forked from apd10/dlrm
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dlrm_s_caffe2.py
1703 lines (1563 loc) · 71.6 KB
/
dlrm_s_caffe2.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
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
#
# Description: an implementation of a deep learning recommendation model (DLRM)
# The model input consists of dense and sparse features. The former is a vector
# of floating point values. The latter is a list of sparse indices into
# embedding tables, which consist of vectors of floating point values.
# The selected vectors are passed to mlp networks denoted by triangles,
# in some cases the vectors are interacted through operators (Ops).
#
# output:
# vector of values
# model: |
# /\
# /__\
# |
# _____________________> Op <___________________
# / | \
# /\ /\ /\
# /__\ /__\ ... /__\
# | | |
# | Op Op
# | ____/__\_____ ____/__\____
# | |_Emb_|____|__| ... |_Emb_|__|___|
# input:
# [ dense features ] [sparse indices] , ..., [sparse indices]
#
# More precise definition of model layers:
# 1) fully connected layers of an mlp
# z = f(y)
# y = Wx + b
#
# 2) embedding lookup (for a list of sparse indices p=[p1,...,pk])
# z = Op(e1,...,ek)
# obtain vectors e1=E[:,p1], ..., ek=E[:,pk]
#
# 3) Operator Op can be one of the following
# Sum(e1,...,ek) = e1 + ... + ek
# Dot(e1,...,ek) = [e1'e1, ..., e1'ek, ..., ek'e1, ..., ek'ek]
# Cat(e1,...,ek) = [e1', ..., ek']'
# where ' denotes transpose operation
#
# References:
# [1] Maxim Naumov, Dheevatsa Mudigere, Hao-Jun Michael Shi, Jianyu Huang,
# Narayanan Sundaram, Jongsoo Park, Xiaodong Wang, Udit Gupta, Carole-Jean Wu,
# Alisson G. Azzolini, Dmytro Dzhulgakov, Andrey Mallevich, Ilia Cherniavskii,
# Yinghai Lu, Raghuraman Krishnamoorthi, Ansha Yu, Volodymyr Kondratenko,
# Stephanie Pereira, Xianjie Chen, Wenlin Chen, Vijay Rao, Bill Jia, Liang Xiong,
# Misha Smelyanskiy, "Deep Learning Recommendation Model for Personalization and
# Recommendation Systems", CoRR, arXiv:1906.00091, 2019
from __future__ import absolute_import, division, print_function, unicode_literals
import functools
# others
import operator
import time
import copy
# data generation
import dlrm_data_pytorch as dp
# numpy
import numpy as np
import sklearn.metrics
# onnx
# The onnx import causes deprecation warnings every time workers
# are spawned during testing. So, we filter out those warnings.
import warnings
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=DeprecationWarning)
try:
import onnx
import caffe2.python.onnx.frontend
except ImportError as error:
print('Unable to import onnx or caffe2.python.onnx.frontend ', error)
# from caffe2.python import data_parallel_model
# caffe2
from caffe2.proto import caffe2_pb2
from caffe2.python import brew, core, dyndep, model_helper, net_drawer, workspace
"""
# auxiliary routine used to split input on the mini-bacth dimension
def where_to_split(mini_batch_size, ndevices, _add_leftover=False):
n = (mini_batch_size + ndevices - 1) // ndevices # ceiling
l = mini_batch_size - n * (ndevices - 1) # leftover
s = [n] * (ndevices - 1)
if _add_leftover:
ls += [l if l > 0 else n]
return ls
"""
### define dlrm in Caffe2 ###
class DLRM_Net(object):
def FeedBlobWrapper(self, tag, val, add_prefix=True, split=False, device_id=-1):
if self.ndevices > 1 and add_prefix:
if split:
# split across devices
mini_batch_size = val.shape[0]
# approach 1: np and caffe2 operators assume the mini-batch size is
# divisible exactly by the number of available devices
if mini_batch_size % self.ndevices != 0:
sys.exit("ERROR: caffe2 net assumes that the mini_batch_size "
+ str(mini_batch_size)
+ " is evenly divisible by the number of available devices"
+ str(self.ndevices))
vals = np.split(val, self.ndevices, axis=0)
"""
# approach 2: np and caffe2 operators do not assume exact divisibility
if args.mini_batch_size != mini_batch_size:
sys.exit("ERROR: caffe2 net was prepared for mini-batch size "
+ str(args.mini_batch_size)
+ " which is different from current mini-batch size "
+ str(mini_batch_size) + " being passed to it. "
+ "This is common for the last mini-batch, when "
+ "mini-batch size does not evenly divided the number of "
+ "elements in the data set.")
