-
Notifications
You must be signed in to change notification settings - Fork 2
/
adamlrm_test.py
582 lines (505 loc) · 24.8 KB
/
adamlrm_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
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
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# 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
#
# http://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.
# ==============================================================================
"""Tests for AdamLRM."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
from tensorflow.python.eager import context
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import test_util
from tensorflow.python.keras import optimizers
#from tensorflow.python.keras.optimizer_v2 import adam
from tensorflow.python.keras.optimizer_v2 import learning_rate_schedule
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.ops import variables
from tensorflow.python.platform import test
import adamlrm
def adam_update_numpy(param,
g_t,
t,
m,
v,
lr=0.001,
beta1=0.9,
beta2=0.999,
epsilon=1e-7):
lr_t = lr * np.sqrt(1 - beta2**(t + 1)) / (1 - beta1**(t + 1))
m_t = beta1 * m + (1 - beta1) * g_t
v_t = beta2 * v + (1 - beta2) * g_t * g_t
param_t = param - lr_t * m_t / (np.sqrt(v_t) + epsilon)
return param_t, m_t, v_t
def adam_update_numpy_amsgrad(param,
g_t,
t,
m,
v,
vhat,
lr=0.001,
beta1=0.9,
beta2=0.999,
epsilon=1e-7):
lr_t = lr * np.sqrt(1 - beta2**(t + 1)) / (1 - beta1**(t + 1))
m_t = beta1 * m + (1 - beta1) * g_t
v_t = beta2 * v + (1 - beta2) * g_t * g_t
vhat_t = np.maximum(vhat, v_t)
param_t = param - lr_t * m_t / (np.sqrt(vhat_t) + epsilon)
return param_t, m_t, v_t, vhat_t
def adam_sparse_update_numpy_amsgrad(param,
indices,
g_t,
t,
m,
v,
vhat,
lr=0.001,
beta1=0.9,
beta2=0.999,
epsilon=1e-7):
m_t, v_t, vhat_t, param_t = (np.copy(m), np.copy(v), np.copy(vhat),
np.copy(param))
lr_t = lr * np.sqrt(1 - beta2**(t + 1)) / (1 - beta1**(t + 1))
m_t_slice = beta1 * m[indices] + (1 - beta1) * g_t
v_t_slice = beta2 * v[indices] + (1 - beta2) * g_t * g_t
m_t[indices] = m_t_slice
v_t[indices] = v_t_slice
v_hat_t = np.maximum(vhat_t, v_t)
v_hat_t_slice = v_hat_t[indices]
param_t_slice = param[indices] - (
lr_t * (m_t_slice / (np.sqrt(v_hat_t_slice) + epsilon)))
param_t[indices] = param_t_slice
return param_t, m_t, v_t, vhat_t
def get_beta_accumulators(opt, dtype):
local_step = math_ops.cast(opt.iterations + 1, dtype)
beta_1_t = math_ops.cast(opt._get_hyper("beta_1"), dtype)
beta_1_power = math_ops.pow(beta_1_t, local_step)
beta_2_t = math_ops.cast(opt._get_hyper("beta_2"), dtype)
beta_2_power = math_ops.pow(beta_2_t, local_step)
return (beta_1_power, beta_2_power)
class AdamLRMOptimizerTest(test.TestCase):
@test_util.run_deprecated_v1
def testSparse(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_session(use_gpu=True):
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.0, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.0, 0.01], dtype=dtype.as_numpy_dtype)
var0 = resource_variable_ops.ResourceVariable(var0_np, name='var0')
var1 = resource_variable_ops.ResourceVariable(var1_np, name='var1')
grads0_np_indices = np.array([0, 2], dtype=np.int32)
grads0 = ops.IndexedSlices(
constant_op.constant(grads0_np[grads0_np_indices]),
constant_op.constant(grads0_np_indices), constant_op.constant([3]))
grads1_np_indices = np.array([0, 2], dtype=np.int32)
grads1 = ops.IndexedSlices(
constant_op.constant(grads1_np[grads1_np_indices]),
constant_op.constant(grads1_np_indices), constant_op.constant([3]))
opt = adamlrm.AdamLRM(lr_multiplier={'var1':1e-2})
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllClose([1.0, 1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 3.0, 4.0], self.evaluate(var1))
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
# Run 3 steps of Adam
for t in range(3):
self.assertAllCloseAccordingToType(0.9**(t + 1),
self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(0.999**(t + 1),
self.evaluate(beta_2_power))
update.run()
var0_np, m0, v0 = adam_update_numpy(var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(var1_np, grads1_np, t, m1, v1,
lr=1e-2*0.001)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
@test_util.run_deprecated_v1
def testSparseDevicePlacement(self):
for index_dtype in [dtypes.int32, dtypes.int64]:
with self.cached_session(force_gpu=test.is_gpu_available()):
