Update.
[mygptrnn.git] / mygpt.py
index 87071c3..95e5527 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:
@@ -181,7 +183,7 @@ 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 +201,11 @@ 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 +312,7 @@ class DumbRec(nn.Module):
 class KVRec(nn.Module):
     def __init__(
         self,
-        dim_in,
+        dim_model,
         dim_qk,
         dim_v,
         nb_heads,
@@ -328,11 +330,11 @@ 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
@@ -441,6 +443,11 @@ class KVRec(nn.Module):
 ##############################
 
 
+# Returns a tensor with an additional index at rank win_dim, that move
+# along the same dimension as dim, on a domain {0...win_size-1}, and
+# dim is restricted on a domain reduced by win_size-1 values.
+
+
 def moving_window(x, dim, win_dim, win_size):
     size, stride = x.size(), x.stride()
     size = size[:dim] + (size[dim] - win_size + 1,) + size[dim + 1 :]
@@ -456,7 +463,7 @@ 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 +483,20 @@ 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.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(
                 (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)
@@ -507,9 +517,10 @@ 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)
-        Dout = self.w_O.size(1)
+        DM = self.w_O.size(1)
         CH = self.caterpillar_height
         CL = self.caterpillar_length
 
@@ -517,23 +528,47 @@ class Caterpillar(nn.Module):
             t0 >= CL and (t1 - t0) % CL == 0
         ), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
 
+        # We cache values to deal efficiently with auto-regression
+
         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)
+
+            self.cache_Y = X.new_zeros(N, T, DM)
 
         ######################################################################
         # 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 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()
 
+        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)
@@ -541,6 +576,12 @@ 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 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))
@@ -548,38 +589,87 @@ 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)
 
+        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
 
         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 +681,7 @@ class Caterpillar(nn.Module):
 class QKVAttention(nn.Module):
     def __init__(
         self,
-        dim_in,
+        dim_model,
         dim_qk,
         dim_v,
         nb_heads=1,
@@ -607,10 +697,10 @@ 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
@@ -684,7 +774,6 @@ class MyGPT(nn.Module):
         nb_blocks,
         nb_lines=None,
         caterpillar_height=None,
-        dim_rec_v=-1,
         causal=False,
         dropout=0.0,
         len_max=1e5,
@@ -692,7 +781,12 @@ 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
@@ -714,7 +808,7 @@ 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,
@@ -723,27 +817,27 @@ 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,
+                    dim_v=dim_model // nb_heads,
                     nb_heads=nb_heads,
                     nb_lines=nb_lines,
                     attention_dropout=dropout,
                 )
             elif attention_layer == "kvrec":
                 return KVRec(
-                    dim_in=dim_model,
+                    dim_model=dim_model,
                     dim_qk=dim_keys,
-                    dim_v=dim_rec_v,
+                    dim_v=dim_model // nb_heads,
                     nb_heads=nb_heads,
                     nb_lines=nb_lines,
                     attention_dropout=dropout,
                 )
             elif attention_layer == "caterpillar":
                 return Caterpillar(
-                    dim_in=dim_model,
+                    dim_model=dim_model,
                     dim_qk=dim_keys,
-                    dim_v=dim_rec_v,
+                    dim_v=dim_model // nb_heads,
                     nb_heads=nb_heads,
                     caterpillar_length=self.caterpillar_length,
                     caterpillar_height=self.caterpillar_height,
@@ -881,7 +975,7 @@ if __name__ == "__main__":
     print("Basic check.")
 
     m = Caterpillar(
-        dim_in=4,
+        dim_model=4,
         dim_qk=3,
         dim_v=7,
         nb_heads=1,