ls = where_to_split(mini_batch_size, self.ndevices)
vals = np.split(val, ls, axis=0)
"""
# feed to multiple devices
for d in range(self.ndevices):
tag_on_device = "gpu_" + str(d) + "/" + tag
_d = core.DeviceOption(workspace.GpuDeviceType, d)
workspace.FeedBlob(tag_on_device, vals[d], device_option=_d)
else:
# feed to multiple devices
for d in range(self.ndevices):
tag_on_device = "gpu_" + str(d) + "/" + tag
_d = core.DeviceOption(workspace.GpuDeviceType, d)
workspace.FeedBlob(tag_on_device, val, device_option=_d)
else:
# feed to a single device (named or not)
if device_id >= 0:
_d = core.DeviceOption(workspace.GpuDeviceType, device_id)
workspace.FeedBlob(tag, val, device_option=_d)
else:
workspace.FeedBlob(tag, val)
def FetchBlobWrapper(self, tag, add_prefix=True, reduce_across=None, device_id=-1):
if self.ndevices > 1 and add_prefix:
# fetch from multiple devices
vals = []
for d in range(self.ndevices):
if tag.__class__ == list:
tag_on_device = tag[d]
else:
tag_on_device = "gpu_" + str(0) + "/" + tag
val = workspace.FetchBlob(tag_on_device)
vals.append(val)
# reduce across devices
if reduce_across == "add":
return functools.reduce(operator.add, vals)
elif reduce_across == "concat":
return np.concatenate(vals)
else:
return vals
else:
# fetch from a single device (named or not)
if device_id >= 0:
tag_on_device = "gpu_" + str(device_id) + "/" + tag
return workspace.FetchBlob(tag_on_device)
else:
return workspace.FetchBlob(tag)
def AddLayerWrapper(self, layer, inp_blobs, out_blobs,
add_prefix=True, reset_grad=False, **kwargs):
# auxiliary routine to adjust tags
def adjust_tag(blobs, on_device):
if blobs.__class__ == str:
_blobs = on_device + blobs
elif blobs.__class__ == list:
_blobs = list(map(lambda tag: on_device + tag, blobs))
else: # blobs.__class__ == model_helper.ModelHelper or something else
_blobs = blobs
return _blobs
if self.ndevices > 1 and add_prefix:
# add layer on multiple devices
ll = []
for d in range(self.ndevices):
# add prefix on_device
on_device = "gpu_" + str(d) + "/"
_inp_blobs = adjust_tag(inp_blobs, on_device)
_out_blobs = adjust_tag(out_blobs, on_device)
# WARNING: reset_grad option was exlusively designed for WeightedSum
# with inp_blobs=[w, tag_one, "", lr], where "" will be replaced
if reset_grad:
w_grad = self.gradientMap[_inp_blobs[0]]
_inp_blobs[2] = w_grad
# add layer to the model
with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, d)):
if kwargs:
new_layer = layer(_inp_blobs, _out_blobs, **kwargs)
else:
new_layer = layer(_inp_blobs, _out_blobs)
ll.append(new_layer)
return ll
else:
# add layer on a single device
# WARNING: reset_grad option was exlusively designed for WeightedSum
# with inp_blobs=[w, tag_one, "", lr], where "" will be replaced
if reset_grad:
w_grad = self.gradientMap[inp_blobs[0]]
inp_blobs[2] = w_grad
# add layer to the model
if kwargs:
new_layer = layer(inp_blobs, out_blobs, **kwargs)
else:
new_layer = layer(inp_blobs, out_blobs)
return new_layer
def create_mlp(self, ln, sigmoid_layer, model, tag):
(tag_layer, tag_in, tag_out) = tag
# build MLP layer by layer
layers = []
weights = []
for i in range(1, ln.size):
n = ln[i - 1]
m = ln[i]
# create tags
tag_fc_w = tag_layer + ":::" + "fc" + str(i) + "_w"
tag_fc_b = tag_layer + ":::" + "fc" + str(i) + "_b"
tag_fc_y = tag_layer + ":::" + "fc" + str(i) + "_y"
tag_fc_z = tag_layer + ":::" + "fc" + str(i) + "_z"
if i == ln.size - 1:
tag_fc_z = tag_out
weights.append(tag_fc_w)
weights.append(tag_fc_b)
# initialize the weights
# approach 1: custom Xavier input, output or two-sided fill
mean = 0.0 # std_dev = np.sqrt(variance)
std_dev = np.sqrt(2 / (m + n)) # np.sqrt(1 / m) # np.sqrt(1 / n)
W = np.random.normal(mean, std_dev, size=(m, n)).astype(np.float32)
std_dev = np.sqrt(1 / m) # np.sqrt(2 / (m + 1))
b = np.random.normal(mean, std_dev, size=m).astype(np.float32)
