X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=12b3631596b702e5f7d4f5e91d797ac4d6487298;hb=HEAD;hp=6e13ff878bae27c818d337a53a413f3ce1a35180;hpb=aa09a883611d6e323b4e8e678b867097fc13afaf;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index 6e13ff8..12b3631 100755 --- a/mygpt.py +++ b/mygpt.py @@ -10,6 +10,8 @@ # with a caching mechanism for keys and values to avoid a O(N^3) cost # for auto-regression. +# This implementation is equipped with RNN layers to replace the MHA + import math, warnings import torch, einops @@ -19,6 +21,8 @@ from torch.nn import functional as F import ffutils +# from blanket import blanket + # import memload ###################################################################### @@ -37,7 +41,7 @@ import ffutils # 1 for the successive tokens. # # Modules able to process brackets may implement a cache that is -# resetted when the input bracket starts at t=0 +# resetted when init_cache is True class BracketedSequence: @@ -82,6 +86,18 @@ class CacheWrapper(nn.Module): ############################## +class NaNChecker(nn.Module): + def __init__(self, name): + super().__init__() + self.name = name + + def forward(self, bs): + x = bs.x if type(bs) is BracketedSequence else bs + assert not x.isnan().any(), f"${self.name} detected NaN" + assert not x.isinf().any(), f"${self.name} detected Inf" + return bs + + class WithResidual(nn.Module): def __init__(self, *f): super().__init__() @@ -124,7 +140,6 @@ class AddPositionalEncoding(nn.Module): import pscan - # X is /.../xTxD A is /.../xT Y_init is /.../xD @@ -145,6 +160,18 @@ 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]) @@ -188,6 +215,8 @@ class DumbRec(nn.Module): nb_lines, attention_dropout=0.0, len_max=1e5, + logger=print, + args=None, ): super().__init__() @@ -201,19 +230,9 @@ class DumbRec(nn.Module): 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 - 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_qw.device) - def forward(self, bs): x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb @@ -221,61 +240,33 @@ class DumbRec(nn.Module): self.rec_v = x_q.new_zeros( x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1) ) - # self.rec_k = x_q.new_zeros( - # x_q.size(0), self.nb_lines, x_q.size(1), self.w_k.size(1) - # ) self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1)) - ###################################################################### - # Prepare the keys - - k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1) - - warnings.warn("rotating key barrel", RuntimeWarning) - k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1) - t_barrel = torch.arange(t0, t1, device=k_star.device) - t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0) - l_barrel = ( - torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel - ) % k_star.size(0) - k_star = k_star[l_barrel, t_barrel] - ###################################################################### # Compute the recurrent state qw = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qw) v = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_v) - # k = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_k) - aw = torch.einsum( - "nhtd,ltd->nhlt", - qw, - k_star, - ) / math.sqrt(self.w_qw.size(1)) + aw = torch.einsum("nhtd,ld->nhlt", qw, self.k_star) / math.sqrt( + self.w_qw.size(1) + ) aw = aw.softmax(dim=2) # nhlt - if self.train: - self.acc_attention += aw.sum(dim=(0, 1, 3)) - self.acc_nb += aw.size(0) * aw.size(1) * aw.size(3) - aw = F.dropout(aw, self.attention_dropout, self.training) A = 1 - aw.sum(dim=1) # nlt V = torch.einsum("nhlt,nhtd->nltd", aw, v).contiguous() - # K = torch.einsum("nhlt,nhtd->nltd", aw, k).contiguous() if t0 == 0: V0 = None - # K0 = None else: V0 = self.rec_v[:, :, t0 - 1] - # K0 = self.rec_k[:, :, t0 - 1] self.rec_v[:, :, t0:t1] = pscan_shape(A, V, V0) - # self.rec_k[:, :, t0:t1] = pscan_shape(A, K, K0) ###################################################################### # compute the readout @@ -285,7 +276,6 @@ class DumbRec(nn.Module): ar = torch.einsum( "nhtd,ld->nhlt", qr, - # self.rec_k[:, :, t0:t1], self.k_star, ) / math.sqrt(self.w_qr.size(1)) @@ -317,6 +307,8 @@ class KVRec(nn.Module): nb_lines, attention_dropout=0.0, len_max=1e5, + logger=print, + args=None, ): super().