File size: 10,682 Bytes
d8f9639
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Seeded samplers for the saved Lecture 7 teaching models."""
import math
import numpy as np
import torch
from torch.nn import functional as F
from scipy.linalg import expm
from learned_bridges import mlp, MaskedModel, CoupledCone, token_reward
from discrete_learning import rate_model


def generate(state, samples=256, seed=107, sample_steps=None):
    if samples < 1 or (sample_steps is not None and sample_steps < 1):
        raise ValueError('Sample counts and step counts must be positive.')
    torch.manual_seed(seed)
    rng = np.random.default_rng(seed)
    method = state['method']
    steps = sample_steps or state.get('steps', 100)
    traces = []
    report = {'method': method, 'sampling_seed': seed, 'samples': samples}
    result = state.get('result', {})

    if method in ['dsb', 'dsbm']:
        coeffs = np.asarray(state['forward'])
        if steps != len(coeffs):
            raise ValueError('This saved affine model uses its fixed training grid; omit --sample-steps.')
        dt = 1/steps
        x = rng.normal(size=samples)
        start = x.copy()
        traces.append(x[:16].tolist())
        for slope, intercept in coeffs:
            drift = slope*x + intercept
            x = drift if method == 'dsb' else x + dt*drift
            x = x + math.sqrt(state['epsilon']*dt)*rng.normal(size=samples)
            traces.append(x[:16].tolist())
        values = x[:, None]
        report.update(mean=float(x.mean()), variance=float(x.var()),
                      endpoint_covariance=float(np.mean((start-start.mean())*(x-x.mean()))))

    elif method == 'sf2m':
        model = mlp(2, 2)
        model.load_state_dict(state['model'])
        model.eval()
        x = torch.randn(samples, 1)
        dt = 1/steps
        traces.append(x[:16, 0].tolist())
        with torch.no_grad():
            for k in range(steps):
                t = torch.full_like(x, (k+.5)*dt)
                velocity, score = model(torch.cat([t, x], 1)).chunk(2, 1)
                x = x + (velocity + .5*score)*dt + math.sqrt(dt)*torch.randn_like(x)
                traces.append(x[:16, 0].tolist())
        values = x.numpy()
        report.update(mean=float(x.mean()), variance=float(x.var(unbiased=False)),
                      approximation='Euler SDE; boundary times extrapolate beyond the training interval [0.02,0.98].')

    elif method == 'tr2d2':
        length, vocab = state['length'], state['vocab']
        if steps != length:
            raise ValueError('The abstract-token example reveals one of four positions per step.')
        model = MaskedModel(length, vocab)
        model.load_state_dict(state['model'])
        model.eval()
        x = torch.full((samples, length), vocab, dtype=torch.long)
        with torch.no_grad():
            for k in range(length):
                x[:, k] = torch.multinomial(model(x)[:, k].softmax(-1), 1).squeeze(1)
                traces.append(x[:16].tolist())
        values = x.numpy()
        report.update(mean_reward=float(token_reward(x).mean()),
                      vocabulary=['circle', 'square', 'triangle'],
                      sampler='Learned denoiser with fixed left-to-right reveal order; no search at inference.')

    elif method == 'branch':
        model = mlp(6, 2)
        model.load_state_dict(state['model'])
        model.eval()
        x0 = .12*torch.randn(samples, 2)
        x = x0[:, None, :].expand(-1, 3, -1).clone().reshape(-1, 2)
        branch = F.one_hot(torch.arange(3), 3).float().repeat(samples, 1)
        dt = 1/steps
        traces.append(x.reshape(samples, 3, 2)[:16].tolist())
        with torch.no_grad():
            for k in range(steps):
                # Midpoint integration of the learned branch-specific velocity.
                t = torch.full((len(x), 1), k*dt)
                v = model(torch.cat([t, branch, x], 1))
                vm = model(torch.cat([t+dt/2, branch, x+dt*v/2], 1))
                x = x + dt*vm
                traces.append(x.reshape(samples, 3, 2)[:16].tolist())
        values = x.reshape(samples, 3, 2).numpy()
        weights = np.array([1., 0., 0.]) + state['growth'].numpy()
        report.update(terminal_branch_weights=weights.tolist(), total_mass=float(weights.sum()),
                      minimum_weight=float(weights.min()), branch_means=values.mean(0).tolist(),
                      approximation='Learned velocity integration; soft growth constraints. Raw branch weights are retained, including negative residuals.')

    elif method == 'entangled':
        model = CoupledCone()
        model.load_state_dict(state['model'])
        model.eval()
        sigma = state['sigma']
        dt = 1/steps
        x = .3*torch.randn(samples, 3, 2)
        alignment = float('inf')
        traces.append(x[:16].tolist())
        with torch.no_grad():
            for k in range(steps):
                control = model(k*dt, x)
                alignment = min(alignment, float((control*(1-x)).sum(-1).min()))
                reference = -.12*x - .35*(x-x.mean(1, keepdim=True))
                x = x + (reference+sigma*control)*dt + sigma*math.sqrt(dt)*torch.randn_like(x)
                traces.append(x[:16].tolist())
        values = x.numpy()
        report.update(mean=float(x.mean()), minimum_bias_alignment=alignment,
                      terminal_potential_mean=float(torch.exp(-.5*((x-1)**2).sum((1,2))/1.5).mean()),
                      approximation='Overdamped interacting-particle Euler chain; terminal potential, not an enforced target marginal.')

