File size: 9,132 Bytes
c8293a4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
"""Complete samplers and guidance extensions around the lecture code modules."""
import math
import numpy as np
import torch
from torch import nn
import torch.nn.functional as F
from lecture_core import (K, MASK, draw, token_ce, udlm_rates,
                          rate_step, geometric_cfg, pareto_filter)
from common import objectives


@torch.no_grad()
def uniform_sample(model, batch, length, steps=100, epsilon=.02):
    """Adaptive reverse Euler on the same truncated interval as udlm_loss.

    Uniform initialization at t=1-epsilon and stopping at t=epsilon
    are endpoint approximations. The returned sequence retains residual
    corruption; UDLM output probabilities are not treated as a clean posterior.
    """
    z = torch.randint(K, (batch, length))
    time = 1. - epsilon
    count = 0
    while time > epsilon + 1e-8:
        t = torch.full((batch,), time)
        rate = udlm_rates(model(z, t).softmax(-1), z, t)
        max_exit = float(rate.sum(-1).max())
        h = min(1. / steps, time - epsilon, .5 / max(max_exit, 1e-8))
        z = rate_step(z, rate, h)
        time -= h
        count += 1
        if count > 100000:
            raise RuntimeError('Adaptive sampler failed to advance.')
    return z


@torch.no_grad()
def block_sample(model, batch, length, size=2, steps=20):
    """Generate one block completely before appending the next one."""
    prefix = torch.empty((batch, 0), dtype=torch.long)
    for start in range(0, length, size):
        width = min(size, length - start)
        block = torch.full((batch, width), MASK)
        grid = torch.linspace(1., 0., steps + 1)
        for t, s in zip(grid[:-1], grid[1:]):
            context = torch.cat((prefix, block), 1)
            p = model(context)[:, start:].softmax(-1)
            candidate = draw(p)
            reveal = (block == MASK) & (torch.rand(block.shape) < (t-s)/t)
            block = torch.where(reveal, candidate, block)
        prefix = torch.cat((prefix, block), 1)
    return prefix


def conditional_loss(model, clean, labels, drop=.15):
    t = torch.rand(len(clean)).clamp_min(1e-4)
    mask = torch.rand(clean.shape) < t[:, None]
    context = clean.masked_fill(mask, MASK)
    condition = labels.clone()
    condition[torch.rand(len(clean)) < drop] = 2
    ce = token_ce(model(context, label=condition), clean)
    return (ce * mask / t[:, None]).sum(1).mean()


@torch.no_grad()
def cfg_sample(model, batch, length, strength=2., label=1, steps=20):
    z = torch.full((batch, length), MASK)
    grid = torch.linspace(1., 0., steps + 1)
    labels = torch.full((batch,), label, dtype=torch.long)
    for t, s in zip(grid[:-1], grid[1:]):
        uncond = model(z).softmax(-1)
        cond = model(z, label=labels).softmax(-1)
        probability = geometric_cfg(uncond, cond, strength)
        candidate = draw(probability)
        reveal = (z == MASK) & (torch.rand(z.shape) < (t-s)/t)
        z = torch.where(reveal, candidate, z)
    return z


class NoisyClassifier(nn.Module):
    """Predict high-GC class from a masked sequence and its noise level."""
    def __init__(self, length):
        super().__init__()
        self.net = nn.Sequential(nn.Linear(length * 5 + 1, 64),
                                 nn.SiLU(), nn.Linear(64, 1))

    def forward(self, z, t):
        features = F.one_hot(z, 5).float() if z.ndim == 2 else z
        return self.net(torch.cat((features.flatten(1), t[:, None]), 1)).squeeze(-1)


def fit_classifier(model, data, labels, steps=200, batch=32):
    opt = torch.optim.Adam(model.parameters(), lr=.003)
    losses = []
    for _ in range(steps):
        idx = torch.randint(len(data), (batch,))
        t = torch.rand(batch)
        noisy = data[idx].masked_fill(torch.rand(batch, data.shape[1]) < t[:, None], MASK)
        loss = F.binary_cross_entropy_with_logits(model(noisy, t), labels[idx].float())
        opt.zero_grad(set_to_none=True)
        loss.backward()
        opt.step()
        losses.append(float(loss.detach()))
    model.eval()
    return losses


def log_success(classifier, z, t, label):
    logit = classifier(z, t)
    return F.logsigmoid(logit if label == 1 else -logit)


def guidance_changes(classifier, z, time, label=1, gradient=False):
    """Log h(candidate)-log h(current), exact or first-order one-hot."""
    batch, length = z.shape
    t = torch.full((batch,), float(time))
    if gradient:
        with torch.enable_grad():
            soft = F.one_hot(z, 5).float().requires_grad_(True)
            value = log_success(classifier, soft, t, label)
            grad = torch.autograd.grad(value.sum(), soft)[0]
        current = grad.gather(-1, z[..., None])
        return grad[..., :K] - current
    with torch.no_grad():
        current = log_success(classifier, z, t, label)
        delta = torch.empty((batch, length, K))
        for i in range(length):
            for a in range(K):
                edited = z.clone()
                edited[:, i] = a
                delta[:, i, a] = log_success(classifier, edited, t, label) - current
        return delta


