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 | import os
from typing import Any, Concatenate, cast, override
from functools import cached_property
from collections.abc import Callable
from pydantic import PositiveFloat, computed_field
os.environ["TF_USE_LEGACY_KERAS"] = "1"
os.environ["KERAS_BACKEND"] = "tensorflow"
os.environ["TF_DETERMINISTIC_OPS"] = "1"
os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"
import tensorflow as tf
from tensorflow import keras as ks
from supaernova.steps.pae.tf import (
loss as snpae_losses,
)
from supaernova.configs.steps import ConfigInputObject, validate_object
from .model import PAEModelConfig
ActivationObject = Callable[[tf.Tensor], tf.Tensor]
RegulariserObject = type[ks.regularizers.Regularizer] | Callable[[tf.Tensor], tf.Tensor]
SchedulerObject = (
type[ks.optimizers.schedules.LearningRateSchedule]
| Callable[[Concatenate[int | tf.Tensor, ...]], tf.Tensor]
)
OptimiserObject = type[ks.optimizers.Optimizer]
LossObject = type[ks.losses.Loss] | Callable[[tf.Tensor, tf.Tensor], tf.Tensor]
def validate_activation(activation: ConfigInputObject[ActivationObject]):
return validate_object(activation, dummy_obj=tf.nn.relu, mod=tf.nn)
def validate_kernel_regulariser(
kernel_regulariser: ConfigInputObject[RegulariserObject],
) -> RegulariserObject:
return validate_object(
kernel_regulariser, dummy_obj=ks.regularizers.Regularizer, mod=ks.regularizers
)
def validate_scheduler(
scheduler: ConfigInputObject[SchedulerObject],
) -> SchedulerObject:
return validate_object(
scheduler,
dummy_obj=ks.optimizers.schedules.LearningRateSchedule,
mod=ks.optimizers.schedules,
)
def validate_optimiser(
optimiser: ConfigInputObject[OptimiserObject],
):
return validate_object(
optimiser, dummy_obj=ks.optimizers.Optimizer, mod=ks.optimizers
)
def validate_loss(
loss: ConfigInputObject[LossObject],
):
err = f"Could not validate loss: {loss}:\n"
for dummy_obj in (ks.losses.Loss, ks.losses.mae):
for mod in (ks.losses, snpae_losses):
try:
return validate_object(loss, dummy_obj=dummy_obj, mod=mod)
except ValueError as e:
err += f"{e}\n"
raise ValueError(err)
def get_loss(
loss_fn: Callable[[tf.Tensor, tf.Tensor], tf.Tensor],
) -> type[ks.losses.Loss]:
@ks.utils.register_keras_serializable("SuPAErnova")
class CustomLoss(ks.losses.Loss):
@override
def call(self, y_true: tf.Tensor, y_pred: tf.Tensor) -> tf.Tensor:
return loss_fn(y_true, y_pred, model=self.model)
return CustomLoss
class TFPAEModelConfig(PAEModelConfig):
# --- Training ---
activation: ConfigInputObject[ActivationObject]
@computed_field
@cached_property
def activation_fn(self) -> ActivationObject:
return validate_activation(self.activation)
kernel_regulariser: ConfigInputObject[RegulariserObject] | None = None
kernel_regulariser_penalty: PositiveFloat | None = None
@computed_field
@cached_property
def kernel_regulariser_cls(self) -> type[ks.regularizers.Regularizer] | None:
if self.kernel_regulariser is None:
return None
regulariser = validate_kernel_regulariser(self.kernel_regulariser)
if isinstance(regulariser, type):
return regulariser
class CustomRegulariser(ks.regularizers.Regularizer):
@override
def __init__(self, *args: Any, **kwargs: Any) -> None:
super().__init__(*args, **kwargs)
@override
def __call__(self, x: tf.Tensor) -> tf.Tensor:
return regulariser(x)
return CustomRegulariser
scheduler: ConfigInputObject[SchedulerObject]
@computed_field
@cached_property
def scheduler_cls(self) -> type[ks.optimizers.schedules.LearningRateSchedule]:
scheduler = validate_scheduler(self.scheduler)
if isinstance(scheduler, type):
return scheduler
class CustomScheduler(ks.optimizers.schedules.LearningRateSchedule):
@override
def __init__(
self,
*,
initial_learning_rate: float,
decay_steps: int,
decay_rate: float,
) -> None:
self.initial_learning_rate: float = initial_learning_rate
self.decay_steps: int = decay_steps
self.decay_rate: float = decay_rate
@override
def __call__(self, step: int | tf.Tensor) -> tf.Tensor:
return scheduler(
step,
initial_learning_rate=self.initial_learning_rate,
decay_steps=self.decay_steps,
decay_rate=self.decay_rate,
)
return CustomScheduler
optimiser: ConfigInputObject[OptimiserObject]
@computed_field
@cached_property
def optimiser_cls(self) -> type[ks.optimizers.Optimizer]:
return cast(
"type[ks.optimizers.Optimizer]",
cast("object", validate_optimiser(self.optimiser)),
)
loss: ConfigInputObject[LossObject]
@computed_field
@cached_property
def loss_cls(self) -> type[ks.losses.Loss]:
loss = validate_loss(self.loss)
if isinstance(loss, type):
loss = loss()
return get_loss(loss)
|