-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
182 lines (155 loc) · 6.45 KB
/
utils.py
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
# AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/00_utils.ipynb.
# %% ../nbs/00_utils.ipynb 3
from __future__ import annotations
from .import_essentials import *
import nbdev
from fastcore.basics import AttrDict
from nbdev.showdoc import BasicMarkdownRenderer
from inspect import isclass
from fastcore.test import *
from jax.core import InconclusiveDimensionOperation
# %% auto 0
__all__ = ['validate_configs', 'save_pytree', 'load_pytree', 'auto_reshaping', 'grad_update', 'gumbel_softmax', 'load_json',
'get_config', 'set_config']
# %% ../nbs/00_utils.ipynb 5
def validate_configs(
configs: dict | BaseParser, # A configuration of the model/dataset.
config_cls: BaseParser, # The desired configuration class.
) -> BaseParser:
"""return a valid configuration object."""
assert isclass(config_cls), f"`config_cls` should be a class."
assert issubclass(config_cls, BaseParser), \
f"{config_cls} should be a subclass of `BaseParser`."
if isinstance(configs, dict):
configs = config_cls(**configs)
if not isinstance(configs, config_cls):
raise TypeError(
f"configs should be either a `dict` or an instance of {config_cls.__name__}.")
return configs
# %% ../nbs/00_utils.ipynb 15
def _is_array(x):
return isinstance(x, np.ndarray) or isinstance(x, jnp.ndarray) or isinstance(x, list)
def save_pytree(pytree, saved_dir):
"""Save a pytree to a directory."""
with open(os.path.join(saved_dir, "data.npy"), "wb") as f:
for x in jax.tree_util.tree_leaves(pytree):
np.save(f, x)
tree_struct = jax.tree_util.tree_map(lambda t: _is_array(t), pytree)
with open(os.path.join(saved_dir, "treedef.json"), "w") as f:
json.dump(tree_struct, f)
# %% ../nbs/00_utils.ipynb 20
def load_pytree(saved_dir):
"""Load a pytree from a saved directory."""
with open(os.path.join(saved_dir, "treedef.json"), "r") as f:
tree_struct = json.load(f)
leaves, treedef = jax.tree_util.tree_flatten(tree_struct)
with open(os.path.join(saved_dir, "data.npy"), "rb") as f:
flat_state = [
np.load(f, allow_pickle=True) if is_arr else np.load(f, allow_pickle=True).item()
for is_arr in leaves
]
return jax.tree_util.tree_unflatten(treedef, flat_state)
# %% ../nbs/00_utils.ipynb 24
def _reshape_x(x: Array):
x_size = x.shape
if len(x_size) > 1 and x_size[0] != 1:
raise ValueError(
f"""Invalid Input Shape: Require `x.shape` = (1, k) or (k, ),
but got `x.shape` = {x.shape}. This method expects a single input instance."""
)
if len(x_size) == 1:
x = x.reshape(1, -1)
return x, x_size
# %% ../nbs/00_utils.ipynb 25
def auto_reshaping(
reshape_argname: str, # The name of the argument to be reshaped.
reshape_output: bool = True, # Whether to reshape the output. Useful to set `False` when returning multiple cfs.
):
"""
Decorator to automatically reshape function's input into (1, k),
and out to input's shape.
"""
def decorator(func):
def wrapper(*args, **kwargs):
kwargs = inspect.getcallargs(func, *args, **kwargs)
if reshape_argname in kwargs:
reshaped_x, x_shape = _reshape_x(kwargs[reshape_argname])
kwargs[reshape_argname] = reshaped_x
else:
raise ValueError(
f"Invalid argument name: `{reshape_argname}` is not a valid argument name.")
# Call the function.
cf = func(**kwargs)
if not isinstance(cf, Array):
raise ValueError(
f"Invalid return type: must be a `jax.Array`, but got `{type(cf).__name__}`.")
if reshape_output:
try:
cf = cf.reshape(x_shape)
except (InconclusiveDimensionOperation, TypeError) as e:
raise ValueError(
f"Invalid return shape: Require `cf.shape` = {cf.shape} "
f"is not compatible with `x.shape` = {x_shape}.")
return cf
return wrapper
return decorator
# %% ../nbs/00_utils.ipynb 30
def grad_update(
grads, # A pytree of gradients.
params, # A pytree of parameters.
opt_state: optax.OptState,
opt: optax.GradientTransformation,
): # Return (upt_params, upt_opt_state)
updates, opt_state = opt.update(grads, opt_state, params)
upt_params = optax.apply_updates(params, updates)
return upt_params, opt_state
# %% ../nbs/00_utils.ipynb 32
def gumbel_softmax(
key: jrand.PRNGKey, # Random key
logits: Array, # Logits for each class. Shape (batch_size, num_classes)
tau: float, # Temperature for the Gumbel softmax
axis: int | tuple[int, ...] = -1, # The axis or axes along which the gumbel softmax should be computed
):
"""The Gumbel softmax function."""
gumbel_noise = jrand.gumbel(key, shape=logits.shape)
y = logits + gumbel_noise
return jax.nn.softmax(y / tau, axis=axis)
# %% ../nbs/00_utils.ipynb 34
def load_json(f_name: str) -> Dict[str, Any]: # file name
with open(f_name) as f:
return json.load(f)
# %% ../nbs/00_utils.ipynb 36
@dataclass
class Config:
rng_reserve_size: int
global_seed: int
@classmethod
def default(cls) -> Config:
return cls(rng_reserve_size=1, global_seed=42)
main_config = Config.default()
# %% ../nbs/00_utils.ipynb 37
def get_config() -> Config:
return main_config
# %% ../nbs/00_utils.ipynb 38
def set_config(
*,
rng_reserve_size: int = None, # The number of random number generators to reserve.
global_seed: int = None, # The global seed for random number generators.
**kwargs
) -> None:
"""Sets the global configurations."""
def set_val(
arg_name: str, # The name of the argument.
arg_value: int, # The value of the argument.
arg_min: int # The minimum value of the argument.
) -> None:
"""Checks the validity of the argument and sets the value."""
if arg_value is None or not hasattr(main_config, arg_name):
return
if not isinstance(arg_value, int):
raise TypeError(f"`{arg_name}` must be an integer, but got {type(arg_value).__name__}.")
if arg_value < arg_min:
raise ValueError(f"`{arg_name}` must be non-negative, but got {arg_value}.")
setattr(main_config, arg_name, arg_value)
set_val('rng_reserve_size', rng_reserve_size, 1)
set_val('global_seed', global_seed, 0)