X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=b137cdbca1cd085ee8dec0185c514118e928b5d3;hb=9112db2ed7d8c262c4ef8298cf6637515675f967;hp=7c8e9f4c894ad332e808d07f008ac4c569046bd1;hpb=f0ea1f2375fa3a0be38970a58185cddee97dccef;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index 7c8e9f4..b137cdb 100755 --- a/mygpt.py +++ b/mygpt.py @@ -126,7 +126,6 @@ class AddPositionalEncoding(nn.Module): import pscan - # X is /.../xTxD A is /.../xT Y_init is /.../xD @@ -147,6 +146,18 @@ def pscan_dim(A, X, Y_init, dim=-2): 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]) @@ -190,6 +201,8 @@ class DumbRec(nn.Module): nb_lines, attention_dropout=0.0, len_max=1e5, + logger=print, + args=None, ): super().__init__() @@ -319,6 +332,8 @@ class KVRec(nn.Module): nb_lines, attention_dropout=0.0, len_max=1e5, + logger=print, + args=None, ): super().__init__() @@ -471,6 +486,8 @@ class Caterpillar(nn.Module): caterpillar_height, attention_dropout=0.0, len_max=1e5, + logger=print, + args=None, ): super().__init__() @@ -485,14 +502,15 @@ class Caterpillar(nn.Module): 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 + self.gate_dropout_replace = args.gate_dropout_replace - self.w_G = randw(nb_heads, caterpillar_height, dim_model, amplitude=1e-5) - self.b_G = nn.Parameter( - torch.full( - (nb_heads, caterpillar_height), -math.log(caterpillar_height - 1) - ) - ) + ###################################################################### + + 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), default_bg)) self.w_K = randw(nb_heads, dim_qk, dim_model) self.w_V = randw(nb_heads, dim_v, dim_model) @@ -500,20 +518,24 @@ class Caterpillar(nn.Module): self.w_O = randw(dim_v * nb_heads, dim_model) self.init_K_rec = randw( - caterpillar_height, caterpillar_length, dim_qk, amplitude=1e-5 + caterpillar_height, + caterpillar_length, + dim_qk, ) self.init_V_rec = randw( - caterpillar_height, caterpillar_length, dim_v, amplitude=1e-5 + caterpillar_height, + caterpillar_length, + dim_v, ) - def reset_inner_loss(self): - self.acc_attention = 0 - self.acc_nb = 0 + # def reset_inner_loss(self): + # self.acc_attention = 0 + # self.acc_nb = 0 - def get_inner_loss(self): - # warnings.warn("l2 regularization", RuntimeWarning) - # return (self.acc_attention / self.acc_nb).pow(2).sum() - return torch.tensor([0], device=self.w_Q.device) + # def get_inner_loss(self): + # warnings.warn("l2 regularization", RuntimeWarning) + # return (self.acc_attention / self.acc_nb).pow(2).sum() + # return torch.tensor([0], device=self.w_Q.device) def forward(self, bs): # Dimensions to make the source a bit clearer, that's needed @@ -526,22 +548,22 @@ 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) @@ -552,8 +574,8 @@ 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. @@ -561,102 +583,107 @@ class Caterpillar(nn.Module): 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 - - dropout_start = ( - ( - torch.rand(G.size(), device=G.device) - .flatten(2, 3) - .sort(dim=2) - .indices - == 0 - ) - .unflatten(2, (CH, t1 - t0)) - .float() - ) + gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V) + gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K) - dropout_tail = dropout_start.cumsum(dim=3) - dropout_start + # We start from cached values, which matters in inference - dropout_active = ( - torch.rand(N, 1, 1, 1, device=G.device) < self.proba_gate_dropout - ).long() + init_rec_V = self.rec_V[:, :, t0 - L : t0] + init_rec_K = self.rec_K[:, :, t0 - L : t0] - dropout_start *= dropout_active - dropout_tail *= dropout_active + # 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. - G = ( - G - + dropout_start * (1 - epsilon - G.detach()) - - dropout_tail * G.detach() - ) + 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).flatten(2, 3) + next_K = pscan_dim(A, gated_K, init_rec_K, dim=2).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] + if self.gate_dropout_sync: + shape_kill = (N, 1, 1) + else: + shape_kill = (N, H, R) + + # Pick a point in each of the NxHxR timeline and set this + # entry and the following to 1 + kill = ( + torch.rand(*shape_kill, t1 - t0, device=G.device).sort(dim=3).indices + == 0 + ).cumsum(dim=3) + + # Keep these mask for only some of the NxHxR + kill = kill * ( + torch.rand(*shape_kill, 1, device=G.device) <= self.gate_dropout_proba + ) - ################################################################# - # Associative scan + # The coefficient to keep are the complementary + mask = 1 - kill - # 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. + masked_next_V, masked_next_K = recurrence(G * mask, V, K) - A = A.unflatten(2, (-1, CL)) - gated_V = gated_V.unflatten(2, (-1, CL)) - gated_K = gated_K.unflatten(2, (-1, CL)) + if self.gate_dropout_replace: + next_V = next_V.detach() + next_K = next_K.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 + (masked_next_V - masked_next_V.detach()) / ( + 1 - self.gate_dropout_proba + ) + next_K = next_K + (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) @@ -695,6 +722,8 @@ class QKVAttention(nn.Module): nb_heads=1, causal=False, attention_dropout=0.0, + logger=print, + args=None, ): super().__init__() @@ -786,6 +815,8 @@ class MyGPT(nn.Module): dropout=0.0, len_max=1e5, attention_layer="kvrec", + logger=print, + args=None, ): super().__init__() @@ -822,6 +853,8 @@ class MyGPT(nn.Module): nb_heads=nb_heads, causal=causal, attention_dropout=dropout, + logger=logger, + args=args, ) elif attention_layer == "dumbrec": return DumbRec( @@ -831,6 +864,8 @@ class MyGPT(nn.Module): nb_heads=nb_heads, nb_lines=nb_lines, attention_dropout=dropout, + logger=logger, + args=args, ) elif attention_layer == "kvrec": return KVRec( @@ -840,6 +875,8 @@ class MyGPT(nn.Module): nb_heads=nb_heads, nb_lines=nb_lines, attention_dropout=dropout, + logger=logger, + args=args, ) elif attention_layer == "caterpillar": return Caterpillar( @@ -850,6 +887,8 @@ class MyGPT(nn.Module): 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}.")