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__()
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.proba_gate_dropout = 0.0
+ default_b_G = kwargs.get("default_b_G")
+ if default_b_G is None:
+ default_b_G = -math.log(caterpillar_height - 1)
+
+ logger(f"default_b_G {default_b_G}")
+
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_b_G))
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,
+ )
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)
# 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. There are CH independent gating values, which means
+ # 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()
# Clip the gating to avoid values greater than 1 when several
G = G / G.sum(1, keepdim=True).clamp(min=1)
- # 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)
+ ######################################################################
+ # Roll the gating indexes
- # We start from cached values, which matters in inference
+ # warnings.warn("rotating barrel", RuntimeWarning)
- init_rec_V = self.rec_V[:, :, t0 - CL : t0]
- init_rec_K = self.rec_K[:, :, t0 - CL : t0]
+ # 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))
######################################################################
+ # 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
+ dropout_head = (
+ (torch.rand(N, H, 1, t1 - t0, device=G.device).sort(dim=3).indices == 0)
+ .expand_as(G)
+ .float()
+ )
+
+ dropout_tail = dropout_head.cumsum(dim=3) - dropout_head
+
+ dropout_active = (
+ torch.rand(N, 1, 1, 1, device=G.device) < self.proba_gate_dropout
+ ).long()
+
+ dropout_head *= dropout_active
+ dropout_tail *= dropout_active
+
+ G = (
+ G
+ + dropout_head * (1 - epsilon - G.detach())
+ - dropout_tail * G.detach()
+ )
+
+ ######################################################################
+
+ # We prepare the arguments for the parallel scan
+
+ A = 1 - G.sum(1)
+
+ # 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-CL, the parallel
- # scan operates with a period of CL. To do so we split the
- # sequence indexing in two axes, the second of size CL, and
+ # 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, CL))
- gated_V = gated_V.unflatten(2, (-1, CL))
- gated_K = gated_K.unflatten(2, (-1, CL))
+ 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)
# 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__()
dropout=0.0,
len_max=1e5,
attention_layer="kvrec",
+ logger=print,
+ **kwargs,
):
super().__init__()
nb_heads=nb_heads,
causal=causal,
attention_dropout=dropout,
+ logger=logger,
+ **kwargs,
)
elif attention_layer == "dumbrec":
return DumbRec(
nb_heads=nb_heads,
nb_lines=nb_lines,
attention_dropout=dropout,
+ logger=logger,
+ **kwargs,
)
elif attention_layer == "kvrec":
return KVRec(
nb_heads=nb_heads,
nb_lines=nb_lines,
attention_dropout=dropout,
+ logger=logger,
+ **kwargs,
)
elif attention_layer == "caterpillar":
return Caterpillar(
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}.")