File size: 6,677 Bytes
d486718 | 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 | from __future__ import annotations
import json
from pathlib import Path
import joblib
import numpy as np
import pandas as pd
import torch
import trackio
from data import generate_pinwheel
from model import RealNVP, parameter_count
from safetensors.torch import save_file
from sklearn.mixture import GaussianMixture
PROJECT_DIR = Path(__file__).resolve().parent
ARTIFACT_DIR = PROJECT_DIR / "artifacts" / "flow-pocket"
DATA_DIR = PROJECT_DIR / "data"
def gaussian_fit(data: np.ndarray) -> dict:
return {"mean": data.mean(0), "covariance": np.cov(data.T)}
def gaussian_nll(data: np.ndarray, fit: dict) -> float:
centered = data - fit["mean"]
covariance = fit["covariance"]
inverse = np.linalg.inv(covariance)
log_determinant = np.linalg.slogdet(covariance)[1]
quadratic = np.einsum("bi,ij,bj->b", centered, inverse, centered)
return float(np.mean(np.log(2 * np.pi) + 0.5 * log_determinant + 0.5 * quadratic))
def gaussian_sample(fit: dict, samples: int, seed: int) -> np.ndarray:
return np.random.default_rng(seed).multivariate_normal(
fit["mean"], fit["covariance"], size=samples
)
def rbf_mmd(first: np.ndarray, second: np.ndarray) -> float:
rng = np.random.default_rng(2043)
first = first[rng.choice(len(first), 1000, replace=False)]
second = second[rng.choice(len(second), 1000, replace=False)]
combined = np.concatenate([first, second])
pairs = rng.choice(len(combined), size=(4000, 2), replace=True)
distances = np.sum(
(combined[pairs[:, 0]] - combined[pairs[:, 1]]) ** 2, axis=1
)
bandwidth = max(float(np.median(distances[distances > 0])), 1e-4)
def kernel_mean(left: np.ndarray, right: np.ndarray) -> float:
distances = ((left[:, None, :] - right[None, :, :]) ** 2).sum(2)
return float(np.exp(-distances / (2 * bandwidth)).mean())
return kernel_mean(first, first) + kernel_mean(second, second) - 2 * kernel_mean(
first, second
)
def main() -> None:
torch.manual_seed(2043)
torch.set_num_threads(1)
train_data = generate_pinwheel(40_000, 2043)
validation_data = generate_pinwheel(5_000, 3043)
test_data = generate_pinwheel(10_000, 4043)
model = RealNVP()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-6)
tensor = torch.from_numpy(train_data)
validation = torch.from_numpy(validation_data)
rng = np.random.default_rng(2043)
trackio.init(
project="flow-pocket",
name="realnvp-pinwheel-v1",
config={
"parameters": parameter_count(model),
"coupling_layers": len(model.layers),
"training_examples": len(train_data),
"training_steps": 4_000,
},
)
best_state = None
best_validation = float("inf")
history = []
for step in range(1, 4_001):
batch = tensor[rng.choice(len(tensor), 512, replace=False)]
loss = -model.log_probability(batch).mean()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5.0)
optimizer.step()
if step % 100 == 0:
model.eval()
with torch.inference_mode():
validation_nll = float(
-model.log_probability(validation).mean()
)
record = {
"training_step": step,
"training_nll": float(loss.detach()),
"validation_nll": validation_nll,
}
history.append(record)
trackio.log(record)
if validation_nll < best_validation:
best_validation = validation_nll
best_state = {
name: parameter.detach().clone()
for name, parameter in model.state_dict().items()
}
model.train()
if best_state is not None:
model.load_state_dict(best_state)
model.eval()
gaussian = gaussian_fit(train_data)
mixture = GaussianMixture(
n_components=5,
covariance_type="full",
random_state=2043,
max_iter=500,
n_init=3,
).fit(train_data)
with torch.inference_mode():
flow_nll = float(
-model.log_probability(torch.from_numpy(test_data)).mean()
)
generated_flow = model.sample(5_000, seed=5043).numpy()
latent, _ = model(torch.from_numpy(test_data[:2_000]))
reconstructed = model.inverse(latent)
cycle_error = float(
torch.max(torch.abs(reconstructed - torch.from_numpy(test_data[:2_000])))
)
generated_gaussian = gaussian_sample(gaussian, 5_000, 6043)
generated_mixture, _ = mixture.sample(5_000)
results = {
"realnvp": {
"parameters": parameter_count(model),
"test_nll": flow_nll,
"sample_mmd": rbf_mmd(generated_flow, test_data),
"maximum_cycle_error": cycle_error,
},
"full_covariance_gaussian": {
"test_nll": gaussian_nll(test_data, gaussian),
"sample_mmd": rbf_mmd(generated_gaussian, test_data),
},
"five_component_gmm": {
"test_nll": float(-mixture.score(test_data)),
"sample_mmd": rbf_mmd(generated_mixture, test_data),
},
}
report = {
"benchmark": "Five-arm pinwheel density estimation",
"training_examples": len(train_data),
"heldout_examples": len(test_data),
"results": results,
"training_history": history,
}
ARTIFACT_DIR.mkdir(parents=True, exist_ok=True)
DATA_DIR.mkdir(parents=True, exist_ok=True)
save_file(model.state_dict(), ARTIFACT_DIR / "realnvp.safetensors")
joblib.dump(
{"gaussian": gaussian, "mixture": mixture},
ARTIFACT_DIR / "classical_controls.joblib",
)
(ARTIFACT_DIR / "evaluation.json").write_text(
json.dumps(report, indent=2), encoding="utf-8"
)
np.savez_compressed(
ARTIFACT_DIR / "generated_samples.npz",
realnvp=generated_flow,
gaussian=generated_gaussian,
gmm=generated_mixture,
)
pd.DataFrame(test_data, columns=["x", "y"]).to_parquet(
DATA_DIR / "pinwheel_test.parquet", index=False
)
trackio.log(
{
"flow_test_nll": results["realnvp"]["test_nll"],
"flow_sample_mmd": results["realnvp"]["sample_mmd"],
"gmm_test_nll": results["five_component_gmm"]["test_nll"],
"gmm_sample_mmd": results["five_component_gmm"]["sample_mmd"],
}
)
trackio.finish()
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
|