Update.
[mygptrnn.git] / mygpt.py
index 87071c3..33c6fee 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -181,7 +181,7 @@ def nsum_shape(X, Y_init):
 class DumbRec(nn.Module):
     def __init__(
         self,
-        dim_in,
+        dim_model,
         dim_qk,
         dim_v,
         nb_heads,
@@ -199,11 +199,11 @@ class DumbRec(nn.Module):
 
         self.k_star = randw(nb_lines, dim_qk)
 
-        self.w_qw = randw(nb_heads, dim_qk, dim_in)
-        self.w_qr = randw(nb_heads, dim_qk, dim_in)
-        # self.w_k = randw(nb_heads, dim_qk, dim_in)
-        self.w_v = randw(nb_heads, dim_v, dim_in)
-        self.w_o = randw(dim_v * nb_heads, dim_in)
+        self.w_qw = randw(nb_heads, dim_qk, dim_model)
+        self.w_qr = randw(nb_heads, dim_qk, dim_model)
+        # self.w_k = randw(nb_heads, dim_qk, dim_model)
+        self.w_v = randw(nb_heads, dim_v, dim_model)
+        self.w_o = randw(dim_v * nb_heads, dim_model)
 
     def reset_inner_loss(self):
         self.acc_attention = 0
@@ -310,7 +310,7 @@ class DumbRec(nn.Module):
 class KVRec(nn.Module):
     def __init__(
         self,
-        dim_in,
+        dim_model,
         dim_qk,
         dim_v,
         nb_heads,
@@ -328,11 +328,11 @@ class KVRec(nn.Module):
 
         self.k_star = randw(nb_lines, dim_qk)
 
-        self.w_qw = randw(nb_heads, dim_qk, dim_in)
-        self.w_qr = randw(nb_heads, dim_qk, dim_in)
-        self.w_k = randw(nb_heads, dim_qk, dim_in)
-        self.w_v = randw(nb_heads, dim_v, dim_in)
-        self.w_o = randw(dim_v * nb_heads, dim_in)
+        self.w_qw = randw(nb_heads, dim_qk, dim_model)
+        self.w_qr = randw(nb_heads, dim_qk, dim_model)
+        self.w_k = randw(nb_heads, dim_qk, dim_model)
+        self.w_v = randw(nb_heads, dim_v, dim_model)
+        self.w_o = randw(dim_v * nb_heads, dim_model)
 
     def reset_inner_loss(self):
         self.acc_attention = 0
@@ -456,7 +456,7 @@ def moving_window(x, dim, win_dim, win_size):
 class Caterpillar(nn.Module):
     def __init__(
         self,
-        dim_in,
+        dim_model,
         dim_qk,
         dim_v,
         nb_heads,
@@ -476,17 +476,17 @@ class Caterpillar(nn.Module):
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
-        self.w_G = randw(nb_heads, caterpillar_height, dim_in)
+        self.w_G = randw(nb_heads, caterpillar_height, dim_model)
         self.b_G = nn.Parameter(
             torch.full(
                 (nb_heads, caterpillar_height), -math.log(caterpillar_height - 1)
             )
         )
 
-        self.w_K = randw(nb_heads, dim_qk, dim_in)
-        self.w_V = randw(nb_heads, dim_v, dim_in)
-        self.w_Q = randw(nb_heads, dim_qk, dim_in)
-        self.w_O = randw(dim_v * nb_heads, dim_in)
+        self.w_K = randw(nb_heads, dim_qk, dim_model)
+        self.w_V = randw(nb_heads, dim_v, dim_model)
+        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)
@@ -519,14 +519,23 @@ class Caterpillar(nn.Module):
 
         if bs.init_cache:
             self.rec_V = X.new_zeros(N, CH, T, DV)
-            self.rec_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :]
             self.rec_K = X.new_zeros(N, CH, T, DK)
+            # We start the recurrent sequences with optimizable
+            # initial values. No idea if it helps.
+            self.rec_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :]
             self.rec_K[:, :, t0 - CL : t0] = self.init_K_rec[None, :, :, :]
+
             self.cache_Y = X.new_zeros(N, T, Dout)
 
