X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=b885e218be6704cd86afed18966e5609e9873369;hb=42831bd654d030b71bca88578d041279018f836c;hp=e7362b749210dcce6c6b92934bd34744d95b770d;hpb=6833683bd343fd687d093d6c47cca8f1909e8b03;p=mygptrnn.git diff --git a/mygpt.py b/mygpt.py index e7362b7..b885e21 100755 --- a/mygpt.py +++ b/mygpt.py @@ -481,8 +481,8 @@ class Caterpillar(nn.Module): self.caterpillar_height = caterpillar_height self.attention_dropout = attention_dropout - warnings.warn("flash back", RuntimeWarning) - self.proba_flashback = 1e-2 + self.proba_flashback = 0.0 + self.proba_gate_dropout = 0.0 self.w_G = randw(nb_heads, caterpillar_height, dim_model) self.b_G = nn.Parameter( @@ -551,7 +551,11 @@ class Caterpillar(nn.Module): torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None] ).sigmoid() - # That bas a bad idea + if self.training and self.proba_gate_dropout > 0.0: + warnings.warn("gate droupout", RuntimeWarning) + epsilon = 0.5 + + # That was a bad idea # G = F.dropout(G, self.attention_dropout, self.training) V = torch.einsum("ntc,hdc->nhtd", X, self.w_V) @@ -559,6 +563,10 @@ class Caterpillar(nn.Module): # We prepare the arguments for the parallel scan + # Clip the gating + warnings.warn("gating clipping", RuntimeWarning) + 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) @@ -585,6 +593,7 @@ class Caterpillar(nn.Module): 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