
+            u = (
+                torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
+            ) * CL
 
-            # mk = (
-            # torch.rand(self.rec_V[:, :, t0:t1].size()) <= self.proba_flashback
-            # ).long()
-            # self.rec_V[:, :, t0:t1] = V[n, src_head, src_time, dv]
-            # self.rec_K[:, :, t0:t1] = K[n, src_head, src_time, dk]
+            src_time = t - u - t0
+            src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
 
-        exit(0)
+            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