X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=12b3631596b702e5f7d4f5e91d797ac4d6487298;hb=HEAD;hp=040845ede9e4307f6e76c8a4a7faadd5bacd9974;hpb=3d7db5b3c1304fdbd599c2a001b5c31df4df2599;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index 040845e..12b3631 100755 --- a/mygpt.py +++ b/mygpt.py @@ -21,6 +21,8 @@ from torch.nn import functional as F import ffutils +# from blanket import blanket + # import memload ###################################################################### @@ -84,6 +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__() @@ -216,19 +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 @@ -236,61 +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 @@ -300,7 +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)) @@ -356,9 +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() @@ -382,12 +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] @@ -500,17 +475,13 @@ class Caterpillar(nn.Module): self.caterpillar_height = caterpillar_height self.attention_dropout = attention_dropout - 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, factor=1.0) + self.w_G = randw(nb_heads, caterpillar_height, dim_model) 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, factor=1) + 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) @@ -583,83 +554,32 @@ class Caterpillar(nn.Module): # 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) - - H = (1 - G).log().sum(1, keepdim=True).exp() + G = G / G.sum(1, keepdim=True).clamp(min=1) ###################################################################### - def recurrence(G, V, K): - # We prepare the arguments for the parallel scan - - A = H - - gated_V = torch.einsum("nhrt,nhtd->nrtd", H * G / (1 - G), V) - gated_K = torch.einsum("nhrt,nhtd->nrtd", H * G / (1 - 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) - - return next_V, next_K + A = 1 - G.sum(dim=1) - ################################################################# + 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) + # 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. - 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 - ) - - # The coefficient to keep are the complementary - mask = 1 - kill - - masked_next_V, masked_next_K = recurrence(G * mask, V, K) - - if self.gate_dropout_replace: - next_V = next_V.detach() - next_K = next_K.detach() + A = A.unflatten(2, (-1, L)) + gated_V = gated_V.unflatten(2, (-1, L)) + gated_K = gated_K.unflatten(2, (-1, L)) - warnings.warn("the rescaling is probably a bad idea", RuntimeWarning) - - 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 - ) + 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 @@ -669,6 +589,8 @@ class Caterpillar(nn.Module): Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q) + # 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 @@ -704,8 +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) @@ -722,6 +642,7 @@ class QKVAttention(nn.Module): dim_v, nb_heads=1, causal=False, + horizon=None, attention_dropout=0.0, logger=print, args=None, @@ -732,6 +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 @@ -775,6 +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 @@ -821,14 +754,17 @@ class MyGPT(nn.Module): ): super().__init__() + 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 @@ -847,59 +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, - logger=logger, - args=args, + 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_model // nb_heads, - nb_heads=nb_heads, - nb_lines=nb_lines, - attention_dropout=dropout, - logger=logger, - args=args, + 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_model // nb_heads, - nb_heads=nb_heads, - nb_lines=nb_lines, - attention_dropout=dropout, - logger=logger, - args=args, + 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_model // nb_heads, - nb_heads=nb_heads, - caterpillar_length=self.caterpillar_length, - caterpillar_height=self.caterpillar_height, - attention_dropout=dropout, - logger=logger, - args=args, + 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,)), @@ -1081,31 +1057,35 @@ if __name__ == "__main__": # t = np.arange(dt, 20.0, dt) # ax.semilogx(t, np.exp(-t / 5.0)) # ax.grid() + ax.set_yscale("log") ###################################################################### - for label, model in [ - # ("nn.Linear", linear), - ("mygpy.QKVAttention", qkv), - ("mygpt.Caterpillar", caterpillar), + 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 - data = [] - for t in range(y.size(1)): - for d in torch.randperm(y.size(2))[:8]: - g = torch.autograd.grad(y[0, t, d], x, retain_graph=True)[0] - sg = g.pow(2).sum().item() - # sg = 0 - # for p in model.parameters(): - # g = torch.autograd.grad(y[0, t, d], p, retain_graph=True)[0] - # sg = sg + g.pow(2).sum().item() - data.append([t, sg]) - - data = torch.tensor(data) - ax.scatter( - data[:, 0], data[:, 1], s=1, label=label - ) # , color='gray', label='Input') + 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")