/
utils.py
488 lines (399 loc) · 16 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
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
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
import numbers
import numpy as np
import theano
import theano.tensor as T
#: Tuple of ``int``-like types for ``isinstance`` checks.
#: Specifically includes long integers and numpy integers.
int_types = (numbers.Integral, np.integer)
def floatX(arr):
"""Converts data to a numpy array of dtype ``theano.config.floatX``.
Parameters
----------
arr : array_like
The data to be converted.
Returns
-------
numpy ndarray
The input array in the ``floatX`` dtype configured for Theano.
If `arr` is an ndarray of correct dtype, it is returned as is.
"""
return np.asarray(arr, dtype=theano.config.floatX)
def shared_empty(dim=2, dtype=None):
"""Creates empty Theano shared variable.
Shortcut to create an empty Theano shared variable with
the specified number of dimensions.
Parameters
----------
dim : int, optional
The number of dimensions for the empty variable, defaults to 2.
dtype : a numpy data-type, optional
The desired dtype for the variable. Defaults to the Theano
``floatX`` dtype.
Returns
-------
Theano shared variable
An empty Theano shared variable of dtype ``dtype`` with
`dim` dimensions.
"""
if dtype is None:
dtype = theano.config.floatX
shp = tuple([1] * dim)
return theano.shared(np.zeros(shp, dtype=dtype))
def as_theano_expression(input):
"""Wrap as Theano expression.
Wraps the given input as a Theano constant if it is not
a valid Theano expression already. Useful to transparently
handle numpy arrays and Python scalars, for example.
Parameters
----------
input : number, numpy array or Theano expression
Expression to be converted to a Theano constant.
Returns
-------
Theano symbolic constant
Theano constant version of `input`.
"""
if isinstance(input, theano.gof.Variable):
return input
else:
try:
return theano.tensor.constant(input)
except Exception as e:
raise TypeError("Input of type %s is not a Theano expression and "
"cannot be wrapped as a Theano constant (original "
"exception: %s)" % (type(input), e))
def collect_shared_vars(expressions):
"""Returns all shared variables the given expression(s) depend on.
Parameters
----------
expressions : Theano expression or iterable of Theano expressions
The expressions to collect shared variables from.
Returns
-------
list of Theano shared variables
All shared variables the given expression(s) depend on, in fixed order
(as found by a left-recursive depth-first search). If some expressions
are shared variables themselves, they are included in the result.
"""
# wrap single expression in list
if isinstance(expressions, theano.Variable):
expressions = [expressions]
# return list of all shared variables
return [v for v in theano.gof.graph.inputs(reversed(expressions))
if isinstance(v, theano.compile.SharedVariable)]
def one_hot(x, m=None):
"""One-hot representation of integer vector.
Given a vector of integers from 0 to m-1, returns a matrix
with a one-hot representation, where each row corresponds
to an element of x.
Parameters
----------
x : integer vector
The integer vector to convert to a one-hot representation.
m : int, optional
The number of different columns for the one-hot representation. This
needs to be strictly greater than the maximum value of `x`.
Defaults to ``max(x) + 1``.
Returns
-------
Theano tensor variable
A Theano tensor variable of shape (``n``, `m`), where ``n`` is the
length of `x`, with the one-hot representation of `x`.
Notes
-----
If your integer vector represents target class memberships, and you wish to
compute the cross-entropy between predictions and the target class
memberships, then there is no need to use this function, since the function
:func:`lasagne.objectives.categorical_crossentropy()` can compute the
cross-entropy from the integer vector directly.
"""
if m is None:
m = T.cast(T.max(x) + 1, 'int32')
return T.eye(m)[T.cast(x, 'int32')]
def unique(l):
"""Filters duplicates of iterable.
Create a new list from l with duplicate entries removed,
while preserving the original order.
Parameters
----------
l : iterable
Input iterable to filter of duplicates.
Returns
-------
list
A list of elements of `l` without duplicates and in the same order.
"""
new_list = []
seen = set()
for el in l:
if el not in seen:
new_list.append(el)
seen.add(el)
return new_list
def as_tuple(x, N, t=None):
"""
Coerce a value to a tuple of given length (and possibly given type).
Parameters
----------
x : value or iterable
N : integer
length of the desired tuple
t : type or tuple of type, optional
required type or types for all elements
Returns
-------
tuple
``tuple(x)`` if `x` is iterable, ``(x,) * N`` otherwise.