self.FeedBlobWrapper(tag_fc_w, W)
self.FeedBlobWrapper(tag_fc_b, b)
# approach 2: caffe2 xavier
# W = self.AddLayerWrapper(
# model.param_init_net.XavierFill,
# [],
# tag_fc_w,
# shape=[m, n]
# )
# b = self.AddLayerWrapper(
# model.param_init_net.ConstantFill,
# [],
# tag_fc_b,
# shape=[m]
# )
# initialize the MLP's momentum for the Adagrad optimizer
if self.emb_optimizer in ["adagrad", "rwsadagrad"]:
# momentum of the weights
self.FeedBlobWrapper(
"momentum_mlp_{}_{}".format(tag_layer, 2 * i - 1),
np.full((m, n), 0, dtype=np.float32)
)
# momentum of the biases
self.FeedBlobWrapper(
"momentum_mlp_{}_{}".format(tag_layer, 2 * i),
np.full((m), 0, dtype=np.float32)
)
# save the blob shapes for latter (only needed if onnx is requested)
if self.save_onnx:
self.onnx_tsd[tag_fc_w] = (onnx.TensorProto.FLOAT, W.shape)
self.onnx_tsd[tag_fc_b] = (onnx.TensorProto.FLOAT, b.shape)
# approach 1: construct fully connected operator using model.net
fc = self.AddLayerWrapper(
model.net.FC, [tag_in, tag_fc_w, tag_fc_b], tag_fc_y
)
# approach 2: construct fully connected operator using brew
# https://github.com/caffe2/tutorials/blob/master/MNIST.ipynb
# fc = brew.fc(model, layer, tag_fc_w, dim_in=m, dim_out=n)
layers.append(fc)
if i == sigmoid_layer:
# approach 1: construct sigmoid operator using model.net
layer = self.AddLayerWrapper(model.net.Sigmoid, tag_fc_y, tag_fc_z)
# approach 2: using brew (which currently does not support sigmoid)
# tag_sigm = tag_layer + ":::" + "sigmoid" + str(i)
# layer = brew.sigmoid(model,fc,tag_sigmoid)
else:
# approach 1: construct relu operator using model.net
layer = self.AddLayerWrapper(model.net.Relu, tag_fc_y, tag_fc_z)
# approach 2: using brew
# tag_relu = tag_layer + ":::" + "relu" + str(i)
# layer = brew.relu(model,fc,tag_relu)
tag_in = tag_fc_z
layers.append(layer)
# WARNING: the dependency between layers is implicit in the tags,
# so only the last layer is added to the layers list. It will
# later be used for interactions.
return layers, weights
def create_emb(self, m, ln, model, tag):
(tag_layer, tag_in, tag_out) = tag
emb_l = []
weights_l = []
vw_l = []
for i in range(0, ln.size):
n = ln[i]
# select device
if self.ndevices > 1:
d = i % self.ndevices
else:
d = -1
# create tags
on_device = "" if self.ndevices <= 1 else "gpu_" + str(d) + "/"
len_s = on_device + tag_layer + ":::" + "sls" + str(i) + "_l"
ind_s = on_device + tag_layer + ":::" + "sls" + str(i) + "_i"
tbl_s = on_device + tag_layer + ":::" + "sls" + str(i) + "_w"
sum_s = on_device + tag_layer + ":::" + "sls" + str(i) + "_z"
weights_l.append(tbl_s)
# initialize the weights
# approach 1a: custom
W = np.random.uniform(low=-np.sqrt(1 / n),
high=np.sqrt(1 / n),
size=(n, m)).astype(np.float32)
# approach 1b: numpy rand
# W = ra.rand(n, m).astype(np.float32)
self.FeedBlobWrapper(tbl_s, W, False, device_id=d)
# approach 2: caffe2 xavier
# with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, d)):
# W = model.param_init_net.XavierFill([], tbl_s, shape=[n, m])
# save the blob shapes for latter (only needed if onnx is requested)
# initialize the embedding's momentum for the Adagrad optimizer
if self.emb_optimizer == "adagrad":
self.FeedBlobWrapper("momentum_emb_{}".format(i),
np.full((n, m), 0), add_prefix=False, device_id=d)
elif self.emb_optimizer == "rwsadagrad":
self.FeedBlobWrapper("momentum_emb_{}".format(i),
np.full((n), 0), add_prefix=False, device_id=d)
if self.save_onnx:
self.onnx_tsd[tbl_s] = (onnx.TensorProto.FLOAT, W.shape)
# create operator
if self.weighted_pooling is not None:
vw_s = on_device + tag_layer + ":::" + "sls" + str(i) + "_v"
psw_s = on_device + tag_layer + ":::" + "sls" + str(i) + "_s"
VW = np.ones(n).astype(np.float32)
self.FeedBlobWrapper(vw_s, VW, False, device_id=d)
if self.weighted_pooling == "learned":
vw_l.append(vw_s)
