X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=760a3c60e086f094d4e86672e040877f94c6a6d2;hb=60cd4f63c55c77f5097ebac146c212763ba11925;hp=aded7967a4c8c0f4d84fdfd39085929b2e42c291;hpb=e56873a0cb64555cbd47e44cdca0ce991765a5fc;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index aded796..760a3c6 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]) @@ -191,7 +202,7 @@ class DumbRec(nn.Module): attention_dropout=0.0, len_max=1e5, logger=print, - **kwargs, + args=None, ): super().__init__() @@ -322,7 +333,7 @@ class KVRec(nn.Module): attention_dropout=0.0, len_max=1e5, logger=print, - **kwargs, + args=None, ): super().__init__() @@ -464,36 +475,6 @@ def moving_window(x, dim, win_dim, win_size): ############################## -class Calibrator: - def __init__(self, w=None, b=None): - self.w = w - self.b = b - self.s, self.s_sq, self.n = 0, 0, 0 - self.mean, self.std = 0, 0 - - def update(self, X): - X = X.detach() - self.s += X.sum(dim=0) - self.s_sq += X.pow(2).sum(dim=0) - self.n += X.size(0) - - def moments(self): - mean = self.s / self.n - std = (self.s_sq / self.n - mean * mean).sqrt() - return mean, std - - def normalize(self): - mean, std = self.moments() - if self.b is not None: - self.b.sub_(mean) - if self.w is not None: - self.w.div_(std) - result = mean - self.mean, std - self.std - self.mean, self.std = mean, std - self.s, self.s_sq, self.n = 0, 0, 0 - return result - - class Caterpillar(nn.Module): def __init__( self, @@ -506,7 +487,7 @@ class Caterpillar(nn.Module): attention_dropout=0.0, len_max=1e5, logger=print, - **kwargs, + args=None, ): super().__init__() @@ -521,27 +502,12 @@ class Caterpillar(nn.Module): self.caterpillar_height = caterpillar_height self.attention_dropout = attention_dropout - ###################################################################### - # sup_args - - x = kwargs.get("gate_dropout") - if x is None: - self.proba_gate_dropout = 0.0 - else: - self.proba_gate_dropout = float(x) - - 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_bg {default_bg}") + 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), default_bg)) @@ -561,10 +527,6 @@ class Caterpillar(nn.Module): dim_v, ) - self.calibrator_G = Calibrator() - self.calibrator_rec_V = Calibrator() - self.calibrator_rec_K = Calibrator() - def reset_inner_loss(self): self.acc_attention = 0 self.acc_nb = 0 @@ -620,90 +582,81 @@ class Caterpillar(nn.Module): torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None] ).sigmoid() - self.calibrator_G.update(G.reshape(-1, G.size(-1))) - - # warnings.warn("softmax gating", RuntimeWarning) + # Clip the gating to avoid values greater than 1 when several + # heads hit the same row - # G = ( - # torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None] - # ).softmax(dim=2) + 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) + gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V) + gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K) - kill = ( - torch.rand(G.size(), device=G.device) <= self.proba_gate_dropout - ).float() + # We start from cached values, which matters in inference - alpha = G / (1 - self.proba_gate_dropout) + init_rec_V = self.rec_V[:, :, t0 - L : t0] + init_rec_K = self.rec_K[:, :, t0 - L : t0] - G = alpha * (1 - kill) + # 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. - ###################################################################### - # Clip the gating to avoid values greater than 1 when several - # heads hit the same row + A = A.unflatten(2, (-1, L)) + gated_V = gated_V.unflatten(2, (-1, L)) + gated_K = gated_K.unflatten(2, (-1, L)) - G = G / G.sum(1, keepdim=True).clamp(min=1) + next_V = pscan_dim(A, gated_V, init_rec_V, dim=2) + next_K = pscan_dim(A, gated_K, init_rec_K, dim=2) - ###################################################################### - # Roll the gating indexes + next_V = next_V.flatten(2, 3) + next_K = next_K.flatten(2, 3) - # warnings.warn("rotating barrel", RuntimeWarning) + return next_V, next_K - # 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 - - 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) + next_V, next_K = recurrence(G, V, K) - # We start from cached values, which matters in inference + if self.training and self.gate_dropout_proba > 0.0: + # G is NxHxRxT where r is the caterpillar's row. - init_rec_V = self.rec_V[:, :, t0 - L : t0] - init_rec_K = self.rec_K[:, :, t0 - L : t0] + warnings.warn("gate dropout", RuntimeWarning) - ################################################################# - # Associative scan + if self.gate_dropout_sync: + shape_kill = (N, 1, 1) + else: + shape_kill = (N, H, R) - # 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. + # 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) - A = A.unflatten(2, (-1, L)) - gated_V = gated_V.unflatten(2, (-1, L)) - gated_K = gated_K.unflatten(2, (-1, L)) + # Keep these mask for only some of the NxHxR + kill = kill * ( + torch.rand(*shape_kill, 1, device=G.device) <= self.gate_dropout_proba + ) - 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 coefficient to keep are the complementary + mask = 1 - kill - next_V = next_V.flatten(2, 3) - next_K = next_K.flatten(2, 3) + masked_next_V, masked_next_K = recurrence(G * mask, V, K) - self.calibrator_rec_V.update( - next_V.permute(0, 1, 3, 2).reshape(-1, next_V.size(2)) - ) - self.calibrator_rec_K.update( - next_K.permute(0, 1, 3, 2).reshape(-1, next_K.size(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 self.rec_K[:, :, t0:t1] = next_K @@ -713,8 +666,8 @@ class Caterpillar(nn.Module): 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( @@ -728,7 +681,7 @@ class Caterpillar(nn.Module): # 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) @@ -768,7 +721,7 @@ class QKVAttention(nn.Module): causal=False, attention_dropout=0.0, logger=print, - **kwargs, + args=None, ): super().__init__() @@ -861,7 +814,7 @@ class MyGPT(nn.Module): len_max=1e5, attention_layer="kvrec", logger=print, - **kwargs, + args=None, ): super().__init__() @@ -899,7 +852,7 @@ class MyGPT(nn.Module): causal=causal, attention_dropout=dropout, logger=logger, - **kwargs, + args=args, ) elif attention_layer == "dumbrec": return DumbRec( @@ -910,7 +863,7 @@ class MyGPT(nn.Module): nb_lines=nb_lines, attention_dropout=dropout, logger=logger, - **kwargs, + args=args, ) elif attention_layer == "kvrec": return KVRec( @@ -921,7 +874,7 @@ class MyGPT(nn.Module): nb_lines=nb_lines, attention_dropout=dropout, logger=logger, - **kwargs, + args=args, ) elif attention_layer == "caterpillar": return Caterpillar( @@ -933,7 +886,7 @@ class MyGPT(nn.Module): caterpillar_height=self.caterpillar_height, attention_dropout=dropout, logger=logger, - **kwargs, + args=args, ) else: raise ValueError(f"Unknown attention type {attention_layer}.")