diff options
Diffstat (limited to 'sdil/local_baselines.py')
| -rw-r--r-- | sdil/local_baselines.py | 241 |
1 files changed, 161 insertions, 80 deletions
diff --git a/sdil/local_baselines.py b/sdil/local_baselines.py index a6a11a3..23b1b5c 100644 --- a/sdil/local_baselines.py +++ b/sdil/local_baselines.py @@ -1,6 +1,7 @@ """ -Other biologically-motivated / non-backprop local learning baselines, sharing -SDILNet's architecture and initialisation for a fair comparison: +Other biologically-motivated / non-backprop local learning baselines. FA and +PEPITA reuse SDILNet's tensor layout; FF and EP retain their native model +classes because their state spaces and objectives are fundamentally different: - FANet : Feedback Alignment (Lillicrap 2016) -- backprop through FIXED random feedback matrices instead of W^T (sequential, layer-wise). DFA is the @@ -12,7 +13,10 @@ SDILNet's architecture and initialisation for a fair comparison: it on negative (wrong label) data; inference picks max total goodness. All train with no global backward graph. Autograd, where used (FF's per-layer -local loss), never crosses layer boundaries -> the update stays local. +local loss), never crosses layer boundaries -> the update stays local. The +implementations follow the authors' published algorithms and reference code; +their deliberately different architectures are exposed rather than hidden +behind an allegedly apples-to-apples SDILNet wrapper. """ import math import torch @@ -25,9 +29,11 @@ from .core import SDILNet, ACTS # Feedback Alignment (sequential, layer-wise random feedback) # -------------------------------------------------------------------------- class FANet(SDILNet): - def __init__(self, *args, b_scale=1.0, **kw): + def __init__(self, *args, b_scale=1.0, feedback_seed=None, **kw): + model_seed = kw.get("seed", 0) super().__init__(*args, **kw) - g = torch.Generator(device="cpu").manual_seed(4242) + g = torch.Generator(device="cpu").manual_seed( + model_seed + 4242 if feedback_seed is None else feedback_seed) # fixed random feedback B[l] with the same shape as W[l], for l=1..L-1 self.B = [None] for l in range(1, self.L): @@ -39,16 +45,22 @@ class FANet(SDILNet): with torch.no_grad(): fwd = self.forward(x) h, u = fwd["h"], fwd["u"] - B = x.shape[0] e = torch.softmax(h[-1], 1) - yoh # grad wrt logits loss = F.cross_entropy(h[-1], y).item() # output layer (exact local) - delta = e # grad wrt u[L-1] - self._upd(self.L - 1, delta, h[self.L - 1], eta, momentum) - # hidden layers: propagate with fixed random B + self._upd(self.L - 1, e, h[self.L - 1], eta, momentum) + + # dh is the error wrt the hidden *state*. On a residual block, + # h'=h+alpha*phi(Wh), the identity branch transports dh exactly and + # only the nonlinear branch uses a fixed random feedback matrix. + dh = e @ self.B[self.L - 1] for l in range(self.L - 2, -1, -1): - delta = (delta @ self.B[l + 1]) * self.act_prime(u[l]) # grad wrt u[l] - self._upd(l, delta, h[l], eta, momentum) + branch_scale = self.res_alpha if self.residual and l >= 1 else 1.0 + du = branch_scale * dh * self.act_prime(u[l]) + self._upd(l, du, h[l], eta, momentum) + if l > 0: + feedback = du @ self.B[l] + dh = dh + feedback if self.residual and l >= 1 else feedback return loss def _upd(self, l, delta, pre, eta, momentum): @@ -65,39 +77,85 @@ class FANet(SDILNet): # PEPITA (forward-only, error-modulated second pass) # -------------------------------------------------------------------------- class PEPITANet(SDILNet): - def __init__(self, *args, f_scale=0.1, **kw): + """Original two-forward-pass PEPITA/ERIN rule. + + Dellaferrera & Kreiman initialise both feedforward weights and the input + error projection with He-uniform samples, then multiply the latter by 0.05. + Their fully-connected implementation has ReLU hidden units, a softmax + output, no biases, and optionally shares one dropout mask across the two + presentations. We keep logits internally but form exactly the same + softmax errors. + """ + + def __init__(self, *args, f_scale=0.05, original_init=True, + use_bias=False, keep_prob=1.0, **kw): + model_seed = kw.get("seed", 0) super().