File size: 7,479 Bytes
da3babb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Flow matching, finite-map identities, Shortcut, MeanFlow, and latent maps.

# %% 1. Learn local motion before learning an interval.
All times run noise -> data except MeanFlow, whose original backward clock
r <= t runs data at zero -> noise at one during training.
"""
import torch
from torch import nn
from torch.nn import functional as F

from common import (MapNet, draw_batch, ema_copy, finite_map, interpolate, mlp,
                    optimize, ordered_times, time_like)


def diagonal_loss(model, data):
    t = torch.rand(len(data), 1, device=data.device)
    x, displacement = interpolate(data, t)
    return F.mse_loss(model(x, t, t), displacement)


def lagrangian_residual(model, velocity, x, s, t):
    # d_t F(s,t,x) = b(t,F(s,t,x)). JVP differentiates only arrival time.
    y, dt = torch.func.jvp(lambda end: finite_map(model, x, s, end),
                         (t,), (torch.ones_like(t),))
    return dt - velocity(y.detach(), t).detach()


def eulerian_residual(model, velocity, x, s, t):
    # (d_s + b_s dot grad_x)F = 0. No full Jacobian is materialized.
    direction = velocity(x, s).detach()
    _, residual = torch.func.jvp(lambda z, start: finite_map(model, z, start, t),
                               (x, s), (direction, torch.ones_like(s)))
    return residual


def semigroup_loss(model, teacher, x, s, t, context=None):
    u = (s + t) / 2
    with torch.no_grad():
        mid = finite_map(teacher, x, s, u, context)
        target = finite_map(teacher, mid, u, t, context)
    prediction = finite_map(model, x, s, t, context)
    return ((prediction - target) / (t - s).clamp_min(.05)).square().mean()


def meanflow_loss(model, data):
    # %% 2. Average backward velocity satisfies u = v - (t-r) D_t u.
    noise = torch.randn_like(data)
    r, t, _ = ordered_times(data, 1.)
    # Include diagonal examples with nonzero probability, as in the paper.
    r = torch.where(torch.rand_like(r) < .75, t, r)
    z = (1 - t) * data + t * noise
    v = noise - data
    average, derivative = torch.func.jvp(model, (z, r, t),
        (v, torch.zeros_like(r), torch.ones_like(t)))
    target = (v - (t - r) * derivative).detach()
    residual = (average - target).square().mean(-1)
    # Detached adaptive weighting controls large self-distillation residuals.
    weight = (residual.detach() + .01).pow(-.5)
    return (weight * residual).mean(), {'mse': residual.mean()}


def shortcut_loss(model, teacher, data):
    t = torch.rand(len(data), 1, device=data.device)
    x, v = interpolate(data, t)
    diag = F.mse_loss(model(x, t, t), v)
    # d is the half-step. Learn the 2d shortcut from two d shortcuts.
    powers = torch.randint(1, 6, (len(data), 1), device=data.device)
    d = 2. ** (-powers)
    start = torch.rand_like(d) * (1 - 2 * d)
    x, _ = interpolate(data, start)
    with torch.no_grad():
        first = teacher(x, start, start + d)
        second = teacher(x + d * first, start + d, start + 2 * d)
        target = .5 * (first + second)
    finite = F.mse_loss(model(x, start, start + 2 * d), target)
    return diag + finite, {'diagonal': diag, 'shortcut': finite}


def train_continuous(method, data, args):
    model = MapNet(data.shape[1], args.width).to(data)
    teacher_logs = []
    teacher = None
    if method in {'fmm-lagrangian', 'fmm-eulerian', 'consistency'}:
        teacher = MapNet(data.shape[1], args.width).to(data)
        teacher_logs = optimize(teacher, lambda _: (diagonal_loss(teacher, draw_batch(data, args.batch_size)), {}),
                                args.teacher_steps, args.lr)
        teacher.requires_grad_(False)
        model.load_state_dict(teacher.state_dict())
    ema = ema_copy(model)