__init__() @@ -339,9 +331,9 @@ class KVRec(nn.Module): 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_qw.device) + # warnings.warn("l2 regularization", RuntimeWarning) + # return (self.acc_attention / self.acc_nb).pow(2).sum() + return torch.tensor([0], device=self.w_qw.device) # warnings.warn("side regularization", RuntimeWarning) # return ( # (0.5 / self.nb_lines - self.acc_attention / self.acc_nb).clamp(min=0).sum() @@ -365,12 +357,12 @@ class KVRec(nn.Module): k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1) - warnings.warn("rotating key barrel", RuntimeWarning) + # warnings.warn("rotating key barrel", RuntimeWarning) k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1) t_barrel = torch.arange(t0, t1, device=k_star.device) t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0) l_barrel = ( - torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel + torch.arange(k_star.size(0), device=k_star.device)[:, None] # + t_barrel ) % k_star.size(0) k_star = k_star[l_barrel, t_barrel] @@ -457,77 +449,6 @@ def moving_window(x, dim, win_dim, win_size): ############################## -# This is one order of magnitude more complicated than I expected, not -# elegant, slow, hopefully not buggy - - -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) - - # Set a origin source time (starting time of the chunck to copy - # here) We set it as the current time minus a multiple of CL to be - # consistent with the "rolling" caterpillar - t = torch.arange(fb_start.size(2), device=fb_start.device)[None, None, :] - src_time = fb_start * ( - t - - CL - * ( - 1 - + ( - torch.rand(fb_start.size(), device=fb_start.device) * (t // CL - 1) - ).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] - fbh_V = fbh[:, :, :, None] - t = fbt_V.clamp(min=0) - n = torch.arange(V.size(0), device=V.device)[:, None, None, None] - d = torch.arange(V.size(3), device=V.device)[None, None, None, :] - 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] - fbh_K = fbh[:, :, :, None] - t = fbt_K.clamp(min=0) - n = torch.arange(K.size(0), device=K.device)[:, None, None, None] - d = torch.arange(K.size(3), device=K.device)[None, None, None, :] - q = K[:, :, t0:t1][n, fbh_K, t, d] - rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0) - - -###################################################################### - class Caterpillar(nn.Module): def __init__( @@ -540,44 +461,49 @@ class Caterpillar(nn.Module): caterpillar_height, attention_dropout=0.0, len_max=1e5, + logger=print, + args=None, ): super().__init__() warnings.warn("Caterpillar", RuntimeWarning) - def randw(*d): - return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1])) + def randw(*d, factor=1): + return nn.Parameter(torch.randn(*d) * factor / math.sqrt(d[-1])) self.caterpillar_length = caterpillar_length self.caterpillar_height = caterpillar_height self.attention_dropout = attention_dropout - warnings.warn("flash back", RuntimeWarning) - self.proba_flashback = 0.1 + ###################################################################### 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.b_G = nn.Parameter(torch.full((nb_heads, caterpillar_height), 0.0)) 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) + 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 - 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 @@ -590,138 +516,97 @@ 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) + V = torch.einsum("ntc,hdc->nhtd", X, self.w_V) + K = torch.einsum("ntc,hdc->nhtd", X, self.w_K) + ###################################################################### # 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 + # 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() - # That bas a bad idea - # G = F.dropout(G, self.attention_dropout, self.training) + # Clip the gating to avoid values greater than 1 when several + # heads hit the same row - V = torch.einsum("ntc,hdc->nhtd", X, self.w_V) - K = torch.einsum("ntc,hdc->nhtd", X, self.w_K) + G = G / G.sum(1, keepdim=True).clamp(min=1) - # 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) - - 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)) - - 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) - - if self.training and self.proba_flashback: - # This piece of code makes the assumption that there is - # nothing informative before t0, otherwise we'd have to - # implement a cache for V and K too. This should not be - # too much of a problem since this is used only during - # train, where full sequence are available - - # insert_flash_back( - # self.rec_V, - # V, - # self.rec_K, - # K, - # t0, - # t1, - # CL, - # proba=self.proba_flashback / CL, - # ) - - n = torch.arange(N, device=X.device)[:, None, None, None] - t = torch.arange(t0, t1, device=X.device)[None, None, :, None] - dv = torch.arange(DV, device=X.device)[None, None, None, :] - dk = torch.arange(DK, device=X.device)[None, None, None, :] - - u = ( - torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL - ) * CL - - src_time = t - u - t0 - src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device) - - mask_V = ( - torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback - ).long() - self.rec_V[:, :, t0:t1] = ( - mask_V * V[n, src_head, src_time, dv] - + (1 - mask_V) * self.rec_V[:, :, t0:t1] - ) + ###################################################################### - mask_K = ( - torch.rand(N, CH, t1 - t0, DK, device=X.device) <= self.proba_flashback - ).long() - self.rec_K[:, :, t0:t1] = ( - mask_K * K[n, src_head, src_time, dk] - + (1 - mask_K) * self.rec_K[:, :, t0:t1] - ) + A = 1 - G.sum(dim=1) + + 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] + + # 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. + + 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) + + 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 + # Q = blanket(Q) + + # 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) @@ -741,8 +626,6 @@ class Caterpillar(nn.Module): 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) @@ -759,7 +642,10 @@ class QKVAttention(nn.Module): dim_v, nb_heads=1, causal=False, + horizon=None, attention_dropout=0.0, + logger=print, + args=None, ): super().__init__() @@ -767,6 +653,7 @@ class QKVAttention(nn.Module): return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1])) self.causal = causal + self.horizon = horizon self.attention_dropout = attention_dropout self.record_attention = False @@ -810,6 +697,17 @@ class QKVAttention(nn.Module): torch.arange(x_q.size(1), device=q.device)[None, None, :, None] < torch.arange(x_q.size(1), device=q.device)[None, None, None, :] ) + + if self.horizon is not None: + self.cache_attzero = torch.logical_or( + self.cache_attzero, + torch.arange(x_q.size(1), device=q.device)[None, None, :, None] + >= torch.arange(x_q.size(1), device=q.device)[ + None, None, None, : + ] + + self.horizon, + ) + a = a.masked_fill( self.cache_attzero[ :, :, bs.first : bs.first + bs.nb, : bs.first + bs.nb @@ -847,17 +745,26 @@ class MyGPT(nn.Module): nb_blocks, nb_lines=None, caterpillar_height=None, - dim_rec_v=-1, causal=False, dropout=0.0, len_max=1e5, - attention_layer="kvrec", + attention_layer="caterpillar", + logger=print, + args=None, ): super().__init__() - assert attention_layer in {"mha", "dumbrec", "kvrec", "caterpillar"} + self.vocabulary_size = vocabulary_size + + assert attention_layer in { + "mha", + "dumbrec", + "kvrec", + "caterpillar", + "attcat", + }, f"Unknown attention operator {attention_layer}." - if attention_layer == "caterpillar": + if attention_layer == "caterpillar" or attention_layer == "attcat": assert nb_lines % caterpillar_height == 0 self.caterpillar_length = nb_lines // caterpillar_height self.caterpillar_height = caterpillar_height @@ -876,51 +783,99 @@ class MyGPT(nn.Module): def attlayer(): if attention_layer == "mha": - return QKVAttention( - dim_model=dim_model, - dim_qk=dim_keys, - dim_v=dim_model // nb_heads, - nb_heads=nb_heads, - causal=causal, - attention_dropout=dropout, + return WithResidual( + CacheWrapper(nn.LayerNorm((dim_model,))), + QKVAttention( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + causal=causal, + attention_dropout=dropout, + logger=logger, + args=args, + ), ) elif attention_layer == "dumbrec": - return DumbRec( - dim_model=dim_model, - dim_qk=dim_keys, - dim_v=dim_rec_v, - nb_heads=nb_heads, - nb_lines=nb_lines, - attention_dropout=dropout, + return WithResidual( + CacheWrapper(nn.LayerNorm((dim_model,))), + DumbRec( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + nb_lines=nb_lines, + attention_dropout=dropout, + logger=logger, + args=args, + ), ) elif attention_layer == "kvrec": - return KVRec( - dim_model=dim_model, - dim_qk=dim_keys, - dim_v=dim_rec_v, - nb_heads=nb_heads, - nb_lines=nb_lines, - attention_dropout=dropout, + return WithResidual( + CacheWrapper(nn.LayerNorm((dim_model,))), + KVRec( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + nb_lines=nb_lines, + attention_dropout=dropout, + logger=logger, + args=args, + ), ) elif attention_layer == "caterpillar": - return