NPE_importance-sampling/run_experiments.py at main · dhia680/NPE_importance-sampling · GitHub

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
import argparse
import os
import sys
import torch
import torch.optim as optim
import numpy as np
import warnings
warnings.filterwarnings('ignore')

sys.path.insert(0, os.path.dirname(__file__))

from models.flow import ConditionalNormalizingFlow
from posteriors.toy import ToyPosterior2D, ToyPosterior3D
from training.objectives import train_forward_kl, train_reverse_kl, train_alpha_divergence, train_annealed
from inference.is_utils import importance_sampling, importance_sampling_faithful_synthetic, run_3d_synthetic_faithful
from plotting.plots import plot_2d_results, plot_3d_results, plot_qualitative_2d, plot_weight_distributions

DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')


def run_2d(out_dir, n_epochs, n_is):
    sigmas = [0.3, 0.1, 0.04, 0.02, 0.01]
    results_fkl, results_rkl, results_a05, results_anneal, results_synth = [], [], [], [], []

    for sigma in sigmas:
        print(f'\n=== 2D | sigma={sigma} ===')
        posterior = ToyPosterior2D(sigma=sigma)

        flow, _ = train_forward_kl(ConditionalNormalizingFlow(2, 0, 8, 128), posterior, n_epochs=n_epochs)
        theta, w, eps, logZ = importance_sampling(flow, posterior, n_is)
        results_fkl.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  fKL:      ε={eps:.4f}')

        flow, _ = train_reverse_kl(ConditionalNormalizingFlow(2, 0, 8, 128), posterior, n_epochs=n_epochs)
        theta, w, eps, logZ = importance_sampling(flow, posterior, n_is)
        results_rkl.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  rKL:      ε={eps:.4f}')

        flow, _ = train_alpha_divergence(ConditionalNormalizingFlow(2, 0, 8, 128), posterior, alpha=0.5, n_epochs=n_epochs)
        theta, w, eps, logZ = importance_sampling(flow, posterior, n_is)
        results_a05.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  α=0.5:    ε={eps:.4f}')

        flow = train_annealed(ToyPosterior2D, sigma_target=sigma, sigma_start=1.0,
                              n_anneal_steps=10, n_epochs_per_step=n_epochs, dim=2)
        theta, w, eps, logZ = importance_sampling(flow, posterior, n_is)
        results_anneal.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  Annealed: ε={eps:.4f}')

        flow_1d = ConditionalNormalizingFlow(1, 0, 6, 64).to(DEVICE)
        data_1d = posterior.sample(50000)[:, :1].to(DEVICE)
        opt     = optim.Adam(flow_1d.parameters(), lr=1e-3)
        for _ in range(n_epochs):
            idx  = torch.randperm(50000)[:512]
            loss = -flow_1d.log_prob(data_1d[idx]).mean()
            opt.zero_grad(); loss.backward(); opt.step()

        theta, w, eps, logZ, _, _ = importance_sampling_faithful_synthetic(flow_1d, posterior, n_samples=n_is)
        results_synth.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  Synthetic:ε={eps:.4f}')

    plot_2d_results(sigmas, results_fkl, results_rkl, results_a05, results_anneal, results_synth, out_dir)
    plot_qualitative_2d(results_fkl, sigmas, sigma_show=0.01, posterior_class=ToyPosterior2D, out_dir=out_dir)
    plot_weight_distributions(results_fkl, sigmas, out_dir)
    print(f'\nFigures saved to {out_dir}/')


def run_3d(out_dir, n_is):
    sigmas_3d = [0.3, 0.1, 0.05, 0.02, 0.01]
    results_3d_full, results_3d_a05, results_3d_anneal, results_3d_synth = [], [], [], []

    for sigma in sigmas_3d:
        print(f'\n=== 3D | sigma={sigma} ===')
        posterior3d = ToyPosterior3D(sigma=sigma)

        flow, _ = train_forward_kl(ConditionalNormalizingFlow(3, 0, 10, 128), posterior3d, n_epochs=500)
        theta, w, eps, logZ = importance_sampling(flow, posterior3d, n_is)
        results_3d_full.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  fKL:      ε={eps:.4f}')

        flow, _ = train_alpha_divergence(ConditionalNormalizingFlow(3, 0, 10, 128), posterior3d,
                                          alpha=0.5, n_epochs=500)
        theta, w, eps, logZ = importance_sampling(flow, posterior3d, n_is)
        results_3d_a05.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  α=0.5:    ε={eps:.4f}')

        flow = train_annealed(ToyPosterior3D, sigma_target=sigma, sigma_start=1.0,
                              n_anneal_steps=8, n_epochs_per_step=400, dim=3, n_layers=10)
        theta, w, eps, logZ = importance_sampling(flow, posterior3d, n_is)
        results_3d_anneal.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  Annealed: ε={eps:.4f}')

        class P2D:
            def sample(self, n): return posterior3d.sample(n)[:, :2]
            def log_prob(self, x): return x.new_zeros(x.shape[0])

        flow_2d, _ = train_forward_kl(ConditionalNormalizingFlow(2, 0, 8, 128), P2D(), n_epochs=400)
        theta, w, eps, logZ, _, _ = run_3d_synthetic_faithful(flow_2d, posterior3d, n_samples=n_is)
        results_3d_synth.append({'sigma': sigma, 'epsilon': eps, 'logZ': logZ, 'theta': theta, 'weights': w})
        print(f'  Synthetic:ε={eps:.4f}')

    plot_3d_results(sigmas_3d, results_3d_full, results_3d_a05, results_3d_anneal, results_3d_synth, out_dir)
    print(f'\nFigures saved to {out_dir}/')


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('dim', choices=['2d', '3d'], help='Which experiment to run')
    parser.add_argument('--out_dir', default='figures', help='Output directory for figures')
    parser.add_argument('--n_epochs', type=int, default=400, help='Training epochs (2D only)')
    parser.add_argument('--n_is',     type=int, default=20000, help='IS sample count')
    parser.add_argument('--seed',     type=int, default=42)
    args = parser.parse_args()

    torch.manual_seed(args.seed)
    np.random.seed(args.seed)
    os.makedirs(args.out_dir, exist_ok=True)
    print(f'Device: {DEVICE}')

    if args.dim == '2d':
        run_2d(args.out_dir, args.n_epochs, args.n_is)
    else:
        run_3d(args.out_dir, args.n_is)


if __name__ == '__main__':
    main()