grad_on_weights = True
else:
grad_on_weights = False
if self.save_onnx:
self.onnx_tsd[vw_s] = (onnx.TensorProto.FLOAT, VW.shape)
if self.ndevices <= 1:
PSW = model.net.Gather([vw_s, ind_s], [psw_s])
EE = model.net.SparseLengthsWeightedSum(
[tbl_s, PSW, ind_s, len_s], [sum_s],
grad_on_weights=grad_on_weights
)
else:
with core.DeviceScope(
core.DeviceOption(workspace.GpuDeviceType, d)
):
PSW = model.net.Gather([vw_s, ind_s], [psw_s])
EE = model.net.SparseLengthsWeightedSum(
[tbl_s, PSW, ind_s, len_s], [sum_s],
grad_on_weights=grad_on_weights
)
else:
if self.ndevices <= 1:
EE = model.net.SparseLengthsSum(
[tbl_s, ind_s, len_s], [sum_s]
)
else:
with core.DeviceScope(
core.DeviceOption(workspace.GpuDeviceType, d)
):
EE = model.net.SparseLengthsSum(
[tbl_s, ind_s, len_s], [sum_s]
)
emb_l.append(EE)
return emb_l, weights_l, vw_l
def create_interactions(self, x, ly, model, tag):
(tag_dense_in, tag_sparse_in, tag_int_out) = tag
if self.arch_interaction_op == "dot":
# concatenate dense and sparse features
tag_int_out_info = tag_int_out + "_info"
T, T_info = model.net.Concat(
x + ly,
[tag_int_out + "_cat_axis0", tag_int_out_info + "_cat_axis0"],
axis=1,
add_axis=1,
)
# perform a dot product
Z = model.net.BatchMatMul([T, T], tag_int_out + "_matmul", trans_b=1)
# append dense feature with the interactions (into a row vector)
# approach 1: all
# Zflat = model.net.Flatten(Z, tag_int_out + "_flatten", axis=1)
# approach 2: unique
Zflat_all = model.net.Flatten(Z, tag_int_out + "_flatten_all", axis=1)
Zflat = model.net.BatchGather(
[Zflat_all, tag_int_out + "_tril_indices"],
tag_int_out + "_flatten"
)
R, R_info = model.net.Concat(
x + [Zflat], [tag_int_out, tag_int_out_info], axis=1
)
elif self.arch_interaction_op == "cat":
# concatenation features (into a row vector)
tag_int_out_info = tag_int_out + "_info"
R, R_info = model.net.Concat(
x + ly, [tag_int_out, tag_int_out_info], axis=1
)
else:
sys.exit("ERROR: --arch-interaction-op="
+ self.arch_interaction_op + " is not supported")
return R
def create_sequential_forward_ops(self):
# embeddings
tag = (self.temb, self.tsin, self.tsout)
self.emb_l, self.emb_w, self.emb_vw = self.create_emb(
self.m_spa, self.ln_emb, self.model, tag
)
# bottom mlp
tag = (self.tbot, self.tdin, self.tdout)
self.bot_l, self.bot_w = self.create_mlp(self.ln_bot, self.sigmoid_bot,
self.model, tag)
# interactions
tag = (self.tdout, self.tsout, self.tint)
Z = self.create_interactions([self.bot_l[-1]], self.emb_l, self.model, tag)
# top mlp
tag = (self.ttop, Z, self.tout)
self.top_l, self.top_w = self.create_mlp(self.ln_top, self.sigmoid_top,
self.model, tag)
# debug prints
# print(self.emb_l)
# print(self.bot_l)
# print(self.top_l)
# setup the last output variable
self.last_output = self.top_l[-1]
def create_parallel_forward_ops(self):
# distribute embeddings (model parallelism)
tag = (self.temb, self.tsin, self.tsout)
self.emb_l, self.emb_w, self.emb_vw = self.create_emb(
self.m_spa, self.ln_emb, self.model, tag
)
# replicate mlp (data parallelism)
tag = (self.tbot, self.tdin, self.tdout)
self.bot_l, self.bot_w = self.create_mlp(self.ln_bot, self.sigmoid_bot,
self.model, tag)
# add communication (butterfly shuffle)
t_list = []
for i, emb_output in enumerate(self.emb_l):
# split input
src_d = i % self.ndevices
lo = [emb_output + "_split_" + str(d) for d in range(self.ndevices)]
# approach 1: np and caffe2 operators assume the mini-batch size is
# divisible exactly by the number of available devices
with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, src_d)):
self.model.net.Split(emb_output, lo, axis=0)
"""
# approach 2: np and caffe2 operators do not assume exact divisibility
ls = where_to_split(args.mini_batch_size, self.ndevices, _add_leftover=True)
with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, src_d)):
