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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
|
import haiku as hk
import einx
from functools import partial
from haiku._src.base import current_module
from typing import Any, Callable, Literal, Optional
thk = einx.tracer.import_("haiku", "hk")
class ParamFactory:
class Concrete(einx.tracer.input.Input):
def __init__(self, name, init, dtype, param_type):
self.name = name
self.init = init
if dtype is None:
module = current_module()
if hasattr(module, "dtype"):
dtype = module.dtype
else:
dtype = "float32"
self.dtype = dtype
self.param_type = param_type
def to_value_and_key(self):
return None, ParamFactory.CacheKey(self.name, self.init, self.dtype, self.param_type)
class CacheKey(einx.tracer.input.CacheKey):
def __init__(self, name, init, dtype, param_type):
self.name = name
self.init = init
self.dtype = dtype
self.param_type = param_type
def __hash__(self):
return hash((self.name, self.init, self.dtype, self.param_type))
def __eq__(self, other):
return (
isinstance(other, ParamFactory.CacheKey)
and self.name == other.name
and self.init == other.init
and self.dtype == other.dtype
and self.param_type == other.param_type
)
def to_tracer(self, backend, virtual_arg):
return (
None,
ParamFactory.Tracer(self.name, self.init, self.dtype, self.param_type, virtual_arg),
)
class Tracer(einx.tracer.TensorFactory):
def __init__(self, name, init, dtype, param_type, depend_on):
self.name = name
self.init = init
self.dtype = dtype
self.param_type = param_type
self.depend_on = depend_on
def __call__(self, shape, kwargs):
name = self.name if not self.name is None else kwargs.get("name", None)
init = self.init if not self.init is None else kwargs.get("init", None)
dtype = self.dtype if not self.dtype is None else kwargs.get("dtype", None)
if name is None:
raise ValueError("Must specify name for tensor factory hk.get_{parameter|state}")
if init is None:
raise ValueError("Must specify init for tensor factory hk.get_{parameter|state}")
elif isinstance(init, str):
if init in "get_at" or init == "rearrange":
init = thk.initializers.RandomNormal(stddev=0.02)
elif init == "add":
init = thk.initializers.Constant(0.0)
elif init == "multiply":
init = thk.initializers.Constant(1.0)
elif init == "dot":
init = thk.initializers.VarianceScaling(
1.0, "fan_in", "truncated_normal", fan_in_axes=kwargs["in_axis"]
)
else:
raise ValueError(f"Don't know which initializer to use for operation '{init}'")
elif isinstance(init, (int, float)):
init = thk.initializers.Constant(init)
if self.param_type == "parameter":
func = thk.get_parameter
elif self.param_type == "state":
func = thk.get_state
else:
assert False
return einx.tracer.apply(
func,
kwargs={"shape": shape, "name": name, "dtype": dtype, "init": init},
output=einx.tracer.Tensor(shape),
depend_on=[self.depend_on],
)
def param(
func: Literal[hk.get_parameter, hk.get_state] = hk.get_parameter,
name: Optional[str] = None,
init: Optional[Any] = None,
dtype: Optional[Any] = None,
):
"""Create a tensor factory for Haiku parameters.
Args:
func: Either ``hk.get_parameter`` or ``hk.get_state``. Defaults to ``hk.get_parameter``.
name: Name of the parameter. If ``None``, uses a default name determined from the calling
operation. Defaults to ``None``.
init: Initializer for the parameter. If ``None``, uses a default init method determined
from the calling operation. Defaults to ``None``.
dtype: Data type of the parameter. If ``None``, uses the ``dtype`` member of the calling
module or ``float32`` if it does not exist. Defaults to ``None``.
Returns:
A tensor factory with the given default parameters.
"""
if func == hk.get_parameter:
param_type = "parameter"
elif func == hk.get_state:
param_type = "state"
else:
raise ValueError(f"Unknown parameter function '{func}'")
return ParamFactory.Concrete(name, init, dtype, param_type)
# Allow passing hk.get_parameter and hk.get_state as tensor factory:
@einx.tracer.input.register_tensor_factory
def tensor_factory(x):
if id(x) == id(hk.get_parameter) or id(x) == id(hk.get_state):
return param(x).to_value_and_key()
else:
return None
class Norm(hk.Module):
"""Normalization layer.
Args:
stats: Einstein string determining the axes along which mean and variance are computed.
