X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=633ad642c19a3045064ef858c0ee494a7c733425;hb=6e87fe0cb8bd8a0042bbf7b2ede9d8ed0372fb6b;hp=d8fd227f63c39a70dded3c55f3c230c3a9d58862;hpb=3e4af6d54fb3d7bd6794035cb79e30ecdcadeb6f;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index d8fd227..633ad64 100755 --- a/mygpt.py +++ b/mygpt.py @@ -476,8 +476,10 @@ class Caterpillar(nn.Module): 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 @@ -497,8 +499,16 @@ class Caterpillar(nn.Module): 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 @@ -520,22 +530,22 @@ class Caterpillar(nn.Module): 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) @@ -546,49 +556,102 @@ class Caterpillar(nn.Module): # 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 - # heads hit the same row + ###################################################################### + # Roll the gating indexes - G = G / G.sum(1, keepdim=True).clamp(min=1) + warnings.warn("rotating barrel", RuntimeWarning) - # We prepare the arguments for the parallel scan + # print(f"SANITY2 {N=} {H=} {R=} {t0=} {t1=} {G.size()=}") - A = 1 - G.sum(1) - gated_V = torch.einsum("nhet,nhtd->netd", G, V) - gated_K = torch.einsum("nhet,nhtd->netd", G, K) + n_barrel = torch.arange(N, device=G.device)[:, None, None, None] + h_barrel = torch.arange(H, device=G.device)[None, :, None, None] + 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 - # We start from cached values, which matters in inference + # GG = G.gather(dim=2,index=r_barrel) + G = G[n_barrel, h_barrel, r_barrel, t_barrel] - init_rec_V = self.rec_V[:, :, t0 - CL : t0] - init_rec_K = self.rec_K[:, :, t0 - CL : t0] + # print("SANITY", (GG-G).abs()) + # exit(0) ###################################################################### + # 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 + + # 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) + + 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) @@ -606,14 +669,14 @@ class Caterpillar(nn.Module): # 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",