Update.

[mygptrnn.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index 492a9bb..0414bb6 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -202,7 +202,7 @@ class DumbRec(nn.Module):
         attention_dropout=0.0,
         len_max=1e5,
         logger=print,
-        **kwargs,
+        args=None,
     ):
         super().__init__()
 
@@ -333,7 +333,7 @@ class KVRec(nn.Module):
         attention_dropout=0.0,
         len_max=1e5,
         logger=print,
-        **kwargs,
+        args=None,
     ):
         super().__init__()
 
@@ -487,7 +487,7 @@ class Caterpillar(nn.Module):
         attention_dropout=0.0,
         len_max=1e5,
         logger=print,
-        **kwargs,
+        args=None,
     ):
         super().__init__()
 
@@ -502,27 +502,12 @@ class Caterpillar(nn.Module):
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
-        ######################################################################
-        # sup_args
-
-        x = kwargs.get("gate_dropout")
-        if x is None:
-            self.proba_gate_dropout = 0.0
-        else:
-            self.proba_gate_dropout = float(x)
-
-        logger(f"self.proba_gate_dropout {self.proba_gate_dropout}")
-
-        x = kwargs.get("default_bg")
-        if x is None:
-            default_bg = -math.log(caterpillar_height - 1)
-        else:
-            default_bg = float(x)
-
-        logger(f"default_bg {default_bg}")
+        self.gate_dropout_proba = args.gate_dropout_proba
+        self.gate_dropout_sync = args.gate_dropout_sync
 
         ######################################################################
 
+        default_bg = -math.log(caterpillar_height - 1)
         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
         self.b_G = nn.Parameter(torch.full((nb_heads, caterpillar_height), default_bg))
 
@@ -617,8 +602,6 @@ class Caterpillar(nn.Module):
             init_rec_V = self.rec_V[:, :, t0 - L : t0]
             init_rec_K = self.rec_K[:, :, t0 - L : t0]
 
-            # Associative scan
-
             # Here there is a trick: Since the stack at position t is
             # computed by updating that at position t-L, the parallel
             # scan operates with a period of L. To do so we split the
@@ -641,24 +624,32 @@ class Caterpillar(nn.Module):
 
         next_V, next_K = recurrence(G, V, K)
 
-        if self.training and self.proba_gate_dropout > 0.0:
+        if self.training and self.gate_dropout_proba > 0.0:
             # G is NxHxRxT where r is the caterpillar's row.
 
             warnings.warn("gate dropout", RuntimeWarning)
 
+            # Pick a point in each of the NxHxR timeline and set this
+            # entry and the following to 1
             kill = (
-                torch.rand(G.size(), device=G.device) <= self.proba_gate_dropout
-            ).float()
+                torch.rand(N, H, R, t1 - t0, device=G.device).sort(dim=3).indices == 0
+            ).cumsum(dim=3)
+
+            # Keep these mask for only some of the NxHxR
+            kill = kill * (
+                torch.rand(N, H, R, 1, device=G.device) <= self.gate_dropout_proba
+            )
 
+            # The coefficient to keep are the complementary
             mask = 1 - kill
 
             masked_next_V, masked_next_K = recurrence(G * mask, V, K)
 
             next_V = next_V.detach() + (masked_next_V - masked_next_V.detach()) / (
-                1 - self.proba_gate_dropout
+                1 - self.gate_dropout_proba
             )
             next_K = next_K.detach() + (masked_next_K - masked_next_K.detach()) / (
-                1 - self.proba_gate_dropout
+                1 - self.gate_dropout_proba
             )
 
         self.rec_V[:, :, t0:t1] = next_V
@@ -669,8 +660,8 @@ class Caterpillar(nn.Module):
 
         Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
 
-        # We build tensors NxHxTxFxL where N is the sample index, H
-        # the head, T the time, F the row in the caterpillar, and L
+        # We build tensors NxHxTxRxL where N is the sample index, H
+        # the head, T the time, R the row in the caterpillar, and L
         # the column in the caterpillar
 
         windowed_V = moving_window(
@@ -684,7 +675,7 @@ class Caterpillar(nn.Module):
         # We have an attention score for each of the RxL values
 
         ar = torch.einsum(
-            "nhtd,nftld->nhtfl",
+            "nhtd,nrtld->nhtrl",
             Q,
             windowed_K,
         ) / math.sqrt(DK)
@@ -724,7 +715,7 @@ class QKVAttention(nn.Module):
         causal=False,
         attention_dropout=0.0,
         logger=print,
-        **kwargs,
+        args=None,
     ):
         super().__init__()
 
@@ -817,7 +808,7 @@ class MyGPT(nn.Module):
         len_max=1e5,
         attention_layer="kvrec",
         logger=print,
-        **kwargs,
+        args=None,
     ):
         super().__init__()
 
@@ -855,7 +846,7 @@ class MyGPT(nn.Module):
                     causal=causal,
                     attention_dropout=dropout,
                     logger=logger,
-                    **kwargs,
+                    args=args,
                 )
             elif attention_layer == "dumbrec":
                 return DumbRec(
@@ -866,7 +857,7 @@ class MyGPT(nn.Module):
                     nb_lines=nb_lines,
                     attention_dropout=dropout,
                     logger=logger,
-                    **kwargs,
+                    args=args,
                 )
             elif attention_layer == "kvrec":
                 return KVRec(
@@ -877,7 +868,7 @@ class MyGPT(nn.Module):
                     nb_lines=nb_lines,
                     attention_dropout=dropout,
                     logger=logger,
-                    **kwargs,
+                    args=args,
                 )
             elif attention_layer == "caterpillar":
                 return Caterpillar(
@@ -889,7 +880,7 @@ class MyGPT(nn.Module):
                     caterpillar_height=self.caterpillar_height,
                     attention_dropout=dropout,
                     logger=logger,
-                    **kwargs,
+                    args=args,
                 )
             else:
                 raise ValueError(f"Unknown attention type {attention_layer}.")