# If a GPU is available, tests that all optimizer ops can be placed on
# it (i.e. they have GPU kernels).
var = variables.Variable([[1.0], [2.0]])
indices = constant_op.constant([0, 1], dtype=index_dtype)
g_sum = lambda: math_ops.reduce_sum(array_ops.gather(var, indices)) # pylint: disable=cell-var-from-loop
optimizer = adamlrm.AdamLRM(3.0)
minimize_op = optimizer.minimize(g_sum, var_list=[var])
variables.global_variables_initializer().run()
minimize_op.run()
@test_util.run_deprecated_v1
def testSparseRepeatedIndices(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_session():
repeated_index_update_var = variables.Variable(
[[1.0], [2.0]], dtype=dtype)
aggregated_update_var = variables.Variable(
[[1.0], [2.0]], dtype=dtype)
grad_repeated_index = ops.IndexedSlices(
constant_op.constant(
[0.1, 0.1], shape=[2, 1], dtype=dtype),
constant_op.constant([1, 1]),
constant_op.constant([2, 1]))
grad_aggregated = ops.IndexedSlices(
constant_op.constant(
[0.2], shape=[1, 1], dtype=dtype),
constant_op.constant([1]),
constant_op.constant([2, 1]))
repeated_update = adamlrm.AdamLRM().apply_gradients(
[(grad_repeated_index, repeated_index_update_var)])
aggregated_update = adamlrm.AdamLRM().apply_gradients(
[(grad_aggregated, aggregated_update_var)])
variables.global_variables_initializer().run()
self.assertAllClose(aggregated_update_var.eval(),
self.evaluate(repeated_index_update_var))
for _ in range(3):
repeated_update.run()
aggregated_update.run()
self.assertAllClose(aggregated_update_var.eval(),
self.evaluate(repeated_index_update_var))
def doTestBasic(self, use_callable_params=False):
for i, dtype in enumerate([dtypes.half, dtypes.float32, dtypes.float64]):
with self.cached_session(use_gpu=True):
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
var0 = resource_variable_ops.ResourceVariable(
var0_np, name="var0_%d" % i)
var1 = resource_variable_ops.ResourceVariable(
var1_np, name="var1_%d" % i)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
learning_rate = lambda: 0.001
beta1 = lambda: 0.9
beta2 = lambda: 0.999
epsilon = lambda: 1e-8
if not use_callable_params:
learning_rate = learning_rate()
beta1 = beta1()
beta2 = beta2()
epsilon = epsilon()
opt = adamlrm.AdamLRM(
learning_rate=learning_rate, lr_multiplier={'var1':1e-2})
if not context.executing_eagerly():
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
# Run 3 steps of Adam
for t in range(3):
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
self.assertAllCloseAccordingToType(0.9**(t + 1),
self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(0.999**(t + 1),
self.evaluate(beta_2_power))
if not context.executing_eagerly():
self.evaluate(update)
else:
opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
var0_np, m0, v0 = adam_update_numpy(var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(var1_np, grads1_np, t, m1, v1,
lr=1e-2*0.001)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
@test_util.run_in_graph_and_eager_modes(reset_test=True)
def testResourceBasic(self):
self.doTestBasic()
def testBasicCallableParams(self):
with context.eager_mode():
self.doTestBasic(use_callable_params=True)
@test_util.run_in_graph_and_eager_modes(reset_test=True)
def testBasicWithAmsgrad(self):
for i, dtype in enumerate([dtypes.half, dtypes.float32, dtypes.float64]):
with self.cached_session(use_gpu=True):
