Update.

[mygptrnn.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index eda8685..bd870bc 100755 (executable)
--- 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:
@@ -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 = 0.1
+        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 +553,11 @@ 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 +565,10 @@ 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)
@@ -585,8 +595,7 @@ class Caterpillar(nn.Module):
         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 
         if self.training and self.proba_flashback > 0.0:
-            # insert_flash_back(self.rec_V,V,self.rec_K,K,t0,t1,CL,proba=self.proba_flashback / CL,)
-
+            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
@@ -605,20 +614,18 @@ class Caterpillar(nn.Module):
             src_time = t - u - t0
             src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
 
-            mask_V = (
+            mask = (
                 torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
             ).long()
+
             self.rec_V[:, :, t0:t1] = (
-                mask_V * V[n, src_head, src_time, dv]
-                + (1 - mask_V) * self.rec_V[:, :, t0:t1]
+                mask * V[n, src_head, src_time, dv]
+                + (1 - mask) * self.rec_V[:, :, t0:t1]
             )
 
-            mask_K = (
-                torch.rand(N, CH, t1 - t0, DK, device=X.device) <= self.proba_flashback
-            ).long()
             self.rec_K[:, :, t0:t1] = (
-                mask_K * K[n, src_head, src_time, dk]
-                + (1 - mask_K) * self.rec_K[:, :, t0:t1]
+                mask * K[n, src_head, src_time, dk]
+                + (1 - mask) * self.rec_K[:, :, t0:t1]
             )
 
         ######################################################################
@@ -775,7 +782,12 @@ 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