# 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
# 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:
##############################
-# This is one order of magnitude more complicated than I expected
-
-
-def flash_back_time_src(N, H, t0, t1, CL, CH, proba, device):
- # starting flash backs
- fb_start = (torch.rand(N, CH, t1 - t0, device=device) <= proba).long()
- fb_start[:, :, -CL:] = 0
- fb_start[:, :, :CL] = 0
-
- # Remove series longer than CL
- fb_body = fb_start.clone()
- fb_body[:, :, CL + 1 :] -= fb_start[:, :, : -(CL + 1)]
- fb_body = fb_body.cumsum(dim=2)
- fb_start = fb_start * (fb_body == 1)
-
- # pick past starting source times
- src_time = (
- fb_start
- * (
- torch.rand(fb_start.size(), device=fb_start.device)
- * (torch.arange(fb_start.size(2), device=fb_start.device) - CL)[
- None, None, :
- ]
- ).long()
- )
- src_time[:, :, CL:] -= src_time.clone()[:, :, :-CL]
- src_time = src_time.cumsum(dim=2)
-
- src_head = fb_start * torch.randint(H, fb_start.size(), device=fb_start.device)
- src_head[:, :, CL:] -= src_head.clone()[:, :, :-CL]
- src_head = src_head.cumsum(dim=2)
-
- # combine
- src_delta = fb_start.clone()
- src_delta[:, :, CL:] -= fb_start[:, :, :-CL]
- src_delta = src_delta.cumsum(dim=2)
- src_delta[:, :, CL:] -= CL * fb_start[:, :, :-CL]
- src_time += src_delta.cumsum(dim=2) - 1
-
- return src_time, src_head
-
-
-def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
- N, H, CH = V.size(0), V.size(1), rec_V.size(1)
-
- fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device)
-
- fbt_V = fbt[:, :, :, None].expand_as(rec_V[:, :, t0:t1])
- fbh_V = fbh[:, :, :, None].expand_as(rec_V[:, :, t0:t1])
- t = fbt_V.clamp(min=0)
- n = torch.arange(V.size(0), device=V.device)[:, None, None, None].expand_as(
- rec_V[:, :, t0:t1]
- )
- d = torch.arange(V.size(3), device=V.device)[None, None, None, :].expand_as(
- rec_V[:, :, t0:t1]
- )
- q = V[:, :, t0:t1][n, fbh_V, t, d]
- rec_V[:, :, t0:t1] = q * (fbt_V >= 0) + rec_V[:, :, t0:t1] * (fbt_V < 0)
-
- fbt_K = fbt[:, :, :, None].expand_as(rec_K[:, :, t0:t1])
- fbh_K = fbh[:, :, :, None].expand_as(rec_K[:, :, t0:t1])
- t = fbt_K.clamp(min=0)
- n = torch.arange(K.size(0), device=K.device)[:, None, None, None].expand_as(
- rec_K[:, :, t0:t1]
- )
- d = torch.arange(K.size(3), device=K.device)[None, None, None, :].expand_as(
- rec_K[:, :, t0:t1]
- )
- q = K[:, :, t0:t1][n, fbh_K, t, d]
- rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0)
-
- # print("SANITY", (fbt_K >=0).float().sum()/fbt_K.numel())
-
-
-######################################################################
-
class Caterpillar(nn.Module):
def __init__(
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.b_G = nn.Parameter(
torch.full(
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)
DM = self.w_O.size(1)
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)
# 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)
self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
- warnings.warn("flash back", RuntimeWarning)
- if self.training:
- insert_flash_back(self.rec_V, V, self.rec_K, K, t0, t1, CL, proba=1e-2 / CL)
+ 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
):
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
projects / mygptrnn.git / blobdiff