# 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:
nb_lines,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ **kwargs,
):
super().__init__()
nb_lines,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ **kwargs,
):
super().__init__()
##############################
-# This is one order of magnitude more complicated than I expected, not
-# elegant, slow, hopefully not buggy
-
-
-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)
-
- # Set a origin source time (starting time of the chunck to copy
- # here) We set it as the current time minus a multiple of CL to be
- # consistent with the "rolling" caterpillar
- t = torch.arange(fb_start.size(2), device=fb_start.device)[None, None, :]
- src_time = fb_start * (
- t
- - CL
- * (
- 1
- + (
- torch.rand(fb_start.size(), device=fb_start.device) * (t // CL - 1)
- ).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]
- fbh_V = fbh[:, :, :, None]
- t = fbt_V.clamp(min=0)
- n = torch.arange(V.size(0), device=V.device)[:, None, None, None]
- d = torch.arange(V.size(3), device=V.device)[None, None, None, :]
- 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]
- fbh_K = fbh[:, :, :, None]
- t = fbt_K.clamp(min=0)
- n = torch.arange(K.size(0), device=K.device)[:, None, None, None]
- d = torch.arange(K.size(3), device=K.device)[None, None, None, :]
- q = K[:, :, t0:t1][n, fbh_K, t, d]
- rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0)
-
-
-######################################################################
+class Calibrator:
+ def __init__(self, w=None, b=None):
+ self.w = w
+ self.b = b
+ self.s, self.s_sq, self.n = 0, 0, 0
+ self.mean, self.std = 0, 0
+
+ def update(self, X):
+ X = X.detach()
+ self.s += X.sum(dim=0)
+ self.s_sq += X.pow(2).sum(dim=0)
+ self.n += X.size(0)
+
+ def moments(self):
+ mean = self.s / self.n
+ std = (self.s_sq / self.n - mean * mean).sqrt()
+ return mean, std
+
+ def normalize(self):
+ mean, std = self.moments()
+ if self.b is not None:
+ self.b.sub_(mean)
+ if self.w is not None:
+ self.w.div_(std)
+ result = mean - self.mean, std - self.std
+ self.mean, self.std = mean, std
+ self.s, self.s_sq, self.n = 0, 0, 0
+ return result
class Caterpillar(nn.Module):
caterpillar_height,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ **kwargs,
):
super().__init__()
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
- warnings.warn("flash back", RuntimeWarning)
- self.proba_flashback = 0.1
+ ######################################################################
+ # 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.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.b_G = nn.Parameter(torch.full((nb_heads, caterpillar_height), default_bg))
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)
+ self.init_K_rec = randw(
+ caterpillar_height,
+ caterpillar_length,
+ dim_qk,
+ )
+ self.init_V_rec = randw(
+ caterpillar_height,
+ caterpillar_length,
+ dim_v,
+ )
+
+ self.calibrator_G = Calibrator()
+ self.calibrator_rec_V = Calibrator()
+ self.calibrator_rec_K = Calibrator()
def reset_inner_loss(self):
self.acc_attention = 0
DV = self.w_V.size(1)
DK = self.w_K.size(1)
DM = self.w_O.size(1)
- CH = self.caterpillar_height
- CL = self.caterpillar_length
+ R = self.caterpillar_height
+ L = self.caterpillar_length
assert (
- t0 >= CL and (t1 - t0) % CL == 0
+ t0 >= L and (t1 - t0) % L == 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_K = X.new_zeros(N, CH, T, DK)
+ self.rec_V = X.new_zeros(N, R, T, DV)
+ self.rec_K = X.new_zeros(N, R, 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.rec_V[:, :, t0 - L : t0, :] = self.init_V_rec[None, :, :, :]
+ self.rec_K[:, :, t0 - L : t0, :] = self.init_K_rec[None, :, :, :]
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
# 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
+ # the new key and value in the R pairs of the current
+ # stack. There are R independent gating values, which means
+ # that the current K/V may 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]
+ torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
).sigmoid()
- # That bas a bad idea
- # G = F.dropout(G, self.attention_dropout, self.training)
+ self.calibrator_G.update(G.reshape(-1, G.size(-1)))
- V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
- K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
+ # warnings.warn("softmax gating", RuntimeWarning)
- # We prepare the arguments for the parallel scan
+ # G = (
+ # torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
+ # ).softmax(dim=2)
- A = 1 - G.sum(1)
- gated_V = torch.einsum("nhet,nhtd->netd", G, V)
- gated_K = torch.einsum("nhet,nhtd->netd", G, K)
+ ######################################################################
+ # The "flashbacks"
- init_rec_V = self.rec_V[:, :, t0 - CL : t0]
- init_rec_K = self.rec_K[:, :, t0 - CL : t0]
+ if self.training and self.proba_gate_dropout > 0.0:
+ # This is a better implementation of "flashbacks".