__init__(*args, **kw) - g = torch.Generator(device="cpu").manual_seed(7777) + g = torch.Generator(device="cpu").manual_seed(model_seed + 7777) n_in = self.sizes[0] - # projection of output error back onto the input (fixed random) - self.Fproj = (torch.randn(n_in, self.n_classes, generator=g) - * (f_scale / math.sqrt(self.n_classes))).to(self.W[0].device, self.dtype) + if original_init: + for l, w in enumerate(self.W): + limit = math.sqrt(6.0 / self.sizes[l]) + w.copy_((torch.rand(w.shape, generator=g) * 2.0 * limit - limit) + .to(w.device, w.dtype)) + limit = math.sqrt(6.0 / n_in) + self.Fproj = ((torch.rand(n_in, self.n_classes, generator=g) * 2.0 * limit - limit) + * f_scale).to(self.W[0].device, self.dtype) + self.use_bias = use_bias + self.keep_prob = keep_prob + + def _pepita_forward(self, x, masks): + h = [x] + for l in range(self.L): + cur = h[-1] @ self.W[l].t() + if self.use_bias: + cur = cur + self.b[l] + if l < self.L - 1: + cur = self.act(cur) + if masks is not None: + cur = cur * masks[l] + h.append(cur) + return h def pepita_step(self, x, y, yoh, eta, momentum=0.0): with torch.no_grad(): - std = self.forward(x) - hs = std["h"] + masks = None + if self.keep_prob < 1.0: + masks = [(torch.rand(x.shape[0], self.sizes[l + 1], device=x.device) + < self.keep_prob).to(x.dtype) / self.keep_prob + for l in range(self.L - 1)] + hs = self._pepita_forward(x, masks) e = torch.softmax(hs[-1], 1) - yoh loss = F.cross_entropy(hs[-1], y).item() x_mod = x + (e @ self.Fproj.t()) # modulate the input by the error - mod = self.forward(x_mod) - hm = mod["h"] + hm = self._pepita_forward(x_mod, masks) + mod_error = torch.softmax(hm[-1], 1) - yoh B = x.shape[0] for l in range(self.L): - # change in this layer's output between clean and modulated passes - dpost = hs[l + 1] - hm[l + 1] - # PEPITA: first layer uses the CLEAN input as presynaptic; deeper - # layers use the modulated presynaptic. - pre = x if l == 0 else hm[l] - dW = -(dpost.t() @ pre) / B - db = -dpost.mean(0) + pre = hm[l] # modulated presynaptic state + if l == self.L - 1: + update = -(mod_error.t() @ pre) / B + update_b = -mod_error.mean(0) + else: + diff = hs[l + 1] - hm[l + 1] + update = -(diff.t() @ pre) / B + update_b = -diff.mean(0) if momentum: - self.mW[l].mul_(momentum).add_(dW); self.mb[l].mul_(momentum).add_(db) - self.W[l] += eta * self.mW[l]; self.b[l] += eta * self.mb[l] + self.mW[l].mul_(momentum).add_(update) + self.W[l] += eta * self.mW[l] + if self.use_bias: + self.mb[l].mul_(momentum).add_(update_b) + self.b[l] += eta * self.mb[l] else: - self.W[l] += eta * dW; self.b[l] += eta * db + self.W[l] += eta * update + if self.use_bias: + self.b[l] += eta * update_b return loss + def forward(self, x, **_): + h = self._pepita_forward(x, masks=None) + return {"h": h, "u": []} + # -------------------------------------------------------------------------- # Forward-Forward (Hinton 2022) @@ -107,8 +165,9 @@ class FFNet: each layer is trained by a local goodness objective; classification sums goodness across layers over the candidate labels.""" - def __init__(self, sizes, act="tanh", device="cpu", seed=0, threshold=2.0, - n_classes=10, dtype=torch.float32, overlay_val=10.0): + def __init__(self, sizes, act="relu", device="cpu", seed=0, threshold=2.0, + n_classes=10, dtype=torch.float32, overlay_val=None, + score_from_layer=1): self.sizes = list(sizes) self.L = len(sizes) - 1 self.n_classes = n_classes @@ -117,18 +176,19 @@ class FFNet: self.act, _ = ACTS[act] self.thr = threshold self.overlay_val = overlay_val + self.score_from_layer = min(score_from_layer, max(0, self.L - 1)) g = torch.Generator(device="cpu").manual_seed(seed) - self.W, self.b, self.mW, self.mb = [], [], [], [] + self.W, self.b, self.optimizers = [], [], [] for i in range(self.L): - w = (torch.randn(sizes[i + 1], sizes[i], generator=g) / math.sqrt(sizes[i]) - ).to(device, dtype) - w.requires_grad_(True) + # Match torch.nn.Linear initialization used by the public reference. + bound = 1.0 / math.sqrt(sizes[i]) + w = (torch.rand(sizes[i + 1], sizes[i], generator=g) * 2.0 * bound - bound).to( + device, dtype).requires_grad_(True) self.W.append(w) - bb = torch.zeros(sizes[i + 1], device=device, dtype=dtype, requires_grad=True) + bb = (torch.rand(sizes[i + 1], generator=g) * 2.0 * bound - bound).to( + device, dtype).requires_grad_(True) self.b.append(bb) - - def __init_overlay_scale__(self): - pass + self.optimizers.append(torch.optim.Adam([w, bb], lr=0.03)) def _overlay(self, x, labels): """Overlay one-hot label onto the first n_classes input features. @@ -136,7 +196,8 @@ class FFNet: (10 of 784 pixels is otherwise swamped by the shared image).""" xo = x.clone() xo[:, :self.n_classes] = 0.0 - xo[torch.arange(x.shape[0]), labels] = self.overlay_val + value = x.max().detach() if self.overlay_val is None else self.overlay_val + xo[torch.arange(x.shape[0], device=x.device), labels] = value return xo @staticmethod @@ -144,28 +205,40 @@ class FFNet: return h / (h.norm(dim=1, keepdim=True) + 1e-8) def _layer_forward(self, l, h_in): - return self.act(h_in @ self.W[l].t() + self.b[l]) + return self.act(self._norm(h_in) @ self.W[l].t() + self.b[l]) + + def _inputs_to_layer(self, l, x): + h = x + with torch.no_grad(): + for k in range(l): + h = self._layer_forward(k, h) + return h.detach() + + def train_layer(self, l, x, y, eta=0.03, negative_labels=None): + """One minibatch update of one greedy FF layer (local autograd only).""" + if negative_labels is None: + negative_labels = ((y + torch.randint( + 1, self.n_classes, y.shape, device=y.device)) % self.n_classes) + hpos = self._inputs_to_layer(l, self._overlay(x, y)) + hneg = self._inputs_to_layer(l, self._overlay(x, negative_labels)) + hp = self._layer_forward(l, hpos) + hn = self._layer_forward(l, hneg) + gp = hp.pow(2).mean(1) + gn = hn.pow(2).mean(1) + loss = (F.softplus(-gp + self.thr) + F.softplus(gn - self.thr)).mean() + opt = self.optimizers[l] + opt.param_groups[0]["lr"] = eta + opt.zero_grad() + loss.backward() + opt.step() + return loss.item() def train_step(self, x, y, eta): - # positive = correct label; negative = a random wrong label + """Compatibility path; canonical runner trains greedily layer-by-layer.""" neg = (y + torch.randint(1, self.n_classes, y.shape, device=y.device)) % self.n_classes - xpos = self._overlay(x, y) - xneg = self._overlay(x, neg) - hpos, hneg = xpos, xneg total = 0.0 for l in range(self.L): - hp = self._layer_forward(l, hpos.detach()) - hn = self._layer_forward(l, hneg.detach()) - gp = hp.pow(2).mean(1) # goodness (positive) - gn = hn.pow(2).mean(1) # goodness (negative) - # push positive goodness above threshold, negative below - loss = (F.softplus(-(gp - self.thr)) + F.softplus(gn - self.thr)).mean() - gW, gb = torch.autograd.grad(loss, [self.W[l], self.b[l]]) - with torch.no_grad(): - self.W[l] -= eta * gW - self.b[l] -= eta * gb - total += loss.item() - hpos, hneg = self._norm(hp).detach(), self._norm(hn).detach() + total += self.train_layer(l, x, y, eta, neg) return total / self.L @torch.no_grad() @@ -176,9 +249,8 @@ class FFNet: good = torch.zeros(x.shape[0], device=x.device) for l in range(self.L): h = self._layer_forward(l, h) - if l > 0: # skip first layer for scoring (Hinton) + if l >= self.score_from_layer: # paper excludes first layer good = good + h.pow(2).mean(1) - h = self._norm(h) scores[:, c] = good return scores.argmax(1) @@ -201,15 +273,20 @@ class EPNet: Updates are local (products of adjacent-layer rho's) with no backprop.""" def __init__(self, sizes, device="cpu", seed=0, beta=0.5, dt=0.5, - T_free=20, T_nudge=8, dtype=torch.float32): + T_free=20, T_nudge=4, dtype=torch.float32, + random_beta_sign=True): self.sizes = list(sizes) self.L = len(sizes) - 1 # number of weight layers self.device = device self.dtype = dtype self.beta, self.dt, self.T_free, self.T_nudge = beta, dt, T_free, T_nudge + self.random_beta_sign = random_beta_sign g = torch.Generator(device="cpu").manual_seed(seed) - self.W = [(torch.randn(sizes[i + 1], sizes[i], generator=g) / math.sqrt(sizes[i]) - ).to(device, dtype) for i in range(self.L)] + self.W = [] + for i in range(self.L): + limit = math.sqrt(6.0 / (sizes[i] + sizes[i + 1])) + self.W.append((torch.rand(sizes[i + 1], sizes[i], generator=g) + * 2.0 * limit - limit).to(device, dtype)) self.b = [torch.zeros(sizes[i + 1], device=device, dtype=dtype) for i in range(self.L)] @staticmethod @@ -218,28 +295,28 @@ class EPNet: @staticmethod def rhop(s): - return ((s > 0) & (s < 1)).to(s.dtype) + # Theano's clip derivative used by the authors is active at the bounds. + return ((s >= 0) & (s <= 1)).to(s.dtype) def _settle(self, x, y=None, beta=0.0, s=None, T=20): rx = self.rho(x) if s is None: - # init in the active region (rhop=0 at the clamp boundaries would - # otherwise freeze the dynamics); a feedforward warm start settles fast. - s = [] - below = rx - for i in range(self.L): - below = (below @ self.W[i].t() + self.b[i]).clamp(0, 1) - s.append(below) + # The reference implementation starts persistent particles at zero. + s = [torch.zeros(x.shape[0], n, device=x.device, dtype=x.dtype) + for n in self.sizes[1:]] for _ in range(T): new = [] for k in range(self.L): below = rx if k == 0 else self.rho(s[k - 1]) - pre = below @ self.W[k].t() + self.b[k] + # -dE/ds for E = ||rho(s)||^2/2 - b*rho(s) + # - rho(s_below) W^T rho(s). + drive = -self.rho(s[k]) + below @ self.W[k].t() + self.b[k] if k < self.L - 1: # top-down from layer above - pre = pre + self.rho(s[k + 1]) @ self.W[k + 1] + drive = drive + self.rho(s[k + 1]) @ self.W[k + 1] if k == self.L - 1 and beta: # nudge output toward target - pre = pre + beta * (y - s[k]) - ds = self.rhop(s[k]) * pre - s[k] + # Original cost is ||s_out-y||^2, hence the factor two. + drive = drive + 2.0 * beta * (y - s[k]) + ds = self.rhop(s[k]) * drive new.append((s[k] + self.dt * ds).clamp(0, 1)) s = new return s @@ -247,15 +324,19 @@ class EPNet: def train_step(self, x, y, yoh, eta): with torch.no_grad(): s0 = self._settle(x, beta=0.0, T=self.T_free) # free phase - sb = self._settle(x, yoh, beta=self.beta, s=[t.clone() for t in s0], T=self.T_nudge) + beta = self.beta + if self.random_beta_sign and torch.randint(0, 2, ()).item() == 0: + beta = -beta + sb = self._settle(x, yoh, beta=beta, s=[t.clone() for t in s0], T=self.T_nudge) B = x.shape[0] for k in range(self.L): below0 = self.rho(x) if k == 0 else self.rho(s0[k - 1]) belowb = self.rho(x) if k == 0 else self.rho(sb[k - 1]) - dW = (self.rho(sb[k]).t() @ belowb - self.rho(s0[k]).t() @ below0) / (self.beta * B) - db = (self.rho(sb[k]) - self.rho(s0[k])).mean(0) / self.beta - self.W[k] += eta * dW - self.b[k] += eta * db + dW = (self.rho(sb[k]).t() @ belowb - self.rho(s0[k]).t() @ below0) / (beta * B) + db = (self.rho(sb[k]) - self.rho(s0[k])).mean(0) / beta + layer_eta = eta[k] if isinstance(eta, (list, tuple)) else eta + self.W[k] += layer_eta * dW + self.b[k] += layer_eta * db # free-phase output as prediction proxy for loss logging return F.mse_loss(s0[-1], yoh).item() |