-
Notifications
You must be signed in to change notification settings - Fork 2.8k
/
jet_test.py
436 lines (376 loc) · 16.6 KB
/
jet_test.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
# Copyright 2020 The JAX Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from functools import reduce, partial
from absl.testing import absltest
import numpy as np
import unittest
import jax
from jax._src import test_util as jtu
import jax.numpy as jnp
import jax.scipy.special
from jax import random
from jax import jacfwd, jit
from jax.example_libraries import stax
from jax.experimental.jet import jet, fact, zero_series
from jax import lax
jax.config.parse_flags_with_absl()
def jvp_taylor(fun, primals, series):
# Computes the Taylor series the slow way, with nested jvp.
order, = set(map(len, series))
primals = tuple(jnp.asarray(p) for p in primals)
def composition(eps):
taylor_terms = [sum(eps ** (i+1) * terms[i] / fact(i + 1)
for i in range(len(terms))) for terms in series]
nudged_args = [(x + t).astype(x.dtype) for x, t in zip(primals, taylor_terms)]
return fun(*nudged_args)
primal_out = fun(*primals)
terms_out = [repeated(jacfwd, i+1)(composition)(0.) for i in range(order)]
return primal_out, terms_out
def repeated(f, n):
def rfun(p):
return reduce(lambda x, _: f(x), range(n), p)
return rfun
def transform(lims, x):
return x * (lims[1] - lims[0]) + lims[0]
class JetTest(jtu.JaxTestCase):
def check_jet(self, fun, primals, series, atol=1e-5, rtol=1e-5,
check_dtypes=True):
# Convert to jax arrays to ensure dtype canonicalization.
primals = jax.tree.map(jnp.asarray, primals)
series = jax.tree.map(jnp.asarray, series)
y, terms = jet(fun, primals, series)
expected_y, expected_terms = jvp_taylor(fun, primals, series)
self.assertAllClose(y, expected_y, atol=atol, rtol=rtol,
check_dtypes=check_dtypes)
self.assertAllClose(terms, expected_terms, atol=atol, rtol=rtol,
check_dtypes=check_dtypes)
def check_jet_finite(self, fun, primals, series, atol=1e-5, rtol=1e-5,
check_dtypes=True):
# Convert to jax arrays to ensure dtype canonicalization.
primals = jax.tree.map(jnp.asarray, primals)
series = jax.tree.map(jnp.asarray, series)
y, terms = jet(fun, primals, series)
expected_y, expected_terms = jvp_taylor(fun, primals, series)
def _convert(x):
return jnp.where(jnp.isfinite(x), x, jnp.nan)
y = _convert(y)
expected_y = _convert(expected_y)
terms = _convert(jnp.asarray(terms))
expected_terms = _convert(jnp.asarray(expected_terms))
self.assertAllClose(y, expected_y, atol=atol, rtol=rtol,
check_dtypes=check_dtypes)
self.assertAllClose(terms, expected_terms, atol=atol, rtol=rtol,
check_dtypes=check_dtypes)
@jtu.skip_on_devices("tpu")
# Default tolerance too tight on A100 after openxla/xla@a58070090
@jax.default_matmul_precision("float32")
def test_dot(self):
M, K, N = 2, 3, 4
order = 3
rng = self.rng()
x1 = rng.randn(M, K)
x2 = rng.randn(K, N)
primals = (x1, x2)
terms_in1 = [rng.randn(*x1.shape) for _ in range(order)]
terms_in2 = [rng.randn(*x2.shape) for _ in range(order)]
series_in = (terms_in1, terms_in2)
self.check_jet(jnp.dot, primals, series_in)
@jtu.skip_on_devices("tpu")
@jax.legacy_prng_key('allow')
def test_conv(self):
order = 3
input_shape = (1, 5, 5, 1)
key = random.PRNGKey(0)
# TODO(duvenaud): Check all types of padding
init_fun, apply_fun = stax.Conv(3, (2, 2), padding='VALID')
_, (W, b) = init_fun(key, input_shape)
rng = self.rng()
