X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=185df3870c1e0c8d35daf9f3e3257340d50fa1a7;hb=57332f677ef5ee535707c1b83a541aa0e79508e6;hp=0e94672777ad3df99016ffd0da855a927f8a6e6a;hpb=597f3a01de92be1a4f05114df11d9e39b3608e29;p=mygptrnn.git

diff --git a/mygpt.py b/mygpt.py
index 0e94672..185df38 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
@@ -37,7 +39,7 @@ import ffutils
 # 1 for the successive tokens.
 #
 # Modules able to process brackets may implement a cache that is
-# resetted when the input bracket starts at t=0
+# resetted when init_cache is True
 
 
 class BracketedSequence:
@@ -474,13 +476,17 @@ class Caterpillar(nn.Module):
 
         warnings.warn("Caterpillar", RuntimeWarning)
 
-        def randw(*d):
-            return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
+        def randw(*d, amplitude=None):
+            if amplitude is None:
+                amplitude = 1 / math.sqrt(d[-1])
+            return nn.Parameter(amplitude * torch.randn(*d))
 
         self.caterpillar_length = caterpillar_length
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
+        self.proba_gate_dropout = 0.0
+
         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
         self.b_G = nn.Parameter(
             torch.full(
@@ -493,8 +499,16 @@ class Caterpillar(nn.Module):
         self.w_Q = randw(nb_heads, dim_qk, dim_model)
         self.w_O = randw(dim_v * nb_heads, dim_model)
 
-        self.init_K_rec = randw(caterpillar_height, caterpillar_length, dim_qk)
-        self.init_V_rec = randw(caterpillar_height, caterpillar_length, dim_v)
+        self.init_K_rec = randw(
+            caterpillar_height,
+            caterpillar_length,
+            dim_qk,
+        )
+        self.init_V_rec = randw(
+            caterpillar_height,
+            caterpillar_length,
+            dim_v,
+        )
 
     def reset_inner_loss(self):
         self.acc_attention = 0
@@ -512,6 +526,7 @@ 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)
@@ -534,38 +549,80 @@ class Caterpillar(nn.Module):
 
             self.cache_Y = X.new_zeros(N, T, DM)
 
+        V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
+        K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
+
         ######################################################################
         # Compute the recurrent state
 
         # 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 all the 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]
+            torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
         ).sigmoid()
 
-        G = F.dropout(G, self.attention_dropout, self.training)
+        ######################################################################
+        # The "flashbacks"
 
-        V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
-        K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
+        if self.training and self.proba_gate_dropout > 0.0:
+            # This is a better implementation of "flashbacks".
+
+            # G is NxHxExT where e is the caterpillar's row.
+
+            warnings.warn("gate dropout", RuntimeWarning)
+            epsilon = 0.5
+
+            dropout_head = (
+                (torch.rand(N, H, 1, t1 - t0, device=G.device).sort(dim=3).indices == 0)
+                .expand_as(G)
+                .float()
+            )
+
+            dropout_tail = dropout_head.cumsum(dim=3) - dropout_head
+
+            dropout_active = (
+                torch.rand(N, 1, 1, 1, device=G.device) < self.proba_gate_dropout
+            ).long()
+
+            dropout_head *= dropout_active
+            dropout_tail *= dropout_active
+
+            G = (
+                G
+                # + dropout_head * (1 - epsilon - G.detach())
+                - dropout_tail * G.detach()
+            )
+
+        ######################################################################
 
         # We prepare the arguments for the parallel scan
 
+        # 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)
+
         A = 1 - G.sum(1)
-        gated_V = torch.einsum("nhet,nhtd->netd", G, V)
-        gated_K = torch.einsum("nhet,nhtd->netd", G, K)
+        gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V)
+        gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K)
+
+        # We start from cached values, which matters in inference
 
         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))
@@ -574,8 +631,6 @@ 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)
 
@@ -725,7 +780,6 @@ class MyGPT(nn.Module):
         nb_blocks,
         nb_lines=None,
         caterpillar_height=None,
-        dim_rec_v=-1,
         causal=False,
         dropout=0.0,
         len_max=1e5,
@@ -733,7 +787,12 @@ class MyGPT(nn.Module):
     ):
         super().__init__()
 
-        assert attention_layer in {"mha", "dumbrec", "kvrec", "caterpillar"}
+        assert attention_layer in {
+            "mha",
+            "dumbrec",
+            "kvrec",
+            "caterpillar",
+        }, f"Unknown attention operator {attention_layer}."
 
         if attention_layer == "caterpillar":
             assert nb_lines % caterpillar_height == 0
@@ -766,7 +825,7 @@ 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,
@@ -775,7 +834,7 @@ 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,
@@ -784,7 +843,7 @@ 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,