Will be passed to ``einx.reduce``.
params: Einstein string determining the axes along which learnable parameters are
applied. Will be passed to ``einx.elementwise``. Defaults to ``"b... [c]"``.
mean: Whether to apply mean normalization. Defaults to ``True``.
var: Whether to apply variance normalization. Defaults to ``True``.
scale: Whether to apply a learnable scale according to ``params``. Defaults to ``True``.
bias: Whether to apply a learnable bias according to ``params``. Defaults to ``True``.
epsilon: A small float added to the variance to avoid division by zero. Defaults
to ``1e-5``.
fastvar: Whether to use a fast variance computation. Defaults to ``True``.
dtype: Data type of the weights. Defaults to ``"float32"``.
decay_rate: Decay rate for exponential moving average of mean and variance. If ``None``,
no moving average is applied. Defaults to ``None``.
name: Name of the module. Defaults to ``None``.
**kwargs: Additional parameters that specify values for single axes, e.g. ``a=4``.
"""
def __init__(
self,
stats: str,
params: str = "b... [c]",
mean: bool = True,
var: bool = True,
scale: bool = True,
bias: bool = True,
epsilon: float = 1e-5,
fastvar: bool = True,
dtype: Any = "float32",
decay_rate: Optional[float] = None,
name: Optional[str] = None,
**kwargs: Any,
):
super().__init__(name=name)
self.stats = stats
self.params = params
self.mean = mean
self.var = var
self.scale = scale
self.bias = bias
self.epsilon = epsilon
self.fastvar = fastvar
self.dtype = dtype
self.decay_rate = decay_rate
self.kwargs = kwargs
def __call__(self, x, training=None):
if self.decay_rate is not None and training is None:
raise ValueError("training must be specified when decay_rate is used")
use_ema = self.decay_rate is not None and (not training or hk.running_init())
x, mean, var = einx.nn.norm(
x,
self.stats,
self.params,
mean=param(hk.get_state, name="mean") if use_ema and self.mean else self.mean,
var=param(hk.get_state, name="var") if use_ema and self.var else self.var,
scale=param(hk.get_parameter, name="scale") if self.scale else None,
bias=param(hk.get_parameter, name="bias") if self.bias else None,
epsilon=self.epsilon,
fastvar=self.fastvar,
**self.kwargs,
)
update_ema = self.decay_rate is not None and training and not hk.running_init()
if update_ema:
if self.mean:
hk.set_state(
"mean", hk.get_state("mean") * self.decay_rate + mean * (1 - self.decay_rate)
)
if self.var:
hk.set_state(
"var", hk.get_state("var") * self.decay_rate + var * (1 - self.decay_rate)
)
return x
class Linear(hk.Module):
"""Linear layer.
Args:
expr: Einstein string determining the axes along which the weight matrix is multiplied.
Will be passed to ``einx.dot``.
bias: Whether to apply a learnable bias. Defaults to ``True``.
dtype: Data type of the weights. Defaults to ``"float32"``.
name: Name of the module. Defaults to ``None``.
**kwargs: Additional parameters that specify values for single axes, e.g. ``a=4``.
"""
def __init__(
self,
expr: str,
bias: bool = True,
dtype: Any = "float32",
name: Optional[str] = None,
**kwargs: Any,
):
super().__init__(name=name)
self.expr = expr
self.bias = bias
self.dtype = dtype
self.kwargs = kwargs
def __call__(self, x):
return einx.nn.linear(
x,
self.expr,
bias=param(hk.get_parameter, name="bias") if self.bias else None,
weight=param(hk.get_parameter, name="weight"),
**self.kwargs,
)
class Dropout(hk.Module):
"""Dropout layer.
Args:
expr: Einstein string determining the axes along which dropout is applied. Will be
passed to ``einx.elementwise``.
drop_rate: Drop rate.
name: Name of the module. Defaults to ``None``.
**kwargs: Additional parameters that specify values for single axes, e.g. ``a=4``.
"""
def __init__(self, expr: str, drop_rate: float, name: Optional[str] = None, **kwargs: Any):
super().__init__(name=name)
self.expr = expr
self.drop_rate = drop_rate
self.kwargs = kwargs
def __call__(self, x, training):
if training:
return einx.nn.dropout(
x,
self.expr,
drop_rate=self.drop_rate,
rng=hk.next_rng_key(),
**self.kwargs,
)
else:
return x
|
