File size: 7,277 Bytes
f771f94 | 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 | from __future__ import annotations
import json
import math
from pathlib import Path
import numpy as np
import pandas as pd
import torch
import trackio
from model import ConditionalNeuralProcess, parameter_count
from safetensors.torch import save_file
PROJECT_DIR = Path(__file__).resolve().parent
ARTIFACT_DIR = PROJECT_DIR / "artifacts" / "neural-process-pocket"
DATA_DIR = PROJECT_DIR / "data"
SEED = 2141
def sample_tasks(
tasks: int,
context_points: int,
target_points: int,
seed: int,
) -> tuple[torch.Tensor, ...]:
generator = torch.Generator().manual_seed(seed)
amplitude = 0.1 + 4.9 * torch.rand(tasks, 1, 1, generator=generator)
phase = 2 * math.pi * torch.rand(tasks, 1, 1, generator=generator)
frequency = 0.8 + 0.4 * torch.rand(tasks, 1, 1, generator=generator)
context_x = -5 + 10 * torch.rand(
tasks, context_points, 1, generator=generator
)
target_x = -5 + 10 * torch.rand(tasks, target_points, 1, generator=generator)
context_y = amplitude * torch.sin(frequency * context_x + phase)
target_y = amplitude * torch.sin(frequency * target_x + phase)
context_y += 0.03 * torch.randn(context_y.shape, generator=generator)
return context_x, context_y, target_x, target_y, amplitude, phase, frequency
def gaussian_nll(
mean: torch.Tensor,
standard_deviation: torch.Tensor,
target: torch.Tensor,
) -> torch.Tensor:
variance = standard_deviation.square()
return (
0.5 * ((target - mean).square() / variance)
+ standard_deviation.log()
+ 0.5 * math.log(2 * math.pi)
).mean()
def rbf_gp(
context_x: np.ndarray,
context_y: np.ndarray,
target_x: np.ndarray,
) -> tuple[np.ndarray, np.ndarray]:
def kernel(left: np.ndarray, right: np.ndarray) -> np.ndarray:
return np.exp(-0.5 * (left[:, None] - right[None, :]) ** 2)
covariance = kernel(context_x, context_x) + np.eye(len(context_x)) * 0.01
cross = kernel(context_x, target_x)
solve_y = np.linalg.solve(covariance, context_y)
mean = cross.T @ solve_y
solve_cross = np.linalg.solve(covariance, cross)
variance = 1.0 - np.sum(cross * solve_cross, axis=0)
return mean, np.sqrt(np.clip(variance + 0.03**2, 0.03**2, None))
@torch.inference_mode()
def evaluate(model: ConditionalNeuralProcess, tasks: int = 500) -> tuple[dict, list]:
context_x, context_y, target_x, target_y, amplitude, phase, frequency = sample_tasks(
tasks, 5, 100, SEED + 20_000
)
mean, std = model(context_x, context_y, target_x)
error = mean - target_y
cnp = {
"rmse": float(error.square().mean().sqrt()),
"gaussian_nll": float(gaussian_nll(mean, std, target_y)),
"coverage_90": float(
((target_y >= mean - 1.645 * std) & (target_y <= mean + 1.645 * std))
.float()
.mean()
),
}
gp_errors = []
gp_nll = []
gp_covered = []
rows = []
for index in range(tasks):
cx = context_x[index, :, 0].numpy()
cy = context_y[index, :, 0].numpy()
tx = target_x[index, :, 0].numpy()
ty = target_y[index, :, 0].numpy()
gp_mean, gp_std = rbf_gp(cx, cy, tx)
gp_errors.extend((gp_mean - ty).tolist())
gp_nll.extend(
(
0.5 * ((ty - gp_mean) / gp_std) ** 2
+ np.log(gp_std)
+ 0.5 * np.log(2 * np.pi)
).tolist()
)
gp_covered.extend(
((ty >= gp_mean - 1.645 * gp_std) & (ty <= gp_mean + 1.645 * gp_std))
.astype(float)
.tolist()
)
rows.append(
{
"amplitude": float(amplitude[index, 0, 0]),
"phase": float(phase[index, 0, 0]),
"frequency": float(frequency[index, 0, 0]),
"context_x": cx.tolist(),
"context_y": cy.tolist(),
}
)
gp_errors_array = np.asarray(gp_errors)
report = {
"conditional_neural_process": cnp,
"fixed_rbf_gaussian_process": {
"rmse": float(np.sqrt(np.mean(gp_errors_array**2))),
"gaussian_nll": float(np.mean(gp_nll)),
"coverage_90": float(np.mean(gp_covered)),
},
"tasks": tasks,
"context_points_per_task": 5,
"targets_per_task": 100,
}
return report, rows
def main() -> None:
torch.manual_seed(SEED)
torch.set_num_threads(1)
model = ConditionalNeuralProcess()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-5)
best = float("inf")
best_state = None
best_step = 0
trackio.init(
project="neural-process-pocket",
name="conditional-neural-process-v1",
config={
"parameters": parameter_count(model),
"training_steps": 5_000,
"context_points": 5,
"tasks_per_step": 64,
},
)
for step in range(1, 5_001):
context_x, context_y, target_x, target_y, *_ = sample_tasks(
64, 5, 50, SEED + step
)
mean, std = model(context_x, context_y, target_x)
loss = gaussian_nll(mean, std, target_y)
optimizer.zero_grad(set_to_none=True)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
optimizer.step()
if step % 250 == 0:
validation, _ = evaluate(model, tasks=100)
validation_nll = validation["conditional_neural_process"]["gaussian_nll"]
trackio.log(
{
"training_step": step,
"training_nll": float(loss.detach()),
**{
f"validation_{key}": value
for key, value in validation[
"conditional_neural_process"
].items()
},
}
)
if validation_nll < best:
best = validation_nll
best_step = step
best_state = {
name: value.detach().cpu().clone()
for name, value in model.state_dict().items()
}
assert best_state is not None
model.load_state_dict(best_state)
benchmark, rows = evaluate(model)
report = {
"model": "Neural Process Pocket",
"parameters": parameter_count(model),
"best_step": best_step,
"benchmark": benchmark,
}
ARTIFACT_DIR.mkdir(parents=True, exist_ok=True)
DATA_DIR.mkdir(parents=True, exist_ok=True)
save_file(model.state_dict(), ARTIFACT_DIR / "model.safetensors")
(ARTIFACT_DIR / "evaluation.json").write_text(
json.dumps(report, indent=2), encoding="utf-8"
)
pd.DataFrame(rows).to_parquet(DATA_DIR / "heldout_tasks.parquet", index=False)
trackio.log(
{
"test_rmse": benchmark["conditional_neural_process"]["rmse"],
"test_nll": benchmark["conditional_neural_process"]["gaussian_nll"],
"test_coverage_90": benchmark["conditional_neural_process"][
"coverage_90"
],
}
)
trackio.finish()
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
|