x = rng.randn(*input_shape).astype(W.dtype)
primals = (W, b, x)
series_in1 = [rng.randn(*W.shape).astype(W.dtype) for _ in range(order)]
series_in2 = [rng.randn(*b.shape).astype(W.dtype) for _ in range(order)]
series_in3 = [rng.randn(*x.shape).astype(W.dtype) for _ in range(order)]
series_in = (series_in1, series_in2, series_in3)
def f(W, b, x):
return apply_fun((W, b), x)
self.check_jet(f, primals, series_in, check_dtypes=False)
def unary_check(self, fun, lims=(-2, 2), order=3, dtype=None, atol=1e-3,
rtol=1e-3):
dims = 2, 3
rng = self.rng()
if dtype is None:
primal_in = transform(lims, rng.rand(*dims))
terms_in = [rng.randn(*dims) for _ in range(order)]
else:
rng = jtu.rand_uniform(rng, *lims)
primal_in = rng(dims, dtype)
terms_in = [rng(dims, dtype) for _ in range(order)]
self.check_jet(fun, (primal_in,), (terms_in,), atol, rtol)
def binary_check(self, fun, lims=None, order=3, finite=True, dtype=None):
lims = lims or [-2, 2]
dims = 2, 3
rng = self.rng()
if isinstance(lims, tuple):
x_lims, y_lims = lims
else:
x_lims, y_lims = lims, lims
if dtype is None:
primal_in = (transform(x_lims, rng.rand(*dims)),
transform(y_lims, rng.rand(*dims)))
series_in = ([rng.randn(*dims) for _ in range(order)],
[rng.randn(*dims) for _ in range(order)])
else:
rng = jtu.rand_uniform(rng, *lims)
primal_in = (rng(dims, dtype),
rng(dims, dtype))
series_in = ([rng(dims, dtype) for _ in range(order)],
[rng(dims, dtype) for _ in range(order)])
if finite:
self.check_jet(fun, primal_in, series_in, atol=1e-4, rtol=1e-4)
else:
self.check_jet_finite(fun, primal_in, series_in, atol=1e-4, rtol=1e-4)
def unary_check_float0(self, fun, lims=(-2, 2), order=3, dtype=None):
# like unary_check but for functions that output integers (so their tangent
# type is float0 arrays)
raise unittest.SkipTest("jet tests must be adapted for integer-output functions")
def binary_check_float0(self, fun, lims=(-2, 2), order=3, finite=True, dtype=None):
# like binary_check but for functions that output integers (so their tangent
# type is float0 arrays)
raise unittest.SkipTest("jet tests must be adapted for integer-output functions")
def expit_check(self, lims=(-2, 2), order=3):
dims = 2, 3
rng = self.rng()
primal_in = transform(lims, rng.rand(*dims))
terms_in = [rng.randn(*dims) for _ in range(order)]
primals = (primal_in, )
series = (terms_in, )
y, terms = jax.experimental.jet._logistic_taylor(primals, series)
expected_y, expected_terms = jvp_taylor(jax.scipy.special.expit, primals, series)
atol = 1e-4
rtol = 1e-4
self.assertAllClose(y, expected_y, atol=atol, rtol=rtol)
self.assertAllClose(terms, expected_terms, atol=atol, rtol=rtol)
@jtu.skip_on_devices("tpu")
def test_int_pow(self):
for p in range(6):
self.unary_check(lambda x: x ** p, lims=[-2, 2])
self.unary_check(lambda x: x ** 10, lims=[0, 0])
@jtu.skip_on_devices("tpu")
def test_is_finite(self): self.unary_check_float0(lax.is_finite)
@jtu.skip_on_devices("tpu")
def test_and(self): self.binary_check_float0(lax.bitwise_and, dtype=np.bool_)
@jtu.skip_on_devices("tpu")
def test_or(self): self.binary_check_float0(lax.bitwise_or, dtype=np.bool_)
@jtu.skip_on_devices("tpu")
def test_xor(self): self.binary_check_float0(jnp.bitwise_xor, dtype=np.bool_)
@jtu.skip_on_devices("tpu")
def test_shift_left(self): self.binary_check_float0(lax.shift_left, dtype=np.int32)
@jtu.skip_on_devices("tpu")
def test_shift_right_a(self): self.binary_check_float0(lax.shift_right_arithmetic, dtype=np.int32)