Raises
------
TypeError
if `type` is given and `x` or any of its elements do not match it
ValueError
if `x` is iterable, but does not have exactly `N` elements
"""
try:
X = tuple(x)
except TypeError:
X = (x,) * N
if (t is not None) and not all(isinstance(v, t) for v in X):
if t == int_types:
expected_type = "int" # easier to understand
elif isinstance(t, tuple):
expected_type = " or ".join(tt.__name__ for tt in t)
else:
expected_type = t.__name__
raise TypeError("expected a single value or an iterable "
"of {0}, got {1} instead".format(expected_type, x))
if len(X) != N:
raise ValueError("expected a single value or an iterable "
"with length {0}, got {1} instead".format(N, x))
return X
def inspect_kwargs(func):
"""
Inspects a callable and returns a list of all optional keyword arguments.
Parameters
----------
func : callable
The callable to inspect
Returns
-------
kwargs : list of str
Names of all arguments of `func` that have a default value, in order
"""
# We try the Python 3.x way first, then fall back to the Python 2.x way
try:
from inspect import signature
except ImportError: # pragma: no cover
from inspect import getargspec
spec = getargspec(func)
return spec.args[-len(spec.defaults):] if spec.defaults else []
else: # pragma: no cover
params = signature(func).parameters
return [p.name for p in params.values() if p.default is not p.empty]
def compute_norms(array, norm_axes=None):
""" Compute incoming weight vector norms.
Parameters
----------
array : numpy array or Theano expression
Weight or bias.
norm_axes : sequence (list or tuple)
The axes over which to compute the norm. This overrides the
default norm axes defined for the number of dimensions
in `array`. When this is not specified and `array` is a 2D array,
this is set to `(0,)`. If `array` is a 3D, 4D or 5D array, it is
set to a tuple listing all axes but axis 0. The former default is
useful for working with dense layers, the latter is useful for 1D,
2D and 3D convolutional layers.
Finally, in case `array` is a vector, `norm_axes` is set to an empty
tuple, and this function will simply return the absolute value for
each element. This is useful when the function is applied to all
parameters of the network, including the bias, without distinction.
(Optional)
Returns
-------
norms : 1D array or Theano vector (1D)
1D array or Theano vector of incoming weight/bias vector norms.
Examples
--------
>>> array = np.random.randn(100, 200)
>>> norms = compute_norms(array)
>>> norms.shape
(200,)
>>> norms = compute_norms(array, norm_axes=(1,))
>>> norms.shape
(100,)
"""
# Check if supported type
if not isinstance(array, theano.Variable) and \
not isinstance(array, np.ndarray):
raise RuntimeError(
"Unsupported type {}. "
"Only theano variables and numpy arrays "
"are supported".format(type(array))
)
# Compute default axes to sum over
ndim = array.ndim
if norm_axes is not None:
sum_over = tuple(norm_axes)
elif ndim == 1: # For Biases that are in 1d (e.g. b of DenseLayer)
sum_over = ()
elif ndim == 2: # DenseLayer
sum_over = (0,)
elif ndim in [3, 4, 5]: # Conv{1,2,3}DLayer
sum_over = tuple(range(1, ndim))
else:
raise ValueError(
"Unsupported tensor dimensionality {}. "
"Must specify `norm_axes`".format(array.ndim)
)
# Run numpy or Theano norm computation
if isinstance(array, theano.Variable):
# Apply theano version if it is a theano variable
if len(sum_over) == 0:
norms = T.abs_(array) # abs if we have nothing to sum over
else:
norms = T.sqrt(T.sum(array**2, axis=sum_over))
elif isinstance(array, np.ndarray):
# Apply the numpy version if ndarray
if len(sum_over) == 0:
norms = abs(array) # abs if we have nothing to sum over
else:
norms = np.sqrt(np.sum(array**2, axis=sum_over))
return norms
def create_param(spec, shape, name=None):
"""
Helper method to create Theano shared variables for layer parameters
and to initialize them.
Parameters
----------
spec : scalar number, numpy array, Theano expression, or callable
Either of the following:
* a scalar or a numpy array with the initial parameter values
* a Theano expression or shared variable representing the parameters
* a function or callable that takes the desired shape of
the parameter array as its single argument and returns
a numpy array, a Theano expression, or a shared variable
representing the parameters.
shape : iterable of int
a tuple or other iterable of integers representing the desired
shape of the parameter array.
name : string, optional
The name to give to the parameter variable. Ignored if `spec`
is or returns a Theano expression or shared variable that
already has a name.
