X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=c8330128e115efc1c1f3773ada90e6b929559ce7;hb=7b2d37d9c7ffb10f9bd81ef6356ef7083614a380;hp=760a3c60e086f094d4e86672e040877f94c6a6d2;hpb=60cd4f63c55c77f5097ebac146c212763ba11925;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index 760a3c6..c833012 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 ###################################################################### @@ -493,23 +495,17 @@ class Caterpillar(nn.Module): warnings.warn("Caterpillar", RuntimeWarning) - def randw(*d, amplitude=None): - if amplitude is None: - amplitude = 1 / math.sqrt(d[-1]) - return nn.Parameter(amplitude * torch.randn(*d)) + 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 - 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)) + 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) @@ -527,14 +523,14 @@ class Caterpillar(nn.Module): 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 @@ -589,74 +585,28 @@ class Caterpillar(nn.Module): ###################################################################### - def recurrence(G, V, K): - # We prepare the arguments for the parallel scan - - A = 1 - G.sum(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) - next_K = pscan_dim(A, gated_K, init_rec_K, dim=2) - - next_V = next_V.flatten(2, 3) - next_K = next_K.flatten(2, 3) + A = 1 - G.sum(dim=1) - return next_V, next_K + 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 - next_V, next_K = recurrence(G, V, K) + init_rec_V = self.rec_V[:, :, t0 - L : t0] + init_rec_K = self.rec_K[:, :, t0 - L : t0] - if self.training and self.gate_dropout_proba > 0.0: - # G is NxHxRxT where r is the caterpillar's row. + # 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. - warnings.warn("gate dropout", RuntimeWarning) + A = A.unflatten(2, (-1, L)) + gated_V = gated_V.unflatten(2, (-1, L)) + gated_K = gated_K.unflatten(2, (-1, L)) - 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) - - 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 - ) + 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 @@ -666,6 +616,8 @@ 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 @@ -701,8 +653,6 @@ 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) @@ -719,6 +669,7 @@ class QKVAttention(nn.Module): dim_v, nb_heads=1, causal=False, + horizon=None, attention_dropout=0.0, logger=print, args=None, @@ -729,6 +680,7 @@ 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 @@ -772,6 +724,17 @@ 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 @@ -812,7 +775,7 @@ class MyGPT(nn.Module): causal=False, dropout=0.0, len_max=1e5, - attention_layer="kvrec", + attention_layer="caterpillar", logger=print, args=None, ): @@ -823,9 +786,10 @@ class MyGPT(nn.Module): "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 @@ -844,59 +808,99 @@ 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,)), @@ -1017,7 +1021,115 @@ 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, @@ -1039,8 +1151,6 @@ 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))