@jtu.skip_on_devices("tpu")
def test_shift_right_l(self): self.binary_check_float0(lax.shift_right_logical, dtype=np.int32)
@jtu.skip_on_devices("tpu")
def test_le(self): self.binary_check_float0(lambda x, y: x <= y)
@jtu.skip_on_devices("tpu")
def test_gt(self): self.binary_check_float0(lambda x, y: x > y)
@jtu.skip_on_devices("tpu")
def test_lt(self): self.binary_check_float0(lambda x, y: x < y)
@jtu.skip_on_devices("tpu")
def test_ge(self): self.binary_check_float0(lambda x, y: x >= y)
@jtu.skip_on_devices("tpu")
def test_eq(self): self.binary_check_float0(lambda x, y: x == y)
@jtu.skip_on_devices("tpu")
def test_ne(self): self.binary_check_float0(lambda x, y: x != y)
@jtu.skip_on_devices("tpu")
def test_not(self): self.unary_check_float0(lax.bitwise_not, dtype=np.bool_)
@jtu.skip_on_devices("tpu")
def test_exp(self): self.unary_check(jnp.exp)
@jtu.skip_on_devices("tpu")
def test_neg(self): self.unary_check(jnp.negative)
@jtu.skip_on_devices("tpu")
def test_floor(self): self.unary_check(jnp.floor)
@jtu.skip_on_devices("tpu")
def test_ceil(self): self.unary_check(jnp.ceil)
@jtu.skip_on_devices("tpu")
def test_trunc(self): self.unary_check(jnp.trunc)
@jtu.skip_on_devices("tpu")
def test_round(self): self.unary_check(lax.round)
@jtu.skip_on_devices("tpu")
def test_sign(self): self.unary_check(lax.sign)
@jtu.skip_on_devices("tpu")
def test_real(self): self.unary_check(lax.real, dtype=np.complex64)
@jtu.skip_on_devices("tpu")
def test_conj(self): self.unary_check(lax.conj, dtype=np.complex64)
@jtu.skip_on_devices("tpu")
def test_imag(self): self.unary_check(lax.imag, dtype=np.complex64)
@jtu.skip_on_devices("tpu")
def test_log(self): self.unary_check(jnp.log, lims=[0.8, 4.0])
@jtu.skip_on_devices("tpu")
def test_gather(self): self.unary_check(lambda x: x[1:])
@jtu.skip_on_devices("tpu")
def test_reduce_max(self): self.unary_check(lambda x: x.max(axis=1))
@jtu.skip_on_devices("tpu")
def test_reduce_min(self): self.unary_check(lambda x: x.min(axis=1))
@jtu.skip_on_devices("tpu")
def test_all_max(self): self.unary_check(jnp.max)
@jtu.skip_on_devices("tpu")
def test_all_min(self): self.unary_check(jnp.min)
@jtu.skip_on_devices("tpu")
def test_stopgrad(self): self.unary_check(lax.stop_gradient)
@jtu.skip_on_devices("tpu")
def test_abs(self): self.unary_check(jnp.abs)
@jtu.skip_on_devices("tpu")
def test_fft(self): self.unary_check(jnp.fft.fft)
@jtu.skip_on_devices("tpu")
def test_log1p(self): self.unary_check(jnp.log1p, lims=[0, 4.])
@jtu.skip_on_devices("tpu")
def test_expm1(self): self.unary_check(jnp.expm1)
@jtu.skip_on_devices("tpu")
def test_sin(self): self.unary_check(jnp.sin)
@jtu.skip_on_devices("tpu")
def test_cos(self): self.unary_check(jnp.cos)
@jtu.skip_on_devices("tpu")
def test_sinh(self): self.unary_check(jnp.sinh)
@jtu.skip_on_devices("tpu")
def test_cosh(self): self.unary_check(jnp.cosh)
@jtu.skip_on_devices("tpu")
def test_tanh(self): self.unary_check(jnp.tanh, lims=[-500, 500], order=5,
atol=5e-3)
@jtu.skip_on_devices("tpu")
def test_logistic(self): self.unary_check(lax.logistic, lims=[-100, 100], order=5)
@jtu.skip_on_devices("tpu")
def test_expit2(self): self.expit_check(lims=[-500, 500], order=5)
@jtu.skip_on_devices("tpu")
def test_sqrt(self): self.unary_check(jnp.sqrt, lims=[0, 5.])
@jtu.skip_on_devices("tpu")
def test_rsqrt(self): self.unary_check(lax.rsqrt, lims=[0, 5000.])
