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 | # Copyright 2025 Patrick Armstrong
from typing import Self, Literal, ClassVar, Annotated
from pathlib import Path
import itertools
from pydantic import (
Field,
PositiveInt,
PositiveFloat,
NonNegativeFloat,
field_validator,
model_validator,
)
from supaernova.analysis.spectra import SpectraPlot
from supaernova.configs.steps.data import DataStepConfig
from supaernova.configs.steps.steps import AbstractStepAnalysis
from supaernova.analysis.distribution import DistributionPlot
from supaernova.configs.steps.backends import AbstractModelConfig
class PAEStepAnalysis(AbstractStepAnalysis):
plot_residual: SpectraPlot | list[SpectraPlot] | None = None
plot_latents: DistributionPlot | list[DistributionPlot] | None = None
class PAEModelConfig(AbstractModelConfig):
# --- Class Variables ---
id: ClassVar[str] = "pae_model"
required_steps: ClassVar[list[str]] = [DataStepConfig.id]
analysis: PAEStepAnalysis = PAEStepAnalysis.model_validate({})
# === Required ===
debug: bool = False
profile: bool = False
@model_validator(mode="after")
def validate_bounds(self) -> Self:
for var in ["redshift", "phase"]:
for data in ["train", "test", "val"]:
min_bound = getattr(self, f"min_{data}_{var}")
max_bound = getattr(self, f"max_{data}_{var}")
if min_bound >= max_bound:
err = f"`max_{data}_{var}`: {max_bound} is not strictly greater than `min_{data}_{var}`: {min_bound}"
self._raise(err)
return self
# --- Network Design ---
architecture: Literal["dense", "convolutional"] = "dense"
encode_dims: tuple[PositiveInt, ...] = (256, 128)
decode_dims: tuple[PositiveInt, ...] = ()
@field_validator("encode_dims", mode="before")
@classmethod
def validate_encode_dims(
cls, value: tuple[PositiveInt, ...]
) -> tuple[PositiveInt, ...]:
if len(value) == 0:
err = "`encode_dims` can not be empty"
cls._raise(err)
if not all(x > y for x, y in itertools.pairwise(value)):
err = f"`encode_dims`: {value} is not monotonically decreasing"
cls._raise(err)
return value
@model_validator(mode="after")
def validate_decode_dims(self) -> Self:
if len(self.decode_dims) == 0:
self.decode_dims = tuple(reversed(self.encode_dims))
if not all(x < y for x, y in itertools.pairwise(self.decode_dims)):
err = f"`decode_dims`: {self.decode_dims} is not monotonically decreasing"
self._raise(err)
return self
physical_latents: bool
n_z_latents: PositiveInt = 3
@model_validator(mode="after")
def validate_n_latents(self) -> Self:
if not self.physical_latents and self.n_z_latents == 0:
err = "You must specify either non-zero `n_z_latents`, or `physical_latents=True`. With both `physical_latents=False` and `n_z_latents=0, there will be no latents to train at all!"
self._raise(err)
return self
# --- Training ---
# Overfitting
batch_normalisation: bool = False
dropout: Annotated[float, Field(ge=0, le=1)] = 0
# Latent training
seperate_latent_training: bool
seperate_z_latent_training: bool
# === Optional ===
seed: int = 12345
validation_frac: Annotated[float, Field(ge=0, le=1)] = 0
batch_size: PositiveInt
save_best: bool = False
patience: float | int = 0.1
# --- Data ---
min_train_redshift: float = 0.02
max_train_redshift: float = 1.0
min_test_redshift: float = 0.02
max_test_redshift: float = 1.0
min_val_redshift: float = 0.02
max_val_redshift: float = 0.1
min_redshift: float = 0.02
max_redshift: float = 1.0
min_train_phase: float = -10
max_train_phase: float = 40
min_test_phase: float = -10
max_test_phase: float = 40
min_val_phase: float = -10
max_val_phase: float = 40
min_phase: float = -10
max_phase: float = 40
# --- Data Offsets ---
phase_offset_scale: float = -0.02
amplitude_offset_scale: NonNegativeFloat = 1.0
mask_fraction: Annotated[float, Field(ge=0, le=1)] = 0.1
# --- Loss ---
loss_residual_penalty: NonNegativeFloat = 0
loss_delta_av_penalty: NonNegativeFloat = 0
loss_delta_m_penalty: NonNegativeFloat = 0
loss_delta_p_penalty: NonNegativeFloat = 0
loss_covariance_penalty: NonNegativeFloat = 50000
loss_decorrelate_all: bool = True
loss_decorrelate_dust: bool = True
loss_clip_delta: PositiveFloat = 25
# --- Stages ---
# ΔAᵥ
delta_av_epochs: PositiveInt = 1000
delta_av_lr: PositiveFloat = 0.005
delta_av_lr_decay_steps: PositiveInt = 300
delta_av_lr_decay_rate: PositiveFloat = 0.95
delta_av_lr_weight_decay_rate: PositiveFloat = 0.0001
# Zs
zs_epochs: PositiveInt = 1000
zs_lr: PositiveFloat = 0.005
zs_lr_decay_steps: PositiveInt = 300
zs_lr_decay_rate: PositiveFloat = 0.95
zs_lr_weight_decay_rate: PositiveFloat = 0.0001
# Δℳ
delta_m_epochs: PositiveInt = 5000
delta_m_lr: PositiveFloat = 0.005
delta_m_lr_decay_steps: PositiveInt = 300
delta_m_lr_decay_rate: PositiveFloat = 0.95
delta_m_lr_weight_decay_rate: PositiveFloat = 0.0001
# Δ𝓅
delta_p_epochs: PositiveInt = 5000
delta_p_lr: PositiveFloat = 0.001
delta_p_lr_decay_steps: PositiveInt = 300
delta_p_lr_decay_rate: PositiveFloat = 0.95
delta_p_lr_weight_decay_rate: PositiveFloat = 0.0001
# Final
final_epochs: PositiveFloat = 5000
final_lr: PositiveFloat = 0.001
final_lr_decay_steps: PositiveInt = 300
final_lr_decay_rate: PositiveFloat = 0.95
final_lr_weight_decay_rate: PositiveFloat = 0.0001
|