X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=fridge;h=194c4e6f672f94b64a176bd5c216f9c6c3cd4340;hb=2434c00a82ebb0b23f45d891cc9f80324e3200bd;hp=ac7f86c5bc9d5159aca0da267eb8dbcb6673eb94;hpb=f3f490def0be8a3ea2b9a0ac60f5bb33c5c45fb5;p=mygptrnn.git

diff --git a/fridge b/fridge
index ac7f86c..194c4e6 100644
--- a/fridge
+++ b/fridge
@@ -81,3 +81,99 @@ def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
 
             # insert_flash_back(self.rec_V,V,self.rec_K,K,t0,t1,CL,proba=self.proba_flashback / CL,)
 
+
+######################################################################
+
+2024 Jan 09 14:24:42 (from mygpt.py)
+
+            # 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]
+            # )
+
+######################################################################
+
+2024 Jan 10 08:10:39 (from mygpt.py)
+
+        # That was a bad idea
+        # G = F.dropout(G, self.attention_dropout, self.training)
+
+
+######################################################################
+
+2024 Jan 10 08:46:13 (from mygpt.py)
+
+        #################################################################
+        # Flashbacks. This version sucks, about to replace it 
+        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]
+            )
+
+
+######################################################################
+
+2024 Jan 13 13:38:31 (from mygpt.py)
+
+        g= F.sigmoid(self.b_G)
+        a=1-g
+
+        print(f"\n\nSANITY {a**T}\n")
+        exit(0)
+