self.caterpillar_height = caterpillar_height
self.attention_dropout = attention_dropout
- self.proba_gate_dropout = 0.0
+ ######################################################################
+ # sup_args
+
+ x = kwargs.get("gate_dropout")
+ if x is None:
+ self.proba_gate_dropout = 0.0
+ else:
+ self.proba_gate_dropout = float(x)
- default_b_G = kwargs.get("default_b_G")
- if default_b_G is None:
- default_b_G = -math.log(caterpillar_height - 1)
+ 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_b_G {default_b_G}")
+ 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), default_b_G))
+ 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)
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
+ # warnings.warn("softmax gating", RuntimeWarning)
- G = G / G.sum(1, keepdim=True).clamp(min=1)
-
- ######################################################################
- # Roll the gating indexes
-
- # warnings.warn("rotating barrel", RuntimeWarning)
-
- # 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))
+ # G = (
+ # torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
+ # ).softmax(dim=2)
######################################################################
# The "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()
- )
+ kill = (
+ torch.rand(G.size(), device=G.device) <= self.proba_gate_dropout
+ ).float()
- dropout_tail = dropout_head.cumsum(dim=3) - dropout_head
+ alpha = G / (1 - self.proba_gate_dropout)
- dropout_active = (
- torch.rand(N, 1, 1, 1, device=G.device) < self.proba_gate_dropout
- ).long()
+ G = alpha * (1 - kill)
- dropout_head *= dropout_active
- dropout_tail *= dropout_active
+ ######################################################################
+ # Clip the gating to avoid values greater than 1 when several
+ # heads hit the same row
- G = (
- G
- + dropout_head * (1 - epsilon - G.detach())
- - dropout_tail * G.detach()
- )
+ G = G / G.sum(1, keepdim=True).clamp(min=1)
######################################################################
+ # Roll the gating indexes
+
+ # warnings.warn("rotating barrel", RuntimeWarning)
+
+ # 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))
# We prepare the arguments for the parallel scan