emb_output_split = self.model.net.Split(
emb_output, lo, split=lp, axis=0
)
"""
# scatter
y = []
for dst_d in range(len(lo)):
src_blob = lo[dst_d]
dst_blob = str(src_blob).replace(
"gpu_" + str(src_d), "gpu_" + str(dst_d), 1
)
if src_blob != dst_blob:
with core.DeviceScope(
core.DeviceOption(workspace.GpuDeviceType, dst_d)
):
blob = self.model.Copy(src_blob, dst_blob)
else:
blob = dst_blob
y.append(blob)
t_list.append(y)
# adjust lists to be ordered per device
x = list(map(lambda x: list(x), zip(*self.bot_l)))
ly = list(map(lambda y: list(y), zip(*t_list)))
# interactions
for d in range(self.ndevices):
on_device = "gpu_" + str(d) + "/"
tag = (on_device + self.tdout, on_device + self.tsout, on_device + self.tint)
with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, d)):
self.create_interactions([x[d][-1]], ly[d], self.model, tag)
# replicate mlp (data parallelism)
tag = (self.ttop, self.tint, self.tout)
self.top_l, self.top_w = self.create_mlp(self.ln_top, self.sigmoid_top,
self.model, tag)
# debug prints
# print(self.model.net.Proto(),end='\n')
# sys.exit("ERROR: debugging")
# setup the last output variable
self.last_output = self.top_l[-1]
def __init__(
self,
m_spa,
ln_emb,
ln_bot,
ln_top,
arch_interaction_op,
arch_interaction_itself=False,
sigmoid_bot=-1,
sigmoid_top=-1,
save_onnx=False,
model=None,
test_net=None,
tag=None,
ndevices=-1,
forward_ops=True,
enable_prof=False,
weighted_pooling=None,
emb_optimizer="sgd"
):
super(DLRM_Net, self).__init__()
# init model
if model is None:
global_init_opt = ["caffe2", "--caffe2_log_level=0"]
if enable_prof:
global_init_opt += [
"--logtostderr=0",
"--log_dir=$HOME",
"--caffe2_logging_print_net_summary=1",
]
workspace.GlobalInit(global_init_opt)
self.set_tags()
self.model = model_helper.ModelHelper(name="DLRM", init_params=True)
self.test_net = None
else:
# WARNING: assume that workspace and tags have been initialized elsewhere
self.set_tags(tag[0], tag[1], tag[2], tag[3], tag[4], tag[5], tag[6],
tag[7], tag[8], tag[9])
self.model = model
self.test_net = test_net
# save arguments
self.m_spa = m_spa
self.ln_emb = ln_emb
self.ln_bot = ln_bot
self.ln_top = ln_top
self.arch_interaction_op = arch_interaction_op
self.arch_interaction_itself = arch_interaction_itself
self.sigmoid_bot = sigmoid_bot
self.sigmoid_top = sigmoid_top
self.save_onnx = save_onnx
self.ndevices = ndevices
self.emb_optimizer = emb_optimizer
if weighted_pooling is not None and weighted_pooling != "fixed":
self.weighted_pooling = "learned"
else:
self.weighted_pooling = weighted_pooling
# onnx types and shapes dictionary
if self.save_onnx:
self.onnx_tsd = {}
# create forward operators
if forward_ops:
if self.ndevices <= 1:
return self.create_sequential_forward_ops()
else:
return self.create_parallel_forward_ops()
def set_tags(
self,
_tag_layer_top_mlp="top",
_tag_layer_bot_mlp="bot",
_tag_layer_embedding="emb",
_tag_feature_dense_in="dense_in",
_tag_feature_dense_out="dense_out",
_tag_feature_sparse_in="sparse_in",
_tag_feature_sparse_out="sparse_out",
_tag_interaction="interaction",
_tag_dense_output="prob_click",
_tag_dense_target="target",
):
# layer tags
self.ttop = _tag_layer_top_mlp
self.tbot = _tag_layer_bot_mlp
self.temb = _tag_layer_embedding
# dense feature tags
self.tdin = _tag_feature_dense_in
self.tdout = _tag_feature_dense_out
# sparse feature tags
self.tsin = _tag_feature_sparse_in
self.tsout = _tag_feature_sparse_out
# output and target tags
self.tint = _tag_interaction
self.ttar = _tag_dense_target
self.tout = _tag_dense_output
def parameters(self):
return self.model
def get_loss(self):
return self.FetchBlobWrapper(self.loss, reduce_across="add")
def get_output(self):
return self.FetchBlobWrapper(self.last_output, reduce_across="concat")
def create(self, X, S_lengths, S_indices, T):
self.create_input(X, S_lengths, S_indices, T)