- # 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.
+ # G is NxHxExT where e is the caterpillar's row.
- A = A.unflatten(2, (-1, CL))
- gated_V = gated_V.unflatten(2, (-1, CL))
- gated_K = gated_K.unflatten(2, (-1, CL))
+ warnings.warn("gate dropout", RuntimeWarning)
- next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
- next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
+ kill = (
+ torch.rand(G.size(), device=G.device) <= self.proba_gate_dropout
+ ).float()
+
+ alpha = G / (1 - self.proba_gate_dropout)
- # Put back the sequence index
+ G = alpha * (1 - kill)
- self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
- self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
+ ######################################################################
+ # Clip the gating to avoid values greater than 1 when several
+ # heads hit the same row
- if self.training and self.proba_flashback > 0.0:
- # insert_flash_back(self.rec_V,V,self.rec_K,K,t0,t1,CL,proba=self.proba_flashback / CL,)
+ G = G / G.sum(1, keepdim=True).clamp(min=1)
- # 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
+ ######################################################################
+ # Roll the gating indexes
- 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, :]
+ # warnings.warn("rotating barrel", RuntimeWarning)
- u = (
- torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
- ) * CL
+ # r_barrel = torch.arange(R, device=G.device)[None, None, :, None]
+ # t_barrel = torch.arange(t1 - t0, device=G.device)[None, None, None, :]
+ # r_barrel = (r_barrel + (t_barrel + t0) // L) % R
+ # G = G.gather(dim=2, index=r_barrel.expand_as(G))
- src_time = t - u - t0
- src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
+ # We prepare the arguments for the parallel scan
- mask_V = (
- torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
- ).long()
- self.rec_V[:, :, t0:t1] = (
- mask_V * V[n, src_head, src_time, dv]
- + (1 - mask_V) * self.rec_V[:, :, t0:t1]
- )
+ A = 1 - G.sum(1)
- mask_K = (
- torch.rand(N, CH, t1 - t0, DK, device=X.device) <= self.proba_flashback
- ).long()
- self.rec_K[:, :, t0:t1] = (
- mask_K * K[n, src_head, src_time, dk]
- + (1 - mask_K) * self.rec_K[:, :, t0:t1]
- )
+ # warnings.warn("harmonic recurrence", RuntimeWarning)
+ # har = torch.arange(t0, t1, device = G.device).float() + 1
+ # A = har / (har + 1)
+ # G = G / har
+
+ gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V)
+ gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K)
+
+ # We start from cached values, which matters in inference
+
+ 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
+ # sequence indexing in two axes, the second of size L, and
+ # run the parallel scan using the first as the sequence index.
+
+ A = A.unflatten(2, (-1, L))
+ gated_V = gated_V.unflatten(2, (-1, L))
+ gated_K = gated_K.unflatten(2, (-1, L))
+
+ next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
+ next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
+
+ next_V = next_V.flatten(2, 3)
+ next_K = next_K.flatten(2, 3)
+
+ self.calibrator_rec_V.update(
+ next_V.permute(0, 1, 3, 2).reshape(-1, next_V.size(2))
+ )
+ self.calibrator_rec_K.update(
+ next_K.permute(0, 1, 3, 2).reshape(-1, next_K.size(2))
+ )
+
+ self.rec_V[:, :, t0:t1] = next_V
+ self.rec_K[:, :, t0:t1] = next_K
######################################################################
# compute the readout
# the column in the caterpillar
windowed_V = moving_window(
- self.rec_V[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
+ self.rec_V[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
)
windowed_K = moving_window(
- self.rec_K[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
+ self.rec_K[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
)
- # We have an attention score for each of the CHxCL values
+ # We have an attention score for each of the RxL values
ar = torch.einsum(
"nhtd,nftld->nhtfl",
nb_heads=1,
causal=False,
attention_dropout=0.0,
+ logger=print,
+ **kwargs,
):
super().__init__()
nb_blocks,
nb_lines=None,
caterpillar_height=None,
- dim_rec_v=-1,
causal=False,
dropout=0.0,
len_max=1e5,
attention_layer="kvrec",
+ logger=print,
+ **kwargs,
):
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
nb_heads=nb_heads,
causal=causal,
attention_dropout=dropout,
+ logger=logger,
+ **kwargs,
)
elif attention_layer == "dumbrec":
return DumbRec(
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,
+ logger=logger,
+ **kwargs,
)
elif attention_layer == "kvrec":
return KVRec(
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,
+ logger=logger,
+ **kwargs,
)
elif attention_layer == "caterpillar":
return Caterpillar(
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,
attention_dropout=dropout,
+ logger=logger,
+ **kwargs,
)
else:
raise ValueError(f"Unknown attention type {attention_layer}.")