PFNs/example_config.py at main · SamuelGabriel/PFNs · 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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
#!/usr/bin/env python3
"""
Example configuration file for PFN training.
This is a Hebo+ prior configuration, as found in the PFNs4BO paper.
This file demonstrates how to configure the MainConfig for training using Python.
"""

import math

import torch
from pfns.model import bar_distribution
from pfns.model.encoders import EncoderConfig, StyleEncoderConfig
from pfns.priors.hyperparameter_sampling import ChoiceDistConfig, UniformFloatDistConfig
from pfns.priors.prior import AdhocPriorConfig
from pfns.train import (
    BatchShapeSamplerConfig,
    MainConfig,
    OptimizerConfig,
    TransformerConfig,
)
from pfns.utils import product_dict

from tqdm import tqdm


def get_config(config_index: int):
    config_dicts = product_dict(
        {
            "sampled_hp_prior": [True, False],
            "emsize": [128, 256, 512],
            "epochs": [100, 400],
            "lr": [1e-4, 3e-4],
        }
    )

    config_dict = list(config_dicts)[config_index]

    sampled_hp_prior = config_dict["sampled_hp_prior"]
    emsize = config_dict["emsize"]
    epochs = config_dict["epochs"]
    lr = config_dict["lr"]

    num_workers = 6
    steps_per_epoch = 1000

    def get_prior_config(sampled_hp_prior=sampled_hp_prior, plotting=False):
        hyperparameters = {
            "lengthscale_mean": 0.7958,  # uniform dist
            "lengthscale_std": 0.7233,  # uniform dist
            "outputscale_mean": 2.1165,  # uniform dist
            "outputscale_std": 2.3021,  # uniform dist
            "add_linear_kernel": False,  # dist as likelihood
            "unused_feature_likelihood": 0.3,  # dist, uniform 0. to .6
            "observation_noise": True,  # dist
            "x_sampler": "normal",  # fixed
            "batch_size_per_gp_sample": 1,
            "hebo_noise_logmean": -4.63,
            "hebo_noise_std": 0.5,
        }
        if sampled_hp_prior:
            hyperparameters.update(
                {
                    "num_hyperparameter_samples_per_batch": -1,
                    "hyperparameter_sampling_add_hps_to_style": "all_sampled",
                    "hyperparameter_sampling_skip_style_prob": 0.1,
                }
            )
            hyperparameters["lengthscale_mean"] = UniformFloatDistConfig(0.5, 1.5)
            hyperparameters["outputscale_mean"] = UniformFloatDistConfig(0.5, 3.0)
            hyperparameters["lengthscale_std"] = UniformFloatDistConfig(0.1, 1.5)
            hyperparameters["outputscale_std"] = UniformFloatDistConfig(0.1, 3.0)
            hyperparameters["unused_feature_likelihood"] = UniformFloatDistConfig(
                0.0, 0.6
            )
            hyperparameters["add_linear_kernel"] = UniformFloatDistConfig(0.0, 1.0)
            hyperparameters["observation_noise"] = ChoiceDistConfig([True, False])
            hyperparameters["hebo_noise_logmean"] = UniformFloatDistConfig(-8.0, -2.0)
            hyperparameters["hebo_noise_std"] = UniformFloatDistConfig(0.1, 5.0)

        prior_config = AdhocPriorConfig(
            prior_names=["hebo_prior"]
            + (["hyperparameter_sampling"] if sampled_hp_prior else []),
            prior_kwargs={
                "num_features": 1 if plotting else 18,
                "hyperparameters": {**hyperparameters},
            },
        )
        return prior_config, hyperparameters

    prior_config, hps = get_prior_config(sampled_hp_prior=sampled_hp_prior)

    gb = prior_config.create_get_batch_method()

    ys = []
    for num_features in tqdm(list(range(1, 11)) * 3):
        ys.append(
            gb(batch_size=16, seq_len=100, num_features=num_features).target_y.flatten()
        )

    ys = torch.cat(ys)
    print(f"{len(ys)=}")

    borders = bar_distribution.get_bucket_borders(1000, ys=ys)

    config = MainConfig(
        priors=[prior_config],
        optimizer=OptimizerConfig("adamw", lr=lr, weight_decay=0.0),
        scheduler="constant",
        model=TransformerConfig(
            criterion=bar_distribution.BarDistributionConfig(
                borders.tolist(), full_support=True
            ),
            emsize=emsize,
            nhead=emsize // 32,
            nhid=emsize * 4,
            nlayers=8,
            encoder=EncoderConfig(
                variable_num_features_normalization=True,
                constant_normalization_mean=0.5,
                constant_normalization_std=1 / math.sqrt(12),
            ),
            y_encoder=EncoderConfig(nan_handling=True),
            attention_between_features=True,
            style_encoder=StyleEncoderConfig(normalize_to_hyperparameters=hps)
            if sampled_hp_prior
            else None,
            y_style_encoder=StyleEncoderConfig(normalize_to_hyperparameters=hps)
            if sampled_hp_prior
            else None,
        ),
        batch_shape_sampler=BatchShapeSamplerConfig(
            batch_size=32,
            max_seq_len=60,
            fixed_num_test_instances=10,
            max_num_features=18,
        ),
        epochs=epochs,
        warmup_epochs=epochs // 10,
        steps_per_epoch=steps_per_epoch,
        num_workers=num_workers,
        train_mixed_precision=False,
    )
    return config


# View with: tensorboard --logdir=runs