File size: 7,104 Bytes
5e4510c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Disable progress bar for cleaner output logs
import os

os.environ["TQDM_DISABLE"] = "1"

# Fixed parameters
n, m, p = 2, 4, 5

# EVOLVE-BLOCK-START
import numpy as np
import jax
import jax.numpy as jnp
import optax
from dataclasses import dataclass
import tqdm


# --- Straight-Through Estimator for Rounding ---
@jax.custom_vjp
def round_to_half_ste(x):
    """Forward pass: snaps values to the nearest half-integer."""
    return jnp.round(x * 2) / 2


def round_ste_fwd(x):
    """Standard forward pass and identity for backward pass."""
    return round_to_half_ste(x), None


def round_ste_bwd(res, g):
    """Backward pass: Identity function, passes gradient straight through."""
    return (g,)


round_to_half_ste.defvjp(round_ste_fwd, round_ste_bwd)
# --- End of STE definition ---


# --- Loss Functions ---
def weighted_l2_loss(reconstructed: jnp.ndarray, target: jnp.ndarray) -> jnp.ndarray:
    error = reconstructed - target
    weights = jnp.where(target != 0, 100.0, 1.0)
    return jnp.mean(weights * (error**2))


def l2_loss_real(x: jnp.ndarray, y: jnp.ndarray) -> jnp.ndarray:
    return jnp.mean((x - y) ** 2)


# --- Hyperparameters ---
@dataclass
class Hyperparameters:
    rank: int = 55
    # Phase 1: Continuous Search
    num_restarts: int = 10
    phase1_steps: int = 80000
    phase1_lr: float = 0.01
    init_scale: float = 0.1
    l1_strength: float = 1e-6
    clamp_range: float = 4.0
    # Phase 2: Discrete Fine-tuning
    phase2_steps: int = 20000
    phase2_lr: float = 1e-4  # A much smaller learning rate for fine-tuning


# --- Optimizer Classes ---
class ContinuousOptimizer:
    """Finds a high-quality approximate continuous solution."""

    def __init__(self, target_tensor: jnp.ndarray, hypers: Hyperparameters):
        self.target_tensor = target_tensor
        self.hypers = hypers
        self.opt = optax.adam(hypers.phase1_lr)

    def _get_constrained_decomposition(self, latent_decomposition: tuple) -> tuple:
        """Applies a scaled tanh to map latent parameters to the desired range."""
        return jax.tree_util.tree_map(
            lambda x: self.hypers.clamp_range * jnp.tanh(x), latent_decomposition
        )

    def _loss_fn(self, latent_decomposition: tuple) -> jnp.ndarray:
        constrained = self._get_constrained_decomposition(latent_decomposition)
        reconstructed = jnp.einsum("ir,jr,kr->ijk", *constrained)
        recon_loss = weighted_l2_loss(reconstructed, self.target_tensor)
        l1_penalty = sum(jnp.mean(jnp.abs(arr)) for arr in constrained)
        return recon_loss + self.hypers.l1_strength * l1_penalty


class DiscreteOptimizer:
    """Refines a continuous solution into an exact discrete one using an STE."""

    def __init__(self, target_tensor: jnp.ndarray, hypers: Hyperparameters):
        self.target_tensor = target_tensor
        self.hypers = hypers
        self.opt = optax.adam(hypers.phase2_lr)

    def _loss_fn(self, continuous_decomposition: tuple) -> jnp.ndarray:
        # Snap the continuous parameters to the discrete grid
        discrete_decomposition = jax.tree_util.tree_map(round_to_half_ste, continuous_decomposition)
        # Compute the loss using only these exact half-integer values
        reconstructed = jnp.einsum("ir,jr,kr->ijk", *discrete_decomposition)
        return l2_loss_real(reconstructed, self.target_tensor)