@jtu.skip_on_devices("tpu")
def test_asinh(self): self.unary_check(lax.asinh, lims=[-100, 100])
@jtu.skip_on_devices("tpu")
def test_acosh(self): self.unary_check(lax.acosh, lims=[-100, 100])
@jtu.skip_on_devices("tpu")
def test_atanh(self): self.unary_check(lax.atanh, lims=[-1, 1])
@jtu.skip_on_devices("tpu")
def test_erf(self): self.unary_check(lax.erf)
@jtu.skip_on_devices("tpu")
def test_erfc(self): self.unary_check(lax.erfc)
@jtu.skip_on_devices("tpu")
def test_erf_inv(self): self.unary_check(lax.erf_inv, lims=[-1, 1])
@jtu.skip_on_devices("tpu")
def test_cumsum(self): self.unary_check(jnp.cumsum)
@jtu.skip_on_devices("tpu")
def test_cumprod(self): self.unary_check(jnp.cumprod)
@jtu.skip_on_devices("tpu")
def test_cummax(self): self.unary_check(partial(lax.cummax, axis=0))
@jtu.skip_on_devices("tpu")
def test_cummin(self): self.unary_check(partial(lax.cummin, axis=0))
@jtu.skip_on_devices("tpu")
def test_dynamic_slice(self): self.unary_check(partial(lax.dynamic_slice, start_indices=(1,2), slice_sizes=(1,1)))
@jtu.skip_on_devices("tpu")
def test_dynamic_update_slice(self): self.unary_check(partial(lax.dynamic_update_slice, start_indices=(1,2), update=np.arange(6.0).reshape(2, 3)))
@jtu.skip_on_devices("tpu")
def test_div(self): self.binary_check(lambda x, y: x / y, lims=[0.8, 4.0])
@jtu.skip_on_devices("tpu")
def test_rem(self): self.binary_check(lax.rem, lims=[0.8, 4.0])
@jtu.skip_on_devices("tpu")
def test_complex(self): self.binary_check(lax.complex)
@jtu.skip_on_devices("tpu")
def test_sub(self): self.binary_check(lambda x, y: x - y)
@jtu.skip_on_devices("tpu")
def test_add(self): self.binary_check(lambda x, y: x + y)
@jtu.skip_on_devices("tpu")
def test_mul(self): self.binary_check(lambda x, y: x * y)
@jtu.skip_on_devices("tpu")
def test_max(self): self.binary_check(lax.max)
@jtu.skip_on_devices("tpu")
def test_min(self): self.binary_check(lax.min)
@jtu.skip_on_devices("tpu")
@jtu.ignore_warning(message="overflow encountered in power")
def test_pow(self): self.binary_check(lambda x, y: x ** y, lims=([0.2, 500], [-500, 500]), finite=False)
@jtu.skip_on_devices("tpu")
def test_atan2(self): self.binary_check(lax.atan2, lims=[-40, 40])
@jtu.skip_on_devices("tpu")
def test_clamp(self):
lims = [-1, 1]
order = 3
dims = 2, 3
# TODO(jakevdp): This test is very sensitive to the inputs, so we use a known
# working seed. We should instead use self.rng(), and make sure that the primal
# points lie outside an epsilon ball of the two critical points in the function.
rng = np.random.RandomState(0)
primal_in = (transform(lims, rng.rand(*dims)),
transform(lims, rng.rand(*dims)),
transform(lims, rng.rand(*dims)))
series_in = ([rng.randn(*dims) for _ in range(order)],
[rng.randn(*dims) for _ in range(order)],
[rng.randn(*dims) for _ in range(order)])
self.check_jet(lax.clamp, primal_in, series_in, atol=1e-4, rtol=1e-4)
def test_process_call(self):
def f(x):
return jit(lambda x: x * x)(x)
self.unary_check(f, rtol=2e-4)
def test_post_process_call(self):
def f(x):
return jit(lambda y: x * y)(2.)
self.unary_check(f, rtol=5e-4)
def test_select(self):
M, K = 2, 3
order = 3
rng = self.rng()
b = rng.rand(M, K) < 0.5
x = rng.randn(M, K)
y = rng.randn(M, K)
primals = (b, x, y)
terms_b = [rng.randn(*b.shape) for _ in range(order)]
terms_x = [rng.randn(*x.shape) for _ in range(order)]
terms_y = [rng.randn(*y.shape) for _ in range(order)]
series_in = (terms_b, terms_x, terms_y)
# Since this nudges bool inputs, we need to allow promotion to float.
with jax.numpy_dtype_promotion('standard'):
self.check_jet(jnp.where, primals, series_in, rtol=5e-4)
def test_inst_zero(self):
def f(x):
return jnp.full_like(x, 2.)
def g(x):
return 2. + 0 * x
x = jnp.ones(1)
order = 3
f_out_primals, f_out_series = jet(f, (x, ), ([jnp.ones_like(x) for _ in range(order)], ))
assert f_out_series is not zero_series
g_out_primals, g_out_series = jet(g, (x, ), ([jnp.ones_like(x) for _ in range(order)], ))
self.assertArraysEqual(g_out_primals, f_out_primals)
self.assertArraysEqual(g_out_series, f_out_series)
def test_add_any(self):
# https://github.com/jax-ml/jax/issues/5217
f = lambda x, eps: x * eps + eps + x
def g(eps):
x = jnp.array(1.)
return jax.grad(f)(x, eps)
jet(g, (1.,), ([1.],)) # doesn't crash
def test_scatter_add(self):
# very basic test from https://github.com/jax-ml/jax/issues/5365
def f(x):
x0 = x[0]
x1 = x[1]
return (x0**5 + x1**5).sum()
def h(eps):
from jax import jacfwd, grad
x = jnp.array([1., 1.])
μ = eps * x
def F(t):
return f(x + t * μ)
return grad(jacfwd(F))(0.)
self.check_jet(h, (0.,), ([1., 2., 3.],), rtol=1e-3)
if __name__ == '__main__':
absltest.main(testLoader=jtu.JaxTestLoader())