Returns
-------
Theano shared variable or Theano expression
A Theano shared variable or expression representing layer parameters.
If a scalar or a numpy array was provided, a shared variable is
initialized to contain this array. If a shared variable or expression
was provided, it is simply returned. If a callable was provided, it is
called, and its output is used to initialize a shared variable.
Notes
-----
This function is called by :meth:`Layer.add_param()` in the constructor
of most :class:`Layer` subclasses. This enables those layers to
support initialization with scalars, numpy arrays, existing Theano shared
variables or expressions, and callables for generating initial parameter
values, Theano expressions, or shared variables.
"""
import numbers # to check if argument is a number
shape = tuple(shape) # convert to tuple if needed
if any(d <= 0 for d in shape):
raise ValueError((
"Cannot create param with a non-positive shape dimension. "
"Tried to create param with shape=%r, name=%r") % (shape, name))
err_prefix = "cannot initialize parameter %s: " % name
if callable(spec):
spec = spec(shape)
err_prefix += "the %s returned by the provided callable"
else:
err_prefix += "the provided %s"
if isinstance(spec, numbers.Number) or isinstance(spec, np.generic) \
and spec.dtype.kind in 'biufc':
spec = np.asarray(spec)
if isinstance(spec, np.ndarray):
if spec.shape != shape:
raise ValueError("%s has shape %s, should be %s" %
(err_prefix % "numpy array", spec.shape, shape))
# We assume parameter variables do not change shape after creation.
# We can thus fix their broadcast pattern, to allow Theano to infer
# broadcastable dimensions of expressions involving these parameters.
bcast = tuple(s == 1 for s in shape)
spec = theano.shared(spec, broadcastable=bcast)
if isinstance(spec, theano.Variable):
# We cannot check the shape here, Theano expressions (even shared
# variables) do not have a fixed compile-time shape. We can check the
# dimensionality though.
if spec.ndim != len(shape):
raise ValueError("%s has %d dimensions, should be %d" %
(err_prefix % "Theano variable", spec.ndim,
len(shape)))
# We only assign a name if the user hasn't done so already.
if not spec.name:
spec.name = name
return spec
else:
if "callable" in err_prefix:
raise TypeError("%s is not a numpy array or a Theano expression" %
(err_prefix % "value"))
else:
raise TypeError("%s is not a numpy array, a Theano expression, "
"or a callable" % (err_prefix % "spec"))
def unroll_scan(fn, sequences, outputs_info, non_sequences, n_steps,
go_backwards=False):
"""
Helper function to unroll for loops. Can be used to unroll theano.scan.
The parameter names are identical to theano.scan, please refer to here
for more information.
Note that this function does not support the truncate_gradient
setting from theano.scan.
Parameters
----------
fn : function
Function that defines calculations at each step.
sequences : TensorVariable or list of TensorVariables
List of TensorVariable with sequence data. The function iterates
over the first dimension of each TensorVariable.
outputs_info : list of TensorVariables
List of tensors specifying the initial values for each recurrent
value.
non_sequences: list of TensorVariables
List of theano.shared variables that are used in the step function.
n_steps: int
Number of steps to unroll.
go_backwards: bool
If true the recursion starts at sequences[-1] and iterates
backwards.
Returns
-------
List of TensorVariables. Each element in the list gives the recurrent
values at each time step.
"""
if not isinstance(sequences, (list, tuple)):
sequences = [sequences]
# When backwards reverse the recursion direction
counter = range(n_steps)
if go_backwards:
counter = counter[::-1]
output = []
prev_vals = outputs_info
for i in counter:
step_input = [s[i] for s in sequences] + prev_vals + non_sequences
out_ = fn(*step_input)
# The returned values from step can be either a TensorVariable,
# a list, or a tuple. Below, we force it to always be a list.
if isinstance(out_, T.TensorVariable):
out_ = [out_]
if isinstance(out_, tuple):
out_ = list(out_)
output.append(out_)
prev_vals = output[-1]
# iterate over each scan output and convert it to same format as scan:
# [[output11, output12,...output1n],
# [output21, output22,...output2n],...]
output_scan = []
for i in range(len(output[0])):
l = map(lambda x: x[i], output)
output_scan.append(T.stack(*l))
return output_scan