    elif method == 'ddsbm':
        model = rate_model()
        model.load_state_dict(state['model'])
        model.eval()
        x = rng.integers(0, 2, samples)
        traces.append(x[:16].tolist())
        with torch.no_grad():
            for k in range(steps):
                t = (k+.5)/steps
                q = model(torch.tensor([[t, 1., 0.], [t, 0., 1.]])).ravel().numpy()
                total = float(q.sum())
                factor = -math.expm1(-total/steps)/total
                flip = factor*q[x]
                x = np.where(rng.random(samples) < flip, 1-x, x)
                traces.append(x[:16].tolist())
        values = x[:, None]
        report.update(terminal_frequencies=np.bincount(x, minlength=2).tolist(),
                      terminal_probability_one=float(x.mean()), target_probability_one=.8,
                      approximation='Exact two-state transitions for rates frozen at each grid midpoint; finite-time approximation to the learned inhomogeneous CTMC.')

    elif method in ['csbm', 'finite-sb']:
        Q = np.asarray(state['transitions'] if method == 'csbm' else result['Q'])
        if sample_steps is not None and sample_steps != len(Q):
            raise ValueError('This finite-chain checkpoint has two transitions.')
        x = rng.integers(0, 2, samples)
        paths = [x.copy()]
        for q in Q:
            x = (rng.random(samples) < q[x, 1]).astype(int)
            paths.append(x.copy())
        values = np.stack(paths, 1)
        report.update(terminal_probability_one=float(x.mean()), target_probability_one=.8)

    elif method in ['ot', 'sinkhorn']:
        name = 'plan' if method == 'ot' else 'entropic_plan'
        pi = np.asarray(result[name])
        pair = rng.choice(4, samples, p=pi.ravel()/pi.sum())
        values = np.stack([pair//2, pair%2], 1)
        cost = np.asarray(result['cost'])
        report.update(coupling=name, exact_expected_cost=float((pi*cost).sum()),
                      sampled_average_cost=float(cost[values[:, 0], values[:, 1]].mean()))

    elif method == 'discrete-imf':
        prob = np.asarray(result['final'])
        values = np.asarray(result['paths'])[rng.choice(len(prob), samples, p=prob/prob.sum())]
        report.update(terminal_probability_one=float(values[:, -1].mean()),
                      path_l1_to_exact_bridge=result['final_path_l1'])

    elif method == 'ctmc-sb':
        G = np.asarray(result['generator'])
        g = np.asarray(result['values'][-1]['h'])
        x = rng.integers(0, 2, samples)
        paths = [x.copy()]
        K = expm(G/steps)
        for k in range(steps):
            hs = expm((1-k/steps)*G)@g
            ht = expm((1-(k+1)/steps)*G)@g
            Q = K*ht[None, :]/hs[:, None]
            if not np.allclose(Q.sum(1), 1):
                raise ArithmeticError('The Doob transition failed to normalize.')
            x = (rng.random(samples) < Q[x, 1]).astype(int)
            paths.append(x.copy())
        values = np.stack(paths, 1)
        report.update(terminal_probability_one=float(x.mean()),
                      sampler='Exact reference-bridge transitions between observation times, up to matrix-exponential precision.')

    elif method == 'gaussian-sb':
        c = result['covariance']
        t = np.linspace(0, 1, steps+1)
        x0 = rng.normal(size=(samples, 1))
        x1 = 2+c*x0+math.sqrt(1-c*c)*rng.normal(size=(samples, 1))
        W = np.c_[np.zeros(samples), np.cumsum(rng.normal(size=(samples, steps))/math.sqrt(steps), axis=1)]
        paths = (1-t)*x0+t*x1+W-t*W[:, -1, None]
        values = paths[:, -1, None]
        traces = paths[:16].T.tolist()
        report.update(mean=float(values.mean()), variance=float(values.var()),
                      sampler='Exact Gaussian endpoint coupling plus Brownian conditional bridges on the grid.')

    elif method == 'reward-tilt':
        values = rng.choice(3, samples, p=result['target'])[:, None]
        report.update(empirical_probabilities=(np.bincount(values[:, 0], minlength=3)/samples).tolist(),
                      target_probabilities=result['target'])

    elif method == 'branch-mass':
        k = rng.choice(3, samples, p=result['weights'][-1])
        x = np.asarray(result['positions'])[-1, k]
        values = np.stack([k, x], 1)
        report.update(exact_terminal_weights=result['weights'][-1],
                      interpretation='Draw endpoints from the explicitly specified analytic branch mixture.')

    elif method == 'cone-geometry':
        d = np.asarray(result['direction'])
        bias = np.asarray(result['bias'])
        values = np.stack([d, bias, d-result['dt']*bias])
        report.update(samples=3, row_meanings=['target displacement', 'bias', 'displacement after the step'],
                      interpretation='A deterministic geometry calculation; these rows are not random samples.')
    else:
        raise ValueError('Unsupported checkpoint method: '+method)

    if not np.isfinite(values).all():
        raise FloatingPointError('Generated values are not finite.')
    report['finite_values'] = True
    return {'values': values.tolist(), 'trajectory_prefix': traces, 'report': report}