    def objective(step):
        batch = draw_batch(data, args.batch_size)
        if method == 'meanflow':
            return meanflow_loss(model, batch)
        if method == 'shortcut':
            return shortcut_loss(model, ema, batch)
        diag = diagonal_loss(model, batch)
        if method == 'flow-matching':
            return diag, {'diagonal': diag}
        s, t, _ = ordered_times(batch)
        x, _ = interpolate(batch, s)
        if method == 'fmm-lagrangian':
            finite = lagrangian_residual(model, lambda z, a: teacher(z, a, a), x, s, t).square().mean()
        elif method == 'fmm-eulerian':
            finite = eulerian_residual(model, lambda z, a: teacher(z, a, a), x, s, t).square().mean()
        elif method == 'consistency':
            # Endpoint consistency distillation on teacher-solved short intervals.
            # At t=1 the residual endpoint parametrization is exactly identity.
            t = (s + .1).clamp_max(1.)
            with torch.no_grad():
                y = integrate_velocity(lambda z, a: teacher(z, a, a), x, s, t, 4)
                target = finite_map(ema, y, t, torch.ones_like(t))
            finite = F.mse_loss(finite_map(model, x, s, torch.ones_like(s)), target)
        else:
            finite = semigroup_loss(model, ema, x, s, t)
        weight = min(1., (step + 1) / max(1, args.train_steps // 5))
        return diag + weight * finite, {'diagonal': diag, 'finite': finite}

    logs = optimize(model, objective, args.train_steps, args.lr, ema)
    # Use the trained student; EMA supplies fixed bootstrap targets during training.
    state = {'model': model.state_dict(), 'dim': data.shape[1]}
    if teacher is not None:
        state['teacher'] = teacher.state_dict()
    return model, state, logs, teacher_logs


@torch.no_grad()
def integrate_velocity(velocity, x, s, t, steps):
    """Heun integration of a velocity, also supports batch-specific time bounds."""
    s, t = time_like(s, x), time_like(t, x)
    h = (t - s) / steps
    for i in range(steps):
        a = s + i * h
        k1 = velocity(x, a)
        k2 = velocity(x + h * k1, a + h)
        x = x + h * .5 * (k1 + k2)
    return x


@torch.no_grad()
def sample_continuous(model, method, noise, steps):
    if method == 'flow-matching':
        return integrate_velocity(lambda z, a: model(z, a, a), noise, 0., 1., steps)
    if method == 'consistency':
        return finite_map(model, noise, 0., 1.)
    x = noise
    for i in range(steps):
        if method == 'meanflow':
            t, r = 1 - i / steps, 1 - (i + 1) / steps
            x = x - (t - r) * model(x, r, t)
        else:
            x = finite_map(model, x, i / steps, (i + 1) / steps)
    return x


class Autoencoder(nn.Module):
    def __init__(self, width):
        super().__init__()
        self.encoder = mlp(3, 2, width)
        self.decoder = mlp(2, 3, width)


def embed_surface(x):
    return torch.cat([x, .3 * (x[:, :1].square() - x[:, 1:].square())], -1)


def train_latent(data, args):
    # %% 3. First fit the representation, then freeze it while fitting its flow.
    surface = embed_surface(data)
    ae = Autoencoder(args.width).to(data)
    def objective(_):
        batch = draw_batch(surface, args.batch_size)
        reconstruction = ae.decoder(ae.encoder(batch))
        return F.mse_loss(reconstruction, batch), {}
    ae_logs = optimize(ae, objective, args.teacher_steps, args.lr)
    ae.requires_grad_(False)
    with torch.no_grad():
        z = ae.encoder(surface)
        mean, std = z.mean(0), z.std(0).clamp_min(.05)
        z = (z - mean) / std
    model, state, logs, _ = train_continuous('self-distill', z, args)
    state.update({'autoencoder': ae.state_dict(), 'latent_mean': mean, 'latent_std': std})
    return model, ae, state, logs, ae_logs