self.create_model(X, S_lengths, S_indices, T)
def create_input(self, X, S_lengths, S_indices, T):
# feed input data to blobs
self.FeedBlobWrapper(self.tdin, X, split=True)
# save the blob shapes for latter (only needed if onnx is requested)
if self.save_onnx:
self.onnx_tsd[self.tdin] = (onnx.TensorProto.FLOAT, X.shape)
for i in range(len(self.emb_l)):
# select device
if self.ndevices > 1:
d = i % self.ndevices
else:
d = -1
# create tags
on_device = "" if self.ndevices <= 1 else "gpu_" + str(d) + "/"
len_s = on_device + self.temb + ":::" + "sls" + str(i) + "_l"
ind_s = on_device + self.temb + ":::" + "sls" + str(i) + "_i"
self.FeedBlobWrapper(len_s, np.array(S_lengths[i]), False, device_id=d)
self.FeedBlobWrapper(ind_s, np.array(S_indices[i]), False, device_id=d)
# save the blob shapes for latter (only needed if onnx is requested)
if self.save_onnx:
lshape = (len(S_lengths[i]),) # =args.mini_batch_size
ishape = (len(S_indices[i]),)
self.onnx_tsd[len_s] = (onnx.TensorProto.INT32, lshape)
self.onnx_tsd[ind_s] = (onnx.TensorProto.INT32, ishape)
# feed target data to blobs
if T is not None:
zeros_fp32 = np.zeros(T.shape).astype(np.float32)
self.FeedBlobWrapper(self.ttar, zeros_fp32, split=True)
# save the blob shapes for latter (only needed if onnx is requested)
if self.save_onnx:
self.onnx_tsd[self.ttar] = (onnx.TensorProto.FLOAT, T.shape)
def create_model(self, X, S_lengths, S_indices, T):
#setup tril indices for the interactions
offset = 1 if self.arch_interaction_itself else 0
num_fea = len(self.emb_l) + 1
tril_indices = np.array([j + i * num_fea
for i in range(num_fea) for j in range(i + offset)])
self.FeedBlobWrapper(self.tint + "_tril_indices", tril_indices)
# create compute graph
if T is not None:
# WARNING: RunNetOnce call is needed only if we use brew and ConstantFill.
# We could use direct calls to self.model functions above to avoid it
workspace.RunNetOnce(self.model.param_init_net)
workspace.CreateNet(self.model.net)
if self.test_net is not None:
workspace.CreateNet(self.test_net)
def run(self, X, S_lengths, S_indices, T, test_net=False, enable_prof=False):
# feed input data to blobs
# dense features
self.FeedBlobWrapper(self.tdin, X, split=True)
# sparse features
for i in range(len(self.emb_l)):
# select device
if self.ndevices > 1:
d = i % self.ndevices
else:
d = -1
# create tags
on_device = "" if self.ndevices <= 1 else "gpu_" + str(d) + "/"
len_s = on_device + self.temb + ":::" + "sls" + str(i) + "_l"
ind_s = on_device + self.temb + ":::" + "sls" + str(i) + "_i"
self.FeedBlobWrapper(len_s, np.array(S_lengths[i]), False, device_id=d)
self.FeedBlobWrapper(ind_s, np.array(S_indices[i]), False, device_id=d)
# feed target data to blobs if needed
if T is not None:
self.FeedBlobWrapper(self.ttar, T, split=True)
# execute compute graph
if test_net:
workspace.RunNet(self.test_net)
else:
if enable_prof:
workspace.C.benchmark_net(self.model.net.Name(), 0, 1, True)
else:
workspace.RunNet(self.model.net)
# debug prints
# print("intermediate")
# print(self.FetchBlobWrapper(self.bot_l[-1]))
# for tag_emb in self.emb_l:
# print(self.FetchBlobWrapper(tag_emb))
# print(self.FetchBlobWrapper(self.tint))
def MSEloss(self, scale=1.0):
# add MSEloss to the model
self.AddLayerWrapper(self.model.SquaredL2Distance, [self.tout, self.ttar], "sd")
self.AddLayerWrapper(self.model.Scale, "sd", "sd2", scale=2.0 * scale)
# WARNING: "loss" is a special tag and should not be changed
self.loss = self.AddLayerWrapper(self.model.AveragedLoss, "sd2", "loss")
def BCEloss(self, scale=1.0, threshold=0.0):
# add BCEloss to the mode
if 0.0 < threshold and threshold < 1.0:
self.AddLayerWrapper(self.model.Clip, self.tout, "tout_c",
min=threshold, max=(1.0 - threshold))
self.AddLayerWrapper(self.model.MakeTwoClass, "tout_c", "tout_2c")
else:
self.AddLayerWrapper(self.model.MakeTwoClass, self.tout, "tout_2c")
