X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=17f2f6d721a6c3848df0e65908589947eb6f3fd9;hb=de0831357e74b1d1a61b1a41890e20ed1a2c9b96;hp=e7362b749210dcce6c6b92934bd34744d95b770d;hpb=6833683bd343fd687d093d6c47cca8f1909e8b03;p=mygptrnn.git

diff --git a/mygpt.py b/mygpt.py
index e7362b7..17f2f6d 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -10,6 +10,8 @@
 # with a caching mechanism for keys and values to avoid a O(N^3) cost
 # for auto-regression.
 
+# This implementation is equipped with RNN layers to replace the MHA
+
 import math, warnings
 
 import torch, einops
@@ -481,8 +483,8 @@ 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(
@@ -543,16 +545,22 @@ 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 multiple pairs of the
+        # stack. There are CH independent gating values, which means
+        # that the current K/V may 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()
 
-        # That bas a bad idea
-        # G = F.dropout(G, self.attention_dropout, self.training)
+        # Clip the gating to avoid values greater than 1 when several
+        # heads hit the same row
+
+        G = G / G.sum(1, keepdim=True).clamp(min=1)
+
+        if self.training and self.proba_gate_dropout > 0.0:
+            warnings.warn("gate dropout", RuntimeWarning)
+            epsilon = 0.5
 
         V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
         K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
@@ -563,14 +571,19 @@ class Caterpillar(nn.Module):
         gated_V = torch.einsum("nhet,nhtd->netd", G, V)
         gated_K = torch.einsum("nhet,nhtd->netd", G, K)
 
+        # Initial recurrent state
+
         init_rec_V = self.rec_V[:, :, t0 - CL : t0]
         init_rec_K = self.rec_K[:, :, t0 - CL : t0]
 
-        # Here there is a trick: Since the stack at time t is computed
-        # by updating that at time t-L, the parallel scan operates
-        # with a period of L. To do so we split the time indexing in
-        # two axes, the second of size CL, and run the parallel scan
-        # using the other as the sequence index.
+        #################################################################
+        # Associative scan
+
+        # Here there is a trick: Since the stack at position t is
+        # computed by updating that at position t-CL, the parallel
+        # scan operates with a period of CL. To do so we split the
+        # sequence indexing in two axes, the second of size CL, and
+        # run the parallel scan using the first as the sequence index.
 
         A = A.unflatten(2, (-1, CL))
         gated_V = gated_V.unflatten(2, (-1, CL))
@@ -579,12 +592,13 @@ class Caterpillar(nn.Module):
         next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
         next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
 
-        # Put back the sequence index
-
         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:
+            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
@@ -763,7 +777,6 @@ class MyGPT(nn.Module):
         nb_blocks,
         nb_lines=None,
         caterpillar_height=None,
-        dim_rec_v=-1,
         causal=False,
         dropout=0.0,
         len_max=1e5,
@@ -809,7 +822,7 @@ class MyGPT(nn.Module):
                 return DumbRec(
                     dim_model=dim_model,
                     dim_qk=dim_keys,
-                    dim_v=dim_rec_v,
+                    dim_v=dim_model // nb_heads,
                     nb_heads=nb_heads,
                     nb_lines=nb_lines,
                     attention_dropout=dropout,
@@ -818,7 +831,7 @@ class MyGPT(nn.Module):
                 return KVRec(
                     dim_model=dim_model,
                     dim_qk=dim_keys,
-                    dim_v=dim_rec_v,
+                    dim_v=dim_model // nb_heads,
                     nb_heads=nb_heads,
                     nb_lines=nb_lines,
                     attention_dropout=dropout,
@@ -827,7 +840,7 @@ class MyGPT(nn.Module):
                 return Caterpillar(
                     dim_model=dim_model,
                     dim_qk=dim_keys,
-                    dim_v=dim_rec_v,
+                    dim_v=dim_model // nb_heads,
                     nb_heads=nb_heads,
                     caterpillar_length=self.caterpillar_length,
                     caterpillar_height=self.caterpillar_height,