Update.

[mygptrnn.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index 0e94672..b885e21 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -37,7 +37,7 @@ 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:
@@ -481,6 +481,9 @@ class Caterpillar(nn.Module):
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
+        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(
             torch.full(
@@ -512,6 +515,7 @@ 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)
@@ -540,20 +544,29 @@ 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 all the pairs of the
+        # that the current K/V could 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()
 
-        G = F.dropout(G, self.attention_dropout, self.training)
+        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)
         K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
 
         # 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)
@@ -579,6 +592,40 @@ class Caterpillar(nn.Module):
         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
+            # 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]
+            )
+
         ######################################################################
         # compute the readout
 
@@ -733,7 +780,12 @@ 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