Caterpillar( - dim_model=dim_model, - dim_qk=dim_keys, - dim_v=dim_rec_v, - nb_heads=nb_heads, - caterpillar_length=self.caterpillar_length, - caterpillar_height=self.caterpillar_height, - attention_dropout=dropout, + return WithResidual( + CacheWrapper(nn.LayerNorm((dim_model,))), + Caterpillar( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + caterpillar_length=self.caterpillar_length, + caterpillar_height=self.caterpillar_height, + attention_dropout=dropout, + logger=logger, + args=args, + ), + ) + elif attention_layer == "attcat": + return nn.Sequential( + WithResidual( + CacheWrapper(nn.LayerNorm((dim_model,))), + QKVAttention( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + causal=causal, + horizon=self.caterpillar_length, + attention_dropout=dropout, + logger=logger, + args=args, + ), + ), + WithResidual( + CacheWrapper(nn.LayerNorm((dim_model,))), + Caterpillar( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + 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}.") for b in range(nb_blocks): trunk_blocks += [ - WithResidual( - CacheWrapper(nn.LayerNorm((dim_model,))), - attlayer(), - ), + attlayer(), WithResidual( CacheWrapper( nn.LayerNorm((dim_model,)), @@ -1041,7 +996,115 @@ class MyGPT(nn.Module): ###################################################################### if __name__ == "__main__": - print("Basic check.") + import argparse + + import numpy as np + import matplotlib.pyplot as plt + import matplotlib.collections as mc + + args = argparse.Namespace( + gate_dropout_proba=0.0, gate_dropout_sync=True, gate_dropout_replace=False + ) + + device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + dim_model, dim_keys, nb_heads = 512, 64, 1 + dropout = 0.1 + + caterpillar = Caterpillar( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + caterpillar_length=16, + caterpillar_height=32, + attention_dropout=dropout, + args=args, + ).to(device) + + qkv = QKVAttention( + dim_model=dim_model, + dim_qk=dim_keys, + dim_v=dim_model // nb_heads, + nb_heads=nb_heads, + causal=True, + attention_dropout=dropout, + args=args, + ).to(device) + + linear = CacheWrapper(nn.Linear(512, 512)).to(device) + + x = torch.randn(1, 256, dim_model) + + x = x.to(device) + x.requires_grad_() + + ###################################################################### + + fig = plt.figure() + fig.set_figheight(6) + fig.set_figwidth(8) + + ax = fig.add_subplot(1, 1, 1) + + # ax.set_xlim(-1.5, 1.5) + # ax.set_ylim(-1.5, 1.5) + # ax.set(aspect=1) + # ax.spines.right.set_visible(False) + # ax.spines.top.set_visible(False) + + # dt = 0.01 + # t = np.arange(dt, 20.0, dt) + # ax.semilogx(t, np.exp(-t / 5.0)) + # ax.grid() + ax.set_yscale("log") + + ###################################################################### + + for label, model, thickness in [ + ("nn.Linear", linear, 0.2), + ("mygpy.QKVAttention", qkv, 1), + ("mygpt.Caterpillar", caterpillar, 2), + ]: + y = model(BracketedSequence(x, 32, x.size(1) - 32, init_cache=True)).x + + for n, p in [("input", x)] + list(model.named_parameters()): + print(f"Processing {model}.{n}") + data = [] + for t in range(y.size(1)): + sg = 0 + for d in torch.randperm(y.size(2))[:8]: + sg += torch.autograd.grad(y[0, t, d], p, retain_graph=True)[0] + assert not sg.isinf().any() + assert not sg.isnan().any() + data.append([t, sg.sum().item()]) + + data = torch.tensor(data) + # cx, cy = data[:, 0], data[:, 1] + cy = data[:, 1].sort().values + cx = torch.linspace(0, 1, cy.size(0)) + ax.plot( + cx, cy, label=label + "." + n, linewidth=thickness + ) # , color='gray', label='Input') + + # ax.legend(frameon=False, loc="top right") + + # Put a legend to the right of the current axis + box = ax.get_position() + ax.set_position([box.x0, box.y0, box.width * 0.8, box.height]) + ax.legend(loc="center left", bbox_to_anchor=(1, 0.5)) + + filename = "plot.pdf" + print(f"saving {filename}") + fig.savefig(filename, bbox_inches="tight") + + # if args.window and hasattr(plt.get_current_fig_manager(), 'window'): + # plt.get_current_fig_manager().window.setGeometry(2, 2, 1024, 768) + # plt.show() + + exit(0) + + ###################################################################### m = Caterpillar( dim_model=4, @@ -1063,8 +1126,6 @@ if __name__ == "__main__": print((y1 - torch.cat([y3a, y3b], dim=1)).abs().max()) exit(0) - device = torch.device("cuda" if torch.cuda.is_available() else "cpu") - vocabulary_size = 128 x = torch.randint(vocabulary_size, (6, 1024))