X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=21ae73901d2525f9865ed48694f6a48dccb461a2;hb=aad820c7d81e962b5f6459093fe558126198f1ed;hp=7f0fb9b6fa506a5136ff4e98c8b9f5a4087420ee;hpb=8a32cb4548bb48ef68adb4df9372fe5f7a80b67c;p=mygptrnn.git

diff --git a/mygpt.py b/mygpt.py
index 7f0fb9b..21ae739 100755
--- 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())
-
 
 ######################################################################
 
@@ -562,6 +552,9 @@ 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(
@@ -593,6 +586,7 @@ 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)
         DM = self.w_O.size(1)
@@ -661,9 +655,37 @@ class Caterpillar(nn.Module):
         self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 
-        warnings.warn("flash back", RuntimeWarning)
-        if self.training:
-            insert_flash_back(self.rec_V, V, self.rec_K, K, t0, t1, CL, proba=1e-2 / CL)
+        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,
+            )
+
+            # 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, :]
+
+            # 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)
+
+            # 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]
+
+        exit(0)
 
         ######################################################################
         # compute the readout