###################################################################### 2024 Jan 07 21:37:48 (from mygpt.py) # 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) ###################################################################### ###################################################################### 2024 Jan 07 21:38:11 (from mygpt.py) # insert_flash_back(self.rec_V,V,self.rec_K,K,t0,t1,CL,proba=self.proba_flashback / CL,) ###################################################################### 2024 Jan 09 14:24:42 (from mygpt.py) # 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] # ) ###################################################################### 2024 Jan 10 08:10:39 (from mygpt.py) # That was a bad idea # G = F.dropout(G, self.attention_dropout, self.training) ###################################################################### 2024 Jan 10 08:46:13 (from mygpt.py) ################################################################# # Flashbacks. This version sucks, about to replace it 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] ) ###################################################################### 2024 Jan 13 13:38:31 (from mygpt.py) g= F.sigmoid(self.b_G) a=1-g print(f"\n\nSANITY {a**T}\n") exit(0) ###################################################################### 2024 Jan 14 13:39:37 (from mygpt.py) epsilon = 0.5 dropout_head = ( (torch.rand(N, H, 1, t1 - t0, device=G.device).sort(dim=3).indices == 0) .expand_as(G) .float() ) 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() ) ###################################################################### 2024 Jan 18 07:39:29 (from mygpt.py) class Calibrator: def __init__(self, w=None, b=None): self.w = w self.b = b self.s, self.s_sq, self.n = 0, 0, 0 self.mean, self.std = 0, 0 def update(self, X): X = X.detach() self.s += X.sum(dim=0) self.s_sq += X.pow(2).sum(dim=0) self.n += X.size(0) def moments(self): mean = self.s / self.n std = (self.s_sq / self.n - mean * mean).sqrt() return mean, std def normalize(self): mean, std = self.moments() if self.b is not None: self.b.sub_(mean) if self.w is not None: self.w.div_(std) result = mean - self.mean, std - self.std self.mean, self.std = mean, std self.s, self.s_sq, self.n = 0, 0, 0 return result ###################################################################### 2024 Jan 18 07:39:34 (from mygpt.py) # self.calibrator_G = Calibrator() # self.calibrator_rec_V = Calibrator() # self.calibrator_rec_K = Calibrator() ###################################################################### 2024 Jan 18 07:39:37 (from mygpt.py) # self.calibrator_G.update(G.reshape(-1, G.size(-1))) ###################################################################### 2024 Jan 18 07:39:42 (from mygpt.py) # self.calibrator_rec_V.update( # next_V.permute(0, 1, 3, 2).reshape(-1, next_V.size(2)) # ) # self.calibrator_rec_K.update( # next_K.permute(0, 1, 3, 2).reshape(-1, next_K.size(2)) # ) ###################################################################### 2024 Jan 18 07:47:12 (from mygpt.py) ###################################################################### # Roll the gating indexes # warnings.warn("rotating barrel", RuntimeWarning) # r_barrel = torch.arange(R, device=G.device)[None, None, :, None] # t_barrel = torch.arange(t1 - t0, device=G.device)[None, None, None, :] # r_barrel = (r_barrel + (t_barrel + t0) // L) % R # G = G.gather(dim=2, index=r_barrel.expand_as(G)) ###################################################################### 2024 Jan 18 07:47:25 (from mygpt.py) # warnings.warn("harmonic recurrence", RuntimeWarning) # har = torch.arange(t0, t1, device = G.device).float() + 1 # A = har / (har + 1) # G = G / har ###################################################################### 2024 Jan 18 08:46:18 (from mygpt.py) # warnings.warn("softmax gating", RuntimeWarning) # G = ( # torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None] # ).softmax(dim=2) ###################################################################### 2024 Jan 21 16:55:24 (from main.py) with open("test.dat", "a") as f: for m filter(lambda m: isinstance(m,mygpt.Catenn.Linear),model.modules()): for p in m.parameters() ] for m in model.modules(): if isinstance(m, mygpt.Caterpillar): ###################################################################### 2024 Feb 13 22:53:52 (from mygpt.py) ###################################################################### # 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] ###################################################################### 2024 Feb 15 23:10:50 (from main.py) def add_memex_v4(batches, memex_proba, marker_token): for input in batches: if torch.rand(1).item() < memex_proba: t = ( torch.arange(2 * input.size(1), device=input.device)[None, :] .expand(input.size(0), -1) .clone() ) u = torch.rand(t.size(), device=t.device) u[:, : input.size(1)] = 1.0 memex_v3_proba_fragment = 1 / 20 u = (u < memex_v3_proba_fragment).long() v = u * torch.randint(input.size(1), u.size()) u[:, input.size(1) + 1 :] = v[:, input.size(1) + 1 :] - u[ :, : input.size(1) - 1 ] * input.size(1) u = u.cumsum().clamp(min=0) u0 = torch.randint(input.size(1), (input.size(0), 1), device=input.device) caterpillar_length = args.nb_lines // args.caterpillar_height u1 = ( u0 + torch.randint( caterpillar_length, (input.size(0), 1), device=input.device ) + 1 ) m0 = (t < u0).long() m1 = (t >= u1).long() * (t < u1 + input.size(1)).long() t = t * m0 + ((-1) * (1 - m0) * (1 - m1)) + (t - u1) * m1 m = (t < 0).long() n = torch.arange(input.size(0), device=input.device)[:, None].expand( -1, t.size(1) ) new_input = input[n, t.clamp(min=0)] new_input = (1 - m) * new_input + m * (marker_token) yield new_input yield input ###################################################################### 2024 Feb 16 17:07:48 (from main.py) # ||gn + lambda * gm|| = max(||gn||,||gm||) # ||gn||^2 + lambda + lambda^2||gm||^2 = max(||gn||^2,||gm||^2) # A = ||gm||^2 B = C = ||gn||^2 - max(||gn||^2, ||gm||^2) ###################################################################### 2024 Feb 16 17:07:51 (from main.py) # A,B,C = gmgm, gngm, gngn - max(gngn,gmgm) # Delta = B*B - 4*A*C # if(delta >= 0): # l = ( -B - sqrt(Delta))/(2*A) # ||gn||+l*rho*||gm|| = max(||gn||,rho*||gm||)