@torch.no_grad()
def classifier_sample(model, classifier, batch, length, strength=1.,
                      label=1, steps=40, gradient=False, epsilon=.005):
    """Rate guidance with adaptive Euler and an explicit endpoint closure."""
    z = torch.full((batch, length), MASK)
    time = 1.
    iterations = 0
    while time > epsilon + 1e-8:
        p = model(z).softmax(-1)
        delta = guidance_changes(classifier, z, time, label, gradient)
        rates = p / time * (strength * delta).clamp(-20, 20).exp()
        rates *= (z == MASK)[..., None]
        exit_rate = rates.sum(-1)
        h = min(1. / steps, time-epsilon, .5 / max(float(exit_rate.max()), 1e-8))
        change = torch.rand(z.shape) < h * exit_rate
        candidate = draw(rates / exit_rate.clamp_min(1e-12)[..., None] + 1e-12)
        z = torch.where(change & (z == MASK), candidate, z)
        time -= h
        iterations += 1
        if iterations > 100000:
            raise RuntimeError('Guided rate integration failed to advance.')
    z = torch.where(z == MASK, draw(model(z).softmax(-1)), z)
    return z


class SearchNode:
    def __init__(self, tokens, parent=None, prior=1.):
        self.tokens, self.parent, self.prior = tokens, parent, prior
        self.children = []
        self.visits = 0
        self.reward = np.zeros(2)


@torch.no_grad()
def peptune_search(model, length=8, iterations=100, branching=4):
    """DNA MCTS: selection, expansion, completion, Pareto rewards, backup.

    All DNA strings are valid. Peptide chemistry, bond-dependent masks,
    RoFormer training, and the PepTune invalid-SMILES penalty are not used.
    """
    root = SearchNode(torch.full((length,), MASK))
    archive = torch.empty((0, length), dtype=torch.long)
    archive_scores = torch.empty((0, 2))
    trace = []
    for iteration in range(iterations):
        node = root
        while node.children:
            def selection(child):
                mean = child.reward.mean() / max(child.visits, 1)
                bonus = 1.5 * child.prior * math.sqrt(node.visits + 1) / (child.visits + 1)
                return mean + bonus
            node = max(node.children, key=selection)
        if (node.tokens == MASK).any():
            p = model(node.tokens[None]).softmax(-1)[0]
            positions = torch.where(node.tokens == MASK)[0]
            seen = set()
            for _ in range(branching):
                pos = int(positions[torch.randint(len(positions), ())])
                token = int(torch.multinomial(p[pos], 1))
                candidate = node.tokens.clone()
                candidate[pos] = token
                key = tuple(candidate.tolist())
                if key not in seen:
                    node.children.append(SearchNode(candidate, node, float(p[pos, token])))
                    seen.add(key)
            node = node.children[0]
        rollout = node.tokens.clone()
        while (rollout == MASK).any():
            prob = model(rollout[None]).softmax(-1)
            position = int(torch.where(rollout == MASK)[0][0])
            rollout[position] = draw(prob)[0, position]
        score = objectives(rollout[None])[0]
        reward = ((score >= archive_scores).float().mean(0).numpy()
                  if len(archive) else np.ones(2))
        archive = torch.cat((archive, rollout[None]))
        archive_scores = torch.cat((archive_scores, score[None]))
        archive, archive_scores = pareto_filter(archive, archive_scores)
        # Equal-score alternatives remain; remove exact repeated sequences only.
        unique = []; seen = set()
        for i, row in enumerate(archive.tolist()):
            key = tuple(row)
            if key not in seen:
                unique.append(i); seen.add(key)
        archive, archive_scores = archive[unique], archive_scores[unique]
        while node is not None:
            node.visits += 1
            node.reward += reward
            node = node.parent
        trace.append(dict(iteration=iteration, archive_size=len(archive), score=score.tolist()))
    return archive, archive_scores, trace