# --- JIT-compatible Train Step ---
def train_step(params, opt_state, optimizer, loss_fn):
    loss, grads = jax.value_and_grad(loss_fn)(params)
    updates, opt_state = optimizer.update(grads, opt_state, params)
    params = optax.apply_updates(params, updates)
    return params, opt_state, loss


def get_matrix_multiplication_tensor(n, m, p):
    T = jnp.zeros((n * m, m * p, n * p))
    for i, j, k in np.ndindex(n, m, p):
        T = T.at[i * m + j, j * p + k, k * n + i].set(1)
    return T


def run():
    hypers = Hyperparameters()
    target_tensor = get_matrix_multiplication_tensor(n, m, p)
    main_key = jax.random.PRNGKey(42)

    # --- PHASE 1: CONTINUOUS EXPLORATION ---
    print(f"\n{'='*20} PHASE 1: Continuous Exploration {'='*20}")
    best_loss_phase1 = float("inf")
    best_latent_decomp = None

    continuous_optimizer = ContinuousOptimizer(target_tensor, hypers)

    # JIT the train_step for the continuous phase
    jit_train_step_continuous = jax.jit(train_step, static_argnums=(2, 3))

    for i in range(hypers.num_restarts):
        print(f"\n--- Restart {i+1}/{hypers.num_restarts} ---")
        main_key, restart_key = jax.random.split(main_key)
        init_fn = jax.nn.initializers.normal(stddev=hypers.init_scale)
        latent_decomp = (
            init_fn(restart_key, (n * m, hypers.rank)),
            init_fn(restart_key, (m * p, hypers.rank)),
            init_fn(restart_key, (n * p, hypers.rank)),
        )
        opt_state = continuous_optimizer.opt.init(latent_decomp)

        for _ in tqdm.tqdm(range(hypers.phase1_steps), desc="Continuous Search"):
            latent_decomp, opt_state, loss = jit_train_step_continuous(
                latent_decomp,
                opt_state,
                continuous_optimizer.opt,
                continuous_optimizer._loss_fn,
            )

        final_loss = l2_loss_real(
            target_tensor,
            jnp.einsum(
                "ir,jr,kr->ijk",
                *continuous_optimizer._get_constrained_decomposition(latent_decomp),
            ),
        )
        print(f"End of Trial | Final continuous loss: {final_loss:.8f}")

        if final_loss < best_loss_phase1:
            best_loss_phase1 = final_loss
            best_latent_decomp = latent_decomp

    # --- PHASE 2: DISCRETE FINE-TUNING ---
    print(f"\n{'='*20} PHASE 2: Discrete Fine-tuning (STE) {'='*20}")
    print(f"Starting with best continuous solution (loss: {best_loss_phase1:.8f})")

    continuous_params = continuous_optimizer._get_constrained_decomposition(best_latent_decomp)

    discrete_optimizer = DiscreteOptimizer(target_tensor, hypers)
    opt_state = discrete_optimizer.opt.init(continuous_params)

    # JIT the train_step for the discrete phase
    jit_train_step_discrete = jax.jit(train_step, static_argnums=(2, 3))

    for step in tqdm.tqdm(range(hypers.phase2_steps), desc="Discrete Fine-tuning"):
        continuous_params, opt_state, loss = jit_train_step_discrete(
            continuous_params, opt_state, discrete_optimizer.opt, discrete_optimizer._loss_fn
        )
        if (step + 1) % 2000 == 0:
            print(f"Step {step+1} | Discrete Loss: {loss:.8f}")
        if loss < 1e-7:
            print("\nFound a perfect solution!")
            break

    final_discrete_decomposition = jax.tree_util.tree_map(round_to_half_ste, continuous_params)
    final_loss = l2_loss_real(
        target_tensor, jnp.einsum("ir,jr,kr->ijk", *final_discrete_decomposition)
    )
    print(f"Search complete. Final discrete loss: {final_loss:.8f}")

    final_decomposition_np = jax.tree_util.tree_map(np.array, final_discrete_decomposition)
    return final_decomposition_np, n, m, p, float(final_loss), hypers.rank


# EVOLVE-BLOCK-END