File size: 5,031 Bytes
2facf1f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import jax
import jax.numpy as jnp
import optax
import numpy as np
from dataclasses import dataclass


@dataclass
class Hyperparameters:
    num_intervals: int = 1000
    learning_rate: float = 0.01
    end_lr_factor: float = 1e-5
    num_steps: int = 80000
    warmup_steps: int = 3000
    smooth_max_temp: float = 100.0  # temperature for log-sum-exp smooth max


class AutocorrelationOptimizer:
    def __init__(self, hypers: Hyperparameters):
        self.hypers = hypers
        self.domain_width = 0.5
        self.dx = self.domain_width / self.hypers.num_intervals

    def _objective_fn(self, f_values: jnp.ndarray, temp: float) -> jnp.ndarray:
        f_non_negative = jax.nn.softplus(f_values)  # smooth non-negativity
        integral_f = jnp.sum(f_non_negative) * self.dx
        eps = 1e-12
        integral_f_safe = jnp.maximum(integral_f, eps)

        N = self.hypers.num_intervals
        padded_f = jnp.pad(f_non_negative, (0, N))
        fft_f = jnp.fft.rfft(padded_f)
        fft_conv = fft_f * fft_f
        conv_f_f = jnp.fft.irfft(fft_conv, n=2 * N)

        scaled_conv_f_f = conv_f_f * self.dx

        # Use log-sum-exp smooth max for better gradients
        smooth_max = jax.nn.logsumexp(temp * scaled_conv_f_f) / temp

        c1_ratio = smooth_max / (integral_f_safe ** 2)
        return c1_ratio

    def _hard_objective(self, f_values: jnp.ndarray) -> jnp.ndarray:
        f_non_negative = jax.nn.softplus(f_values)
        integral_f = jnp.sum(f_non_negative) * self.dx
        eps = 1e-12
        integral_f_safe = jnp.maximum(integral_f, eps)

        N = self.hypers.num_intervals
        padded_f = jnp.pad(f_non_negative, (0, N))
        fft_f = jnp.fft.rfft(padded_f)
        fft_conv = fft_f * fft_f
        conv_f_f = jnp.fft.irfft(fft_conv, n=2 * N)
        scaled_conv_f_f = conv_f_f * self.dx
        max_conv = jnp.max(scaled_conv_f_f)
        c1_ratio = max_conv / (integral_f_safe ** 2)
        return c1_ratio

    def train_step(self, f_values, opt_state, temp):
        loss, grads = jax.value_and_grad(self._objective_fn)(f_values, temp)
        updates, opt_state = self.optimizer.update(grads, opt_state, f_values)
        f_values = optax.apply_updates(f_values, updates)
        hard_loss = self._hard_objective(f_values)
        return f_values, opt_state, loss, hard_loss

    def run_optimization(self, seed=42):
        schedule = optax.warmup_cosine_decay_schedule(
            init_value=0.0,
            peak_value=self.hypers.learning_rate,
            warmup_steps=self.hypers.warmup_steps,
            decay_steps=self.hypers.num_steps - self.hypers.warmup_steps,
            end_value=self.hypers.learning_rate * self.hypers.end_lr_factor,
        )
        self.optimizer = optax.adam(learning_rate=schedule)

        key = jax.random.PRNGKey(seed)
        N = self.hypers.num_intervals

        # Better initialization: triangle-like shape centered
        x = jnp.linspace(0, 1, N)
        # Start with a bump function shape
        f_values = jnp.exp(-20.0 * (x - 0.5) ** 2)
        f_values = f_values + 0.05 * jax.random.uniform(key, (N,))

        opt_state = self.optimizer.init(f_values)

        train_step_jit = jax.jit(self.train_step)

        best_loss = jnp.inf
        best_f = f_values

        for step in range(self.hypers.num_steps):
            # Anneal temperature: start low, increase
            progress = min(step / (self.hypers.num_steps * 0.5), 1.0)
            temp = 10.0 + progress * (self.hypers.smooth_max_temp - 10.0)

            f_values, opt_state, loss, hard_loss = train_step_jit(f_values, opt_state, temp)

            if hard_loss < best_loss:
                best_loss = hard_loss
                best_f = f_values

            if step % 5000 == 0 or step == self.hypers.num_steps - 1:
                print(f"Step {step:5d} | C1(smooth) ≈ {loss:.8f} | C1(hard) ≈ {hard_loss:.8f} | temp={temp:.1f}")

        print(f"Best C1 found: {best_loss:.8f}")

        # Convert softplus to actual non-negative values
        final_f = jax.nn.softplus(best_f)
        final_c1 = float(best_loss)
        return final_f, final_c1


def run():
    best_c1 = float('inf')
    best_result = None

    # Try multiple seeds and configs
    configs = [
        (1000, 0.01, 80000, 42),
        (1000, 0.005, 80000, 123),
        (1500, 0.008, 60000, 42),
    ]

    for n_intervals, lr, steps, seed in configs:
        print(f"\n--- Config: N={n_intervals}, lr={lr}, steps={steps}, seed={seed} ---")
        hypers = Hyperparameters(
            num_intervals=n_intervals,
            learning_rate=lr,
            num_steps=steps,
        )
        optimizer = AutocorrelationOptimizer(hypers)
        optimized_f, final_c1 = optimizer.run_optimization(seed=seed)

        if final_c1 < best_c1:
            best_c1 = final_c1
            f_values_np = np.array(optimized_f)
            best_result = (f_values_np, best_c1, best_c1, hypers.num_intervals)
            print(f"*** New best: C1 = {best_c1:.10f}")

    return best_result