# Initialize variables for numpy implementation.
m0, v0, v0hat, m1, v1, v1hat = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
var0 = resource_variable_ops.ResourceVariable(
var0_np, name="var0_%d" % i)
var1 = resource_variable_ops.ResourceVariable(
var1_np, name="var1_%d" % i)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
opt = adamlrm.AdamLRM(amsgrad=True, lr_multiplier={'var1':1e-2})
if not context.executing_eagerly():
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
# Run 3 steps of Adam
for t in range(3):
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
self.assertAllCloseAccordingToType(0.9**(t + 1),
self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(0.999**(t + 1),
self.evaluate(beta_2_power))
if not context.executing_eagerly():
self.evaluate(update)
else:
opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
var0_np, m0, v0, v0hat = adam_update_numpy_amsgrad(
var0_np, grads0_np, t, m0, v0, v0hat)
var1_np, m1, v1, v1hat = adam_update_numpy_amsgrad(
var1_np, grads1_np, t, m1, v1, v1hat,
lr=1e-2*0.001)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
@test_util.run_in_graph_and_eager_modes
def testSparseWithAmsgrad(self):
# dtypes.half does not work on gpu + eager.
for dtype in [dtypes.float32, dtypes.float64]:
with self.cached_session():
m0 = np.array([[0.0], [0.0]])
v0 = np.array([[0.0], [0.0]])
v0hat = np.array([[0.0], [0.0]])
indices_np = np.array([1])
indices = constant_op.constant(indices_np, dtype=dtypes.int32)
var0_np = np.array([[1.0], [2.0]], dtype=dtype.as_numpy_dtype)
repeated_index_update_var = variables.Variable(var0_np, dtype=dtype)
aggregated_update_var = variables.Variable(var0_np, dtype=dtype)
grads0_np = np.array([[0.2]], dtype=dtype.as_numpy_dtype)
grad_repeated_index = ops.IndexedSlices(
constant_op.constant([0.1, 0.1], shape=[2, 1], dtype=dtype),
constant_op.constant([1, 1]), constant_op.constant([2, 1]))
grad_aggregated = ops.IndexedSlices(grads0_np, indices,
constant_op.constant([2, 1]))
opt_repeated = adamlrm.AdamLRM(amsgrad=True)
opt_aggregated = adamlrm.AdamLRM(amsgrad=True)
if not context.executing_eagerly():
repeated_update = opt_repeated.apply_gradients(
[(grad_repeated_index, repeated_index_update_var)])
aggregated_update = opt_aggregated.apply_gradients(
[(grad_aggregated, aggregated_update_var)])
self.evaluate(variables.global_variables_initializer())
self.assertAllClose(
self.evaluate(aggregated_update_var),
self.evaluate(repeated_index_update_var))
for t in range(3):
if not context.executing_eagerly():
self.evaluate(repeated_update)
self.evaluate(aggregated_update)
else:
opt_repeated.apply_gradients(
[(grad_repeated_index, repeated_index_update_var)])
opt_aggregated.apply_gradients(
[(grad_aggregated, aggregated_update_var)])
var0_np, m0, v0, v0hat = adam_sparse_update_numpy_amsgrad(
var0_np, indices_np, grads0_np, t, m0, v0, v0hat)
# Validate updated params
self.assertAllCloseAccordingToType(
var0_np, self.evaluate(aggregated_update_var))
self.assertAllCloseAccordingToType(
self.evaluate(aggregated_update_var),
self.evaluate(repeated_index_update_var))
@test_util.run_deprecated_v1
def testBasicWithLearningRateDecay(self):
for i, dtype in enumerate([dtypes.half, dtypes.float32, dtypes.float64]):
with self.cached_session(use_gpu=True):
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
var0 = resource_variable_ops.ResourceVariable(
var0_np, name="var0_%d" % i)
var1 = resource_variable_ops.ResourceVariable(
var1_np, name="var1_%d" % i)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
learning_rate = 0.001
beta_1 = 0.9
beta_2 = 0.999
epsilon = 1e-7
decay = 0.5
opt = adamlrm.AdamLRM(
learning_rate=learning_rate,
beta_1=beta_1,
beta_2=beta_2,
epsilon=epsilon,
decay=decay,
lr_multiplier={'var1':1e-2})
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
# Run 3 steps of Adam
for t in range(3):
self.evaluate(update)
lr_np = learning_rate / (1 + decay * t)
var0_np, m0, v0 = adam_update_numpy(
var0_np, grads0_np, t, m0, v0, lr=lr_np)
var1_np, m1, v1 = adam_update_numpy(
var1_np, grads1_np, t, m1, v1, lr=1e-2*lr_np)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
@test_util.run_deprecated_v1
def testBasicWithLearningRateInverseTimeDecay(self):
for i, dtype in enumerate([dtypes.half, dtypes.float32, dtypes.float64]):