self.AddLayerWrapper(self.model.LabelCrossEntropy, ["tout_2c", self.ttar], "sd")
# WARNING: "loss" is a special tag and should not be changed
if scale == 1.0:
self.loss = self.AddLayerWrapper(self.model.AveragedLoss, "sd", "loss")
else:
self.AddLayerWrapper(self.model.Scale, "sd", "sd2", scale=scale)
self.loss = self.AddLayerWrapper(self.model.AveragedLoss, "sd2", "loss")
def sgd_optimizer(self, learning_rate,
T=None, _gradientMap=None, sync_dense_params=True):
# create one, it and lr tags (or use them if already present)
if T is not None:
(tag_one, tag_it, tag_lr) = T
else:
(tag_one, tag_it, tag_lr) = ("const_one", "optim_it", "optim_lr")
# approach 1: feed values directly
# self.FeedBlobWrapper(tag_one, np.ones(1).astype(np.float32))
# self.FeedBlobWrapper(tag_it, np.zeros(1).astype(np.int64))
# it = self.AddLayerWrapper(self.model.Iter, tag_it, tag_it)
# lr = self.AddLayerWrapper(self.model.LearningRate, tag_it, tag_lr,
# base_lr=-1 * learning_rate, policy="fixed")
# approach 2: use brew
self.AddLayerWrapper(self.model.param_init_net.ConstantFill,
[], tag_one, shape=[1], value=1.0)
self.AddLayerWrapper(brew.iter, self.model, tag_it)
self.AddLayerWrapper(self.model.LearningRate, tag_it, tag_lr,
base_lr=-1 * learning_rate, policy="fixed")
# save the blob shapes for latter (only needed if onnx is requested)
if self.save_onnx:
self.onnx_tsd[tag_one] = (onnx.TensorProto.FLOAT, (1,))
self.onnx_tsd[tag_it] = (onnx.TensorProto.INT64, (1,))
# create gradient maps (or use them if already present)
if _gradientMap is not None:
self.gradientMap = _gradientMap
else:
if self.loss.__class__ == list:
self.gradientMap = self.model.AddGradientOperators(self.loss)
else:
self.gradientMap = self.model.AddGradientOperators([self.loss])
# update weights
# approach 1: builtin function
# optimizer.build_sgd(self.model, base_learning_rate=learning_rate)
# approach 2: custom code
# top MLP weight and bias
for w in self.top_w:
# allreduce across devices if needed
if sync_dense_params and self.ndevices > 1:
grad_blobs = [
self.gradientMap["gpu_{}/".format(d) + w]
for d in range(self.ndevices)
]
self.model.NCCLAllreduce(grad_blobs, grad_blobs)
# update weights
self.AddLayerWrapper(self.model.WeightedSum,
[w, tag_one, "", tag_lr], w, reset_grad=True)
# bottom MLP weight and bias
for w in self.bot_w:
# allreduce across devices if needed
if sync_dense_params and self.ndevices > 1:
grad_blobs = [
self.gradientMap["gpu_{}/".format(d) + w]
for d in range(self.ndevices)
]
self.model.NCCLAllreduce(grad_blobs, grad_blobs)
# update weights
self.AddLayerWrapper(self.model.WeightedSum,
[w, tag_one, "", tag_lr], w, reset_grad=True)
# update embeddings
for i, w in enumerate(self.emb_w):
# select device
if self.ndevices > 1:
d = i % self.ndevices
# create tags
on_device = "" if self.ndevices <= 1 else "gpu_" + str(d) + "/"
_tag_one = on_device + tag_one
_tag_lr = on_device + tag_lr
# pickup gradient
w_grad = self.gradientMap[w]
# update weights
if self.ndevices > 1:
with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, d)):
self.model.ScatterWeightedSum([w, _tag_one, w_grad.indices,
w_grad.values, _tag_lr], w)
else:
self.model.ScatterWeightedSum([w, _tag_one, w_grad.indices,
w_grad.values, _tag_lr], w)
# update per sample weights
if self.weighted_pooling == "learned":
for i, w in enumerate(self.emb_vw):
# select device
if self.ndevices > 1:
d = i % self.ndevices
# create tags
on_device = "" if self.ndevices <= 1 else "gpu_" + str(d) + "/"
_tag_one = on_device + tag_one
_tag_lr = on_device + tag_lr
# pickup gradient
w_grad = self.gradientMap[w]
# update weights
if self.ndevices > 1:
with core.DeviceScope(
core.DeviceOption(workspace.GpuDeviceType, d)
):
self.model.ScatterWeightedSum(
[w, _tag_one, w_grad.indices,
w_grad.values, _tag_lr], w
)
else:
