X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=185df3870c1e0c8d35daf9f3e3257340d50fa1a7;hb=57332f677ef5ee535707c1b83a541aa0e79508e6;hp=95e552720e06031d0f0b95c4c3fc6beed0b4eae7;hpb=75e1ddcb8de30a4a7be16c80c4f258da662837a6;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index 95e5527..185df38 100755 --- a/mygpt.py +++ b/mygpt.py @@ -476,14 +476,15 @@ class Caterpillar(nn.Module): warnings.warn("Caterpillar", RuntimeWarning) - def randw(*d): - return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1])) + def randw(*d, amplitude=None): + if amplitude is None: + amplitude = 1 / math.sqrt(d[-1]) + return nn.Parameter(amplitude * torch.randn(*d)) self.caterpillar_length = caterpillar_length self.caterpillar_height = caterpillar_height self.attention_dropout = attention_dropout - self.proba_flashback = 0.0 self.proba_gate_dropout = 0.0 self.w_G = randw(nb_heads, caterpillar_height, dim_model) @@ -498,8 +499,16 @@ class Caterpillar(nn.Module): 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 @@ -540,47 +549,80 @@ class Caterpillar(nn.Module): 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 + # stack. There are CH 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() + ###################################################################### + # The "flashbacks" + if self.training and self.proba_gate_dropout > 0.0: - warnings.warn("gate droupout", RuntimeWarning) + # This is a better implementation of "flashbacks". + + # G is NxHxExT where e is the caterpillar's row. + + warnings.warn("gate dropout", RuntimeWarning) epsilon = 0.5 - # That was a bad idea - # G = F.dropout(G, self.attention_dropout, self.training) + dropout_head = ( + (torch.rand(N, H, 1, t1 - t0, device=G.device).sort(dim=3).indices == 0) + .expand_as(G) + .float() + ) - V = torch.einsum("ntc,hdc->nhtd", X, self.w_V) - K = torch.einsum("ntc,hdc->nhtd", X, self.w_K) + dropout_tail = dropout_head.cumsum(dim=3) - dropout_head + + dropout_active = ( + torch.rand(N, 1, 1, 1, device=G.device) < self.proba_gate_dropout + ).long() + + dropout_head *= dropout_active + dropout_tail *= dropout_active + + G = ( + G + # + dropout_head * (1 - epsilon - G.detach()) + - dropout_tail * G.detach() + ) + + ###################################################################### # We prepare the arguments for the parallel scan - # Clip the gating - warnings.warn("gating clipping", RuntimeWarning) + # 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) A = 1 - G.sum(1) - gated_V = torch.einsum("nhet,nhtd->netd", G, V) - gated_K = torch.einsum("nhet,nhtd->netd", G, 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 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. + ################################################################# + # Associative scan + + # Here there is a trick: Since the stack at position t is + # computed by updating that at position t-CL, the parallel + # scan operates with a period of CL. To do so we split the + # sequence indexing in two axes, the second of size CL, and + # run the parallel scan using the first as the sequence index. A = A.unflatten(2, (-1, CL)) gated_V = gated_V.unflatten(2, (-1, CL)) @@ -589,45 +631,9 @@ 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) - if self.training and self.proba_flashback > 0.0: - warnings.warn("flash back", RuntimeWarning) - # 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 - - 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 = ( - torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback - ).long() - - self.rec_V[:, :, t0:t1] = ( - mask * V[n, src_head, src_time, dv] - + (1 - mask) * self.rec_V[:, :, t0:t1] - ) - - self.rec_K[:, :, t0:t1] = ( - mask * K[n, src_head, src_time, dk] - + (1 - mask) * self.rec_K[:, :, t0:t1] - ) - ###################################################################### # compute the readout