X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=7c8e9f4c894ad332e808d07f008ac4c569046bd1;hb=f0ea1f2375fa3a0be38970a58185cddee97dccef;hp=17f2f6d721a6c3848df0e65908589947eb6f3fd9;hpb=de0831357e74b1d1a61b1a41890e20ed1a2c9b96;p=mygptrnn.git

diff --git a/mygpt.py b/mygpt.py
index 17f2f6d..7c8e9f4 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -476,17 +476,18 @@ 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_flashback = 0.0
         self.proba_gate_dropout = 0.0
 
-        self.w_G = randw(nb_heads, caterpillar_height, dim_model)
+        self.w_G = randw(nb_heads, caterpillar_height, dim_model, amplitude=1e-5)
         self.b_G = nn.Parameter(
             torch.full(
                 (nb_heads, caterpillar_height), -math.log(caterpillar_height - 1)
@@ -498,8 +499,12 @@ 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, amplitude=1e-5
+        )
+        self.init_V_rec = randw(
+            caterpillar_height, caterpillar_length, dim_v, amplitude=1e-5
+        )
 
     def reset_inner_loss(self):
         self.acc_attention = 0
@@ -540,6 +545,9 @@ 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
 
@@ -550,28 +558,61 @@ class Caterpillar(nn.Module):
         # 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()
 
-        # 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)
+        ######################################################################
+        # The "flashbacks"
 
         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
 
-        V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
-        K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
+            dropout_start = (
+                (
+                    torch.rand(G.size(), device=G.device)
+                    .flatten(2, 3)
+                    .sort(dim=2)
+                    .indices
+                    == 0
+                )
+                .unflatten(2, (CH, t1 - t0))
+                .float()
+            )
+
+            dropout_tail = dropout_start.cumsum(dim=3) - dropout_start
+
+            dropout_active = (
+                torch.rand(N, 1, 1, 1, device=G.device) < self.proba_gate_dropout
+            ).long()
+
+            dropout_start *= dropout_active
+            dropout_tail *= dropout_active
+
+            G = (
+                G
+                + dropout_start * (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)
 
-        # Initial recurrent state
+        # 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]
@@ -595,42 +636,6 @@ 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