self.model.ScatterWeightedSum(
[w, _tag_one, w_grad.indices, w_grad.values, _tag_lr], w
)
def adagrad_optimizer(self, learning_rate,
T=None, _gradientMap=None, sync_dense_params=True,
epsilon=1e-10, decay_=0.0, weight_decay_=0.0):
# create one, it and lr tags (or use them if already present)
if T is not None:
(tag_one, tag_it, tag_lr) = T
else:
(tag_one, tag_it, tag_lr) = ("const_one", "optim_it", "optim_lr")
# approach 1: feed values directly
# self.FeedBlobWrapper(tag_one, np.ones(1).astype(np.float32))
# self.FeedBlobWrapper(tag_it, np.zeros(1).astype(np.int64))
# it = self.AddLayerWrapper(self.model.Iter, tag_it, tag_it)
# lr = self.AddLayerWrapper(self.model.LearningRate, tag_it, tag_lr,
# base_lr=-1 * learning_rate, policy="fixed")
# approach 2: use brew
self.AddLayerWrapper(self.model.param_init_net.ConstantFill,
[], tag_one, shape=[1], value=1.0)
self.AddLayerWrapper(brew.iter, self.model, tag_it)
self.AddLayerWrapper(self.model.LearningRate, tag_it, tag_lr,
base_lr=-1 * learning_rate, policy="fixed")
# save the blob shapes for latter (only needed if onnx is requested)
if self.save_onnx:
self.onnx_tsd[tag_one] = (onnx.TensorProto.FLOAT, (1,))
self.onnx_tsd[tag_it] = (onnx.TensorProto.INT64, (1,))
# create gradient maps (or use them if already present)
if _gradientMap is not None:
self.gradientMap = _gradientMap
else:
if self.loss.__class__ == list:
self.gradientMap = self.model.AddGradientOperators(self.loss)
else:
self.gradientMap = self.model.AddGradientOperators([self.loss])
# update weights
# approach 1: builtin function
# optimizer.build_sgd(self.model, base_learning_rate=learning_rate)
# approach 2: custom code
# top MLP weight and bias
for i, w in enumerate(self.top_w):
# allreduce across devices if needed
if sync_dense_params and self.ndevices > 1:
grad_blobs = [
self.gradientMap["gpu_{}/".format(d) + w]
for d in range(self.ndevices)
]
self.model.NCCLAllreduce(grad_blobs, grad_blobs)
# update weights
self.model.Adagrad(
[
w,
"momentum_mlp_top_{}".format(i + 1),
self.gradientMap[w],
tag_lr
],
[w, "momentum_mlp_top_{}".format(i + 1)],
epsilon=epsilon,
decay_=decay_,
weight_decay_=weight_decay_
)
# bottom MLP weight and bias
for i, w in enumerate(self.bot_w):
# allreduce across devices if needed
if sync_dense_params and self.ndevices > 1:
grad_blobs = [
self.gradientMap["gpu_{}/".format(d) + w]
for d in range(self.ndevices)
]
self.model.NCCLAllreduce(grad_blobs, grad_blobs)
# update weights
self.model.Adagrad(
[
w,
"momentum_mlp_bot_{}".format(i + 1),
self.gradientMap[w],
tag_lr
],
[w, "momentum_mlp_bot_{}".format(i + 1)],
epsilon=epsilon,
decay_=decay_,
weight_decay_=weight_decay_
)
# update embeddings
for i, w in enumerate(self.emb_w):
# select device
if self.ndevices > 1:
d = i % self.ndevices
# create tags
on_device = "" if self.ndevices <= 1 else "gpu_" + str(d) + "/"
_tag_one = on_device + tag_one
_tag_lr = on_device + tag_lr
# pickup gradient
w_grad = self.gradientMap[w]
# update weights
def add_optimizer():
self.model.Unique(
w_grad.indices,
["unique_w_grad_indices", "remapping_w_grad_indices"]
)
self.model.UnsortedSegmentSum(
[w_grad.values, "remapping_w_grad_indices"],
"unique_w_grad_values"
)
if self.emb_optimizer == "adagrad":
self.model.SparseAdagrad(
[
w,
"momentum_emb_{}".format(i),
"unique_w_grad_indices",
"unique_w_grad_values",
_tag_lr
],
[w, "momentum_emb_{}".format(i)],
epsilon=epsilon,
decay_=decay_,
weight_decay_=weight_decay_
)
elif self.emb_optimizer == "rwsadagrad":
self.model.RowWiseSparseAdagrad(
[
w,
"momentum_emb_{}".format(i),
"unique_w_grad_indices",
"unique_w_grad_values",
_tag_lr
],
[w, "momentum_emb_{}".format(i)],
epsilon=epsilon,
decay_=decay_,
weight_decay_=weight_decay_
)
if self.ndevices > 1:
with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, d)):
add_optimizer()
else:
add_optimizer()