         ######################################################################
         # 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
+        # recurrent state, or not at all.
+
         G = (
             torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None]
         ).sigmoid()
@@ -534,6 +543,8 @@ class Caterpillar(nn.Module):
         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
+
         A = 1 - G.sum(1)
         gated_V = torch.einsum("nhet,nhtd->netd", G, V)
         gated_K = torch.einsum("nhet,nhtd->netd", G, K)
@@ -541,6 +552,12 @@ class Caterpillar(nn.Module):
         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 alone as the sequence index.
+
         A = A.unflatten(2, (-1, CL))
         gated_V = gated_V.unflatten(2, (-1, CL))
         gated_K = gated_K.unflatten(2, (-1, CL))
@@ -548,6 +565,8 @@ 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)
 
@@ -556,30 +575,43 @@ class Caterpillar(nn.Module):
 
         Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
 
-        uv = moving_window(
+        # We build tensors NxHxTxFxL where N is the sample index, H
+        # the head, T the time, F the row in the caterpillar, and L
+        # the column in the caterpillar
+
+        windowed_V = moving_window(
             self.rec_V[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
         )
 
-        uk = moving_window(
+        windowed_K = moving_window(
             self.rec_K[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
         )
 
+        # We have an attention score for each of the CHxCL values
+
         ar = torch.einsum(
             "nhtd,nftld->nhtfl",
             Q,
-            uk,
+            windowed_K,
         ) / math.sqrt(DK)
 
+        # softmax can operate only on one dimension, hence the
+        # flattening
+
         ar = ar.flatten(3).softmax(dim=3).view(ar.size())
 
         ar = F.dropout(ar, self.attention_dropout, self.training)
 
+        # Compute the output for each head, flatten to concatenate
+
         Y = torch.einsum(
             "nhtfl,nftld->nthd",
             ar,
-            uv,
+            windowed_V,
         ).flatten(2)
 
+        # Compute the final output
+
         self.cache_Y[:, t0:t1] = Y @ self.w_O
 
         return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache)
@@ -591,7 +623,7 @@ class Caterpillar(nn.Module):
 class QKVAttention(nn.Module):
     def __init__(
         self,
-        dim_in,
+        dim_model,
         dim_qk,
         dim_v,
         nb_heads=1,
@@ -607,10 +639,10 @@ class QKVAttention(nn.Module):
         self.attention_dropout = attention_dropout
         self.record_attention = False
 
-        self.w_q = randw(nb_heads, dim_qk, dim_in)
-        self.w_k = randw(nb_heads, dim_qk, dim_in)
-        self.w_v = randw(nb_heads, dim_v, dim_in)
-        self.w_o = randw(dim_v * nb_heads, dim_in)
+        self.w_q = randw(nb_heads, dim_qk, dim_model)
+        self.w_k = randw(nb_heads, dim_qk, dim_model)
+        self.w_v = randw(nb_heads, dim_v, dim_model)
+        self.w_o = randw(dim_v * nb_heads, dim_model)
 
     def forward(self, bs):
         x_q = bs.x
@@ -714,7 +746,7 @@ class MyGPT(nn.Module):
         def attlayer():
             if attention_layer == "mha":
                 return QKVAttention(
-                    dim_in=dim_model,
+                    dim_model=dim_model,
                     dim_qk=dim_keys,
                     dim_v=dim_model // nb_heads,
                     nb_heads=nb_heads,
@@ -723,7 +755,7 @@ class MyGPT(nn.Module):
                 )
             elif attention_layer == "dumbrec":
                 return DumbRec(
-                    dim_in=dim_model,
+                    dim_model=dim_model,
                     dim_qk=dim_keys,
                     dim_v=dim_rec_v,
                     nb_heads=nb_heads,
@@ -732,7 +764,7 @@ class MyGPT(nn.Module):
                 )
             elif attention_layer == "kvrec":
                 return KVRec(
-                    dim_in=dim_model,
+                    dim_model=dim_model,
                     dim_qk=dim_keys,
                     dim_v=dim_rec_v,
                     nb_heads=nb_heads,
@@ -741,7 +773,7 @@ class MyGPT(nn.Module):
                 )
             elif attention_layer == "caterpillar":
                 return Caterpillar(
-                    dim_in=dim_model,
+                    dim_model=dim_model,
                     dim_qk=dim_keys,
                     dim_v=dim_rec_v,
                     nb_heads=nb_heads,
@@ -881,7 +913,7 @@ if __name__ == "__main__":
     print("Basic check.")
 
     m = Caterpillar(
-        dim_in=4,
+        dim_model=4,
         dim_qk=3,
         dim_v=7,
         nb_heads=1,