Update.

diff --git a/fridge b/fridge
index ac7f86c..d28cc89 100644 (file)
--- a/fridge
+++ b/fridge
@@ -81,3 +81,39 @@ 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]
+            # )

diff --git a/mygpt.py b/mygpt.py
index e7362b7..b885e21 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -481,8 +481,8 @@ class Caterpillar(nn.Module):
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
-        warnings.warn("flash back", RuntimeWarning)
-        self.proba_flashback = 1e-2
+        self.proba_flashback = 0.0
+        self.proba_gate_dropout = 0.0
 
         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
         self.b_G = nn.Parameter(
@@ -551,7 +551,11 @@ class Caterpillar(nn.Module):
             torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None]
         ).sigmoid()
 
-        # That bas a bad idea
+        if self.training and self.proba_gate_dropout > 0.0:
+            warnings.warn("gate droupout", RuntimeWarning)
+            epsilon = 0.5
+
+        # That was a bad idea
         # G = F.dropout(G, self.attention_dropout, self.training)
 
         V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
@@ -559,6 +563,10 @@ class Caterpillar(nn.Module):
 
         # We prepare the arguments for the parallel scan
 
+        # Clip the gating
+        warnings.warn("gating clipping", RuntimeWarning)
+        G = G / G.sum(1, keepdim=True).clamp(min=1)
+
         A = 1 - G.sum(1)
         gated_V = torch.einsum("nhet,nhtd->netd", G, V)
         gated_K = torch.einsum("nhet,nhtd->netd", G, K)
@@ -585,6 +593,7 @@ class Caterpillar(nn.Module):
         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