import pscan
-
# X is /.../xTxD A is /.../xT Y_init is /.../xD
return Y
+def pscan_rgrad(grad_Y, A, X, Y_init, dim=-2, eps=1e-2):
+ with torch.no_grad():
+ s_A, s_X = 0, 0
+ for t in range(X.size(dim) - 1, 0, -1):
+ delta = (grad_Y[t] - s_A) / A[t].grad
+ s_A += A[t].grad * delta
+ A[t].grad = delta
+ delta = (grad_Y[t] - s_X) / X[t].grad
+ s_X += X[t].grad * delta
+ X[t].grad = delta
+
+
def pscan_shape(A, X, Y_init):
s = X.size()
A = A.reshape(-1, s[-2])
nb_lines,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ args=None,
):
super().__init__()
nb_lines,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ args=None,
):
super().__init__()
caterpillar_height,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ args=None,
):
super().__init__()
self.caterpillar_height = caterpillar_height
self.attention_dropout = attention_dropout
- self.proba_gate_dropout = 0.0
+ self.gate_dropout_proba = args.gate_dropout_proba
+ self.gate_dropout_sync = args.gate_dropout_sync
+ ######################################################################
+
+ default_bg = -math.log(caterpillar_height - 1)
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)
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.
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
+ # heads hit the same row
+
+ G = G / G.sum(1, keepdim=True).clamp(min=1)
+
######################################################################
- # The "flashbacks"
- if self.training and self.proba_gate_dropout > 0.0:
- # This is a better implementation of "flashbacks".
+ def recurrence(G, V, K):
+ # We prepare the arguments for the parallel scan
- # G is NxHxExT where e is the caterpillar's row.
+ A = 1 - G.sum(1)
- warnings.warn("gate dropout", RuntimeWarning)
- epsilon = 0.5
+ gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V)
+ gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K)
- dropout_head = (
- (torch.rand(N, H, 1, t1 - t0, device=G.device).sort(dim=3).indices == 0)
- .expand_as(G)
- .float()
- )
+ # We start from cached values, which matters in inference
- dropout_tail = dropout_head.cumsum(dim=3) - dropout_head
+ init_rec_V = self.rec_V[:, :, t0 - L : t0]
+ init_rec_K = self.rec_K[:, :, t0 - L : t0]
- dropout_active = (
- torch.rand(N, 1, 1, 1, device=G.device) < self.proba_gate_dropout
- ).long()
+ # 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.
- dropout_head *= dropout_active
- dropout_tail *= dropout_active
+ A = A.unflatten(2, (-1, L))
+ gated_V = gated_V.unflatten(2, (-1, L))
+ gated_K = gated_K.unflatten(2, (-1, L))
- G = (
- G
- # + dropout_head * (1 - epsilon - G.detach())
- - dropout_tail * G.detach()
- )
+ 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)
- # We prepare the arguments for the parallel scan
+ return next_V, next_K
- # Clip the gating to avoid values greater than 1 when several
- # heads hit the same row
+ #################################################################
- G = G / G.sum(1, keepdim=True).clamp(min=1)
+ next_V, next_K = recurrence(G, V, K)
- A = 1 - G.sum(1)
- gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V)
- gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K)
+ if self.training and self.gate_dropout_proba > 0.0:
+ # G is NxHxRxT where r is the caterpillar's row.
- # We start from cached values, which matters in inference
+ warnings.warn("gate dropout", RuntimeWarning)
- init_rec_V = self.rec_V[:, :, t0 - CL : t0]
- init_rec_K = self.rec_K[:, :, t0 - CL : t0]
+ # Pick a point in each of the NxHxR timeline and set this
+ # entry and the following to 1
+ kill = (
+ torch.rand(N, H, R, t1 - t0, device=G.device).sort(dim=3).indices == 0
+ ).cumsum(dim=3)
- #################################################################
- # Associative scan
+ # Keep these mask for only some of the NxHxR
+ kill = kill * (
+ torch.rand(N, H, R, 1, device=G.device) <= self.gate_dropout_proba
+ )
- # 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
- # run the parallel scan using the first as the sequence index.
+ # The coefficient to keep are the complementary
+ mask = 1 - kill
- A = A.unflatten(2, (-1, CL))
- gated_V = gated_V.unflatten(2, (-1, CL))
- gated_K = gated_K.unflatten(2, (-1, CL))
+ masked_next_V, masked_next_K = recurrence(G * mask, V, K)
- 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.detach() + (masked_next_V - masked_next_V.detach()) / (
+ 1 - self.gate_dropout_proba
+ )
+ next_K = next_K.detach() + (masked_next_K - masked_next_K.detach()) / (
+ 1 - self.gate_dropout_proba
+ )
- self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
- self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
+ self.rec_V[:, :, t0:t1] = next_V
+ self.rec_K[:, :, t0:t1] = next_K
######################################################################
# compute the readout
Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
- # We build tensors NxHxTxFxL where N is the sample index, H
- # the head, T the time, F the row in the caterpillar, and L
+ # We build tensors NxHxTxRxL where N is the sample index, H
+ # the head, T the time, R 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
+ 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",
+ "nhtd,nrtld->nhtrl",
Q,
windowed_K,
) / math.sqrt(DK)
nb_heads=1,
causal=False,
attention_dropout=0.0,
+ logger=print,
+ args=None,
):
super().__init__()
dropout=0.0,
len_max=1e5,
attention_layer="kvrec",
+ logger=print,
+ args=None,
):
super().__init__()
nb_heads=nb_heads,
causal=causal,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
elif attention_layer == "dumbrec":
return DumbRec(
nb_heads=nb_heads,
nb_lines=nb_lines,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
elif attention_layer == "kvrec":
return KVRec(
nb_heads=nb_heads,
nb_lines=nb_lines,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
elif attention_layer == "caterpillar":
return Caterpillar(
caterpillar_length=self.caterpillar_length,
caterpillar_height=self.caterpillar_height,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
else:
raise ValueError(f"Unknown attention type {attention_layer}.")
projects / mygptrnn.git / blobdiff