X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=f97af49bbb60edb3eca5d26ab96c6e5cccc9dd07;hb=0d25f8a86e80850cf6a6e27d419f7b043c6028f1;hp=90102bf1bf1328e94a5b44ac1af5bc148211befe;hpb=4395f9a90218819997c706de9505cda1c86ad507;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index 90102bf..f97af49 100755 --- a/mygpt.py +++ b/mygpt.py @@ -16,7 +16,6 @@ import torch, einops from torch import nn from torch.nn import functional as F -from functorch.dim import dims import ffutils @@ -182,7 +181,7 @@ def nsum_shape(X, Y_init): class DumbRec(nn.Module): def __init__( self, - dim_in, + dim_model, dim_qk, dim_v, nb_heads, @@ -200,11 +199,11 @@ class DumbRec(nn.Module): self.k_star = randw(nb_lines, dim_qk) - self.w_qw = randw(nb_heads, dim_qk, dim_in) - self.w_qr = randw(nb_heads, dim_qk, dim_in) - # self.w_k = randw(nb_heads, dim_qk, dim_in) - self.w_v = randw(nb_heads, dim_v, dim_in) - self.w_o = randw(dim_v * nb_heads, dim_in) + self.w_qw = randw(nb_heads, dim_qk, dim_model) + self.w_qr = randw(nb_heads, dim_qk, dim_model) + # self.w_k = randw(nb_heads, dim_qk, dim_model) + self.w_v = randw(nb_heads, dim_v, dim_model) + self.w_o = randw(dim_v * nb_heads, dim_model) def reset_inner_loss(self): self.acc_attention = 0 @@ -311,7 +310,7 @@ class DumbRec(nn.Module): class KVRec(nn.Module): def __init__( self, - dim_in, + dim_model, dim_qk, dim_v, nb_heads, @@ -329,11 +328,11 @@ class KVRec(nn.Module): self.k_star = randw(nb_lines, dim_qk) - self.w_qw = randw(nb_heads, dim_qk, dim_in) - self.w_qr = randw(nb_heads, dim_qk, dim_in) - self.w_k = randw(nb_heads, dim_qk, dim_in) - self.w_v = randw(nb_heads, dim_v, dim_in) - self.w_o = randw(dim_v * nb_heads, dim_in) + self.w_qw = randw(nb_heads, dim_qk, dim_model) + self.w_qr = randw(nb_heads, dim_qk, dim_model) + self.w_k = randw(nb_heads, dim_qk, dim_model) + self.w_v = randw(nb_heads, dim_v, dim_model) + self.w_o = randw(dim_v * nb_heads, dim_model) def reset_inner_loss(self): self.acc_attention = 0 @@ -352,8 +351,6 @@ class KVRec(nn.Module): def forward(self, bs): x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb - # n,h,l,t,d = dims(5) - if bs.init_cache: self.rec_v = x_q.new_zeros( x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1) @@ -444,6 +441,11 @@ class KVRec(nn.Module): ############################## +# Returns a tensor with an additional index at rank win_dim, that move +# along the same dimension as dim, on a domain {0...win_size-1}, and +# dim is restricted on a domain reduced by win_size-1 values. + + def moving_window(x, dim, win_dim, win_size): size, stride = x.size(), x.stride() size = size[:dim] + (size[dim] - win_size + 1,) + size[dim + 1 :] @@ -455,11 +457,87 @@ def moving_window(x, dim, win_dim, win_size): ############################## +# This is one order of magnitude more complicated than I expected + + +def flash_back_time_src(N, H, t0, t1, CL, CH, proba, device): + # starting flash backs + fb_start = (torch.rand(N, CH, t1 - t0, device=device) <= proba).long() + fb_start[:, :, -CL:] = 0 + fb_start[:, :, :CL] = 0 + + # Remove series longer than CL + fb_body = fb_start.clone() + fb_body[:, :, CL + 1 :] -= fb_start[:, :, : -(CL + 1)] + fb_body = fb_body.cumsum(dim=2) + fb_start = fb_start * (fb_body == 1) + + # pick past starting source times + src_time = ( + fb_start + * ( + torch.rand(fb_start.size(), device=fb_start.device) + * (torch.arange(fb_start.size(2), device=fb_start.device) - CL)[ + None, None, : + ] + ).long() + ) + src_time[:, :, CL:] -= src_time.clone()[:, :, :-CL] + src_time = src_time.cumsum(dim=2) + + src_head = fb_start * torch.randint(H, fb_start.size(), device=fb_start.device) + src_head[:, :, CL:] -= src_head.clone()[:, :, :-CL] + src_head = src_head.cumsum(dim=2) + + # combine + src_delta = fb_start.clone() + src_delta[:, :, CL:] -= fb_start[:, :, :-CL] + src_delta = src_delta.cumsum(dim=2) + src_delta[:, :, CL:] -= CL * fb_start[:, :, :-CL] + src_time += src_delta.cumsum(dim=2) - 1 + + return src_time, src_head + + +def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba): + N, H, CH = V.size(0), V.size(1), rec_V.size(1) + + fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device) + + fbt_V = fbt[:, :, :, None].expand_as(rec_V[:, :, t0:t1]) + fbh_V = fbh[:, :, :, None].expand_as(rec_V[:, :, t0:t1]) + t = fbt_V.clamp(min=0) + n = torch.arange(V.size(0), device=V.device)[:, None, None, None].expand_as( + rec_V[:, :, t0:t1] + ) + d = torch.arange(V.size(3), device=V.device)[None, None, None, :].expand_as( + rec_V[:, :, t0:t1] + ) + q = V[:, :, t0:t1][n, fbh_V, t, d] + rec_V[:, :, t0:t1] = q * (fbt_V >= 0) + rec_V[:, :, t0:t1] * (fbt_V < 0) + + fbt_K = fbt[:, :, :, None].expand_as(rec_K[:, :, t0:t1]) + fbh_K = fbh[:, :, :, None].expand_as(rec_K[:, :, t0:t1]) + t = fbt_K.clamp(min=0) + n = torch.arange(K.size(0), device=K.device)[:, None, None, None].expand_as( + rec_K[:, :, t0:t1] + ) + d = torch.arange(K.size(3), device=K.device)[None, None, None, :].expand_as( + rec_K[:, :, t0:t1] + ) + q = K[:, :, t0:t1][n, fbh_K, t, d] + rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0) + + # print("SANITY", (fbt_K >=0).float().sum()/fbt_K.numel()) + + +###################################################################### + class Caterpillar(nn.Module): def __init__( self, - dim_in, + dim_model, dim_qk, dim_v, nb_heads, @@ -479,17 +557,17 @@ class Caterpillar(nn.Module): self.caterpillar_height = caterpillar_height self.attention_dropout = attention_dropout - self.w_G = randw(nb_heads, caterpillar_height, dim_in) + 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.w_K = randw(nb_heads, dim_qk, dim_in) - self.w_V = randw(nb_heads, dim_v, dim_in) - self.w_Q = randw(nb_heads, dim_qk, dim_in) - self.w_O = randw(dim_v * nb_heads, dim_in) + self.w_K = randw(nb_heads, dim_qk, dim_model) + self.w_V = randw(nb_heads, dim_v, dim_model) + 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) @@ -512,7 +590,7 @@ class Caterpillar(nn.Module): T = bs.x.size(1) DV = self.w_V.size(1) DK = self.w_K.size(1) - Dout = self.w_O.size(1) + DM = self.w_O.size(1) CH = self.caterpillar_height CL = self.caterpillar_length @@ -520,23 +598,39 @@ class Caterpillar(nn.Module): t0 >= CL and (t1 - t0) % CL == 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_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :] self.rec_K = X.new_zeros(N, CH, 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.cache_Y = X.new_zeros(N, T, Dout) + + 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. The CH gating values are independent, which means + # that the current K/V could 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] ).sigmoid() + # That bas a bad idea + # G = F.dropout(G, self.attention_dropout, self.training) + V = torch.einsum("ntc,hdc->nhtd", X, self.w_V) K = torch.einsum("ntc,hdc->nhtd", X, self.w_K) + # We prepare the arguments for the parallel scan + A = 1 - G.sum(1) gated_V = torch.einsum("nhet,nhtd->netd", G, V) gated_K = torch.einsum("nhet,nhtd->netd", G, K) @@ -544,6 +638,12 @@ class Caterpillar(nn.Module): init_rec_V = self.rec_V[:, :, t0 - CL : t0] init_rec_K = self.rec_K[:, :, t0 - CL : t0] + # Here there is a trick: Since the stack at time t is computed + # by updating that at time t-L, the parallel scan operates + # with a period of L. To do so we split the time indexing in + # two axes, the second of size CL, and run the parallel scan + # using the other 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)) @@ -551,38 +651,57 @@ class Caterpillar(nn.Module): next_V = pscan_dim(A, gated_V, init_rec_V, dim=2) next_K = pscan_dim(A, gated_K, init_rec_K, dim=2) + # Put back the sequence index + self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3) self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3) + warnings.warn("flash back", RuntimeWarning) + if self.training: + insert_flash_back(self.rec_V, V, self.rec_K, K, t0, t1, CL, proba=1e-2 / CL) + ###################################################################### # compute the readout Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q) - uv = moving_window( + # We build tensors NxHxTxFxL where N is the sample index, H + # the head, T the time, F 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 ) - uk = moving_window( + windowed_K = moving_window( self.rec_K[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL ) + # We have an attention score for each of the CHxCL values + ar = torch.einsum( "nhtd,nftld->nhtfl", Q, - uk, + windowed_K, ) / math.sqrt(DK) + # softmax can operate only on one dimension, hence the + # flattening + ar = ar.flatten(3).softmax(dim=3).view(ar.size()) ar = F.dropout(ar, self.attention_dropout, self.training) + # Compute the output for each head, flatten to concatenate + Y = torch.einsum( "nhtfl,nftld->nthd", ar, - uv, + windowed_V, ).flatten(2) + # Compute the final output + self.cache_Y[:, t0:t1] = Y @ self.w_O return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache) @@ -594,7 +713,7 @@ class Caterpillar(nn.Module): class QKVAttention(nn.Module): def __init__( self, - dim_in, + dim_model, dim_qk, dim_v, nb_heads=1, @@ -610,10 +729,10 @@ class QKVAttention(nn.Module): self.attention_dropout = attention_dropout self.record_attention = False - self.w_q = randw(nb_heads, dim_qk, dim_in) - self.w_k = randw(nb_heads, dim_qk, dim_in) - self.w_v = randw(nb_heads, dim_v, dim_in) - self.w_o = randw(dim_v * nb_heads, dim_in) + self.w_q = randw(nb_heads, dim_qk, dim_model) + self.w_k = randw(nb_heads, dim_qk, dim_model) + self.w_v = randw(nb_heads, dim_v, dim_model) + self.w_o = randw(dim_v * nb_heads, dim_model) def forward(self, bs): x_q = bs.x @@ -717,7 +836,7 @@ class MyGPT(nn.Module): def attlayer(): if attention_layer == "mha": return QKVAttention( - dim_in=dim_model, + dim_model=dim_model, dim_qk=dim_keys, dim_v=dim_model // nb_heads, nb_heads=nb_heads, @@ -726,7 +845,7 @@ class MyGPT(nn.Module): ) elif attention_layer == "dumbrec": return DumbRec( - dim_in=dim_model, + dim_model=dim_model, dim_qk=dim_keys, dim_v=dim_rec_v, nb_heads=nb_heads, @@ -735,7 +854,7 @@ class MyGPT(nn.Module): ) elif attention_layer == "kvrec": return KVRec( - dim_in=dim_model, + dim_model=dim_model, dim_qk=dim_keys, dim_v=dim_rec_v, nb_heads=nb_heads, @@ -744,7 +863,7 @@ class MyGPT(nn.Module): ) elif attention_layer == "caterpillar": return Caterpillar( - dim_in=dim_model, + dim_model=dim_model, dim_qk=dim_keys, dim_v=dim_rec_v, nb_heads=nb_heads, @@ -884,7 +1003,7 @@ if __name__ == "__main__": print("Basic check.") m = Caterpillar( - dim_in=4, + dim_model=4, dim_qk=3, dim_v=7, nb_heads=1,