with self.cached_session(use_gpu=True):
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
var0 = resource_variable_ops.ResourceVariable(
var0_np, name="var0_%d" % i)
var1 = resource_variable_ops.ResourceVariable(
var1_np, name="var1_%d" % i)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
learning_rate = 0.001
decay = 0.5
lr_schedule = learning_rate_schedule.InverseTimeDecay(
learning_rate, decay_steps=1.0, decay_rate=decay)
beta_1 = 0.9
beta_2 = 0.999
epsilon = 1e-7
opt = adamlrm.AdamLRM(
learning_rate=lr_schedule,
beta_1=beta_1,
beta_2=beta_2,
epsilon=epsilon,
lr_multiplier={'var1':1e-2})
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
# Run 3 steps of Adam
for t in range(3):
self.evaluate(update)
lr_np = learning_rate / (1 + decay * t)
var0_np, m0, v0 = adam_update_numpy(
var0_np, grads0_np, t, m0, v0, lr=lr_np)
var1_np, m1, v1 = adam_update_numpy(
var1_np, grads1_np, t, m1, v1, lr=1e-2*lr_np)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
@test_util.run_deprecated_v1
def testTensorLearningRate(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_session(use_gpu=True):
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
var0 = variables.Variable(var0_np, name='var0')
var1 = variables.Variable(var1_np, name='var1')
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
opt = adamlrm.AdamLRM(constant_op.constant(0.001),
lr_multiplier={'var1':1e-2})
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
# Run 3 steps of Adam
for t in range(3):
self.assertAllCloseAccordingToType(0.9**(t + 1),
self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(0.999**(t + 1),
self.evaluate(beta_2_power))
update.run()
var0_np, m0, v0 = adam_update_numpy(var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(var1_np, grads1_np, t, m1, v1,
lr=1e-2*0.001)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
@test_util.run_deprecated_v1
def testSharing(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_session(use_gpu=True):
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
var0 = variables.Variable(var0_np, name='var0')
var1 = variables.Variable(var1_np, name='var1')
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
opt = adamlrm.AdamLRM(lr_multiplier={'var1':1e-2})
update1 = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
update2 = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
# Run 3 steps of intertwined Adam1 and Adam2.
for t in range(3):
self.assertAllCloseAccordingToType(0.9**(t + 1),
self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(0.999**(t + 1),
self.evaluate(beta_2_power))
if t % 2 == 0:
update1.run()
else:
update2.run()
var0_np, m0, v0 = adam_update_numpy(var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(var1_np, grads1_np, t, m1, v1,
lr=1e-2*0.001)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
def testSlotsUniqueEager(self):
with context.eager_mode():
v1 = resource_variable_ops.ResourceVariable(1.)
v2 = resource_variable_ops.ResourceVariable(1.)
opt = adamlrm.AdamLRM(1.)
opt.minimize(lambda: v1 + v2, var_list=[v1, v2])
# There should be iteration, and two unique slot variables for v1 and v2.
self.assertEqual(
5, len(set([v.experimental_ref() for v in opt.variables()])))
self.assertEqual(
self.evaluate(opt.variables()[0]), self.evaluate(opt.iterations))
def testSetWeightsFromV1AdamLRMWithoutMinimize(self):
keras_v1_adam = optimizers.Adam()
keras_v2_adam = adamlrm.AdamLRM()
keras_v2_adam.set_weights(keras_v1_adam.get_weights())
keras_v1_iteration = keras_v1_adam.iterations
keras_v2_iteration = keras_v2_adam.iterations
self.evaluate(variables.global_variables_initializer())
self.assertEqual(
self.evaluate(keras_v1_iteration), self.evaluate(keras_v2_iteration))
def testConstructAdamLRMWithLR(self):
opt = adamlrm.AdamLRM(lr=1.0)
opt_2 = adamlrm.AdamLRM(learning_rate=0.1, lr=1.0)
opt_3 = adamlrm.AdamLRM(learning_rate=0.1)
self.assertIsInstance(opt.lr, variables.Variable)
self.assertIsInstance(opt_2.lr, variables.Variable)
self.assertIsInstance(opt_3.lr, variables.Variable)
self.evaluate(variables.global_variables_initializer())
self.assertAllClose(self.evaluate(opt.lr), (1.0))
self.assertAllClose(self.evaluate(opt_2.lr), (1.0))
self.assertAllClose(self.evaluate(opt_3.lr), (0.1))
if __name__ == "__main__":
test.main()