-
Notifications
You must be signed in to change notification settings - Fork 4.3k
/
distributed_multi_learner_test.py
270 lines (215 loc) · 11.7 KB
/
distributed_multi_learner_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
"""
This test extends the bmuf_metrics_aggregation_test and tests multiple learners in the distributed training.
"""
import pytest
import cntk
import numpy as np
import sys, os
import argparse
import re
import platform
sys.path.append(os.path.dirname(__file__))
cntk.cntk_py.set_fixed_random_seed(1)
from distributed_learner_test import mpiexec_execute
from bmuf_metrics_aggregation_test import get_minibatch
feat_dim = 5
label_dim = 3
cell_dim = 5
seq_len = 20
num_batches = 101
progress_freq =10
class SingleDataParallelTrainer():
def __init__(self, frame_mode=False):
self.create_model(frame_mode)
self.create_trainer()
def create_model(self, frame_mode=False):
if frame_mode:
self.feat = cntk.input_variable(shape=(feat_dim,))
self.label = cntk.input_variable((label_dim,))
net = cntk.layers.Sequential([cntk.layers.Dense(cell_dim), cntk.layers.Dense(label_dim)])
self.output = net(self.feat)
else:
#sequence mode
self.feat = cntk.sequence.input_variable(shape=(feat_dim,))
self.label = cntk.sequence.input_variable((label_dim,))
net = cntk.layers.Sequential([cntk.layers.Recurrence(cntk.layers.LSTM(shape=label_dim, cell_shape=(cell_dim,)))])
self.output = net(self.feat)
self.ce = cntk.cross_entropy_with_softmax(self.output, self.label)
self.err = cntk.classification_error(self.output, self.label)
def create_trainer(self):
try:
lr_per_sample = cntk.learning_parameter_schedule_per_sample(0.007)
learner = cntk.data_parallel_distributed_learner(cntk.sgd(self.output.parameters, lr_per_sample))
comm_rank = cntk.distributed.Communicator.rank()
self.trainer = cntk.Trainer(self.output, (self.ce, self.err), [learner], [cntk.logging.ProgressPrinter(freq=progress_freq, tag="Training", rank=comm_rank)])
except RuntimeError:
self.trainer = None
return
class TwoDataParallelTrainer(SingleDataParallelTrainer):
def __init__(self, frame_mode=False):
SingleDataParallelTrainer.__init__(self, frame_mode)
def create_trainer(self):
try:
lr_per_sample = cntk.learning_parameter_schedule_per_sample(0.007)
p = self.output.parameters
# Three of four parameters are learned by first data_parallel_distributed_learner.
learner1 = cntk.data_parallel_distributed_learner(cntk.sgd([p[0],p[1],p[2]], lr_per_sample))
# New API to mark which learner is to use for metric aggregaion.
learner1.set_as_metric_aggregator()
# The last parameter is learned by another data_parallel_distributed_learner.
learner2 = cntk.data_parallel_distributed_learner(cntk.sgd([p[3]], lr_per_sample))
comm_rank = cntk.distributed.Communicator.rank()
self.trainer = cntk.Trainer(self.output, (self.ce, self.err), [learner1, learner2], [cntk.logging.ProgressPrinter(freq=progress_freq, tag="Training", rank=comm_rank)])
except RuntimeError:
self.trainer = None
return
class MultiLearnerTrainer(SingleDataParallelTrainer):
def __init__(self, frame_mode=False):
SingleDataParallelTrainer.__init__(self, frame_mode)
def create_trainer(self):
try:
p = self.output.parameters
# Three of four parameters are learned by block_momentum_distributed_learner.
bmd_learner = cntk.block_momentum_distributed_learner(cntk.momentum_sgd([p[0],p[1],p[2]], cntk.learning_parameter_schedule(0.0001), cntk.momentum_as_time_constant_schedule(1000)),
block_size=1000, block_learning_rate=0.01, block_momentum_as_time_constant=1000)
# New API to mark which learner is to use for metric aggregaion.
bmd_learner.set_as_metric_aggregator()
# The last parameter is learned by the data_parallel_distributed_learner.
momentum_schedule = cntk.momentum_schedule_per_sample(0.9990913221888589)
lr_per_sample = cntk.learning_parameter_schedule_per_sample(0.007)
dpd_learner = cntk.data_parallel_distributed_learner(cntk.momentum_sgd([p[3]], lr_per_sample, momentum_schedule, True))
comm_rank = cntk.distributed.Communicator.rank()
self.trainer = cntk.Trainer(self.output, (self.ce, self.err), [bmd_learner, dpd_learner], [cntk.logging.ProgressPrinter(freq=progress_freq, tag="Training", rank=comm_rank)])
except RuntimeError:
self.trainer = None
return
def mpi_worker_multi_learner(trainer, working_dir, checkpoint_dir, mb_source):
comm_rank = cntk.distributed.Communicator.rank()
np.random.seed(comm_rank)
num_paritions = cntk.Communicator.num_workers();
partition_index = cntk.Communicator.rank();
checkpoint_performed = False
for i, data in enumerate(get_minibatch(trainer, working_dir, mb_source, num_paritions, partition_index)):
trainer.trainer.train_minibatch(data)
if i % 50 == 0:
trainer.trainer.summarize_training_progress()
if not checkpoint_performed and not checkpoint_dir == "":
trainer.trainer.save_checkpoint(checkpoint_dir)
trainer.trainer.restore_from_checkpoint(checkpoint_dir)
checkpoint_performed = True
def get_loss_perepoch_perworker(log_line, num_workers):
# [0]Finished Epoch[1]: [Training] loss = 1.663636 * 10, metric = 52.40% * 10 0.890s ( 11.2 samples/s);
regex_pattern = r"\[(?P<worker_rank>\d)\].*? Epoch\[(?P<epoch>\d+)\].*? loss = (?P<loss>\d+\.\d+) \* (?P<samples>\d+).*? metric = (?P<metric>\d+\.\d+)"
loss_perepoch_perworker = {i:{} for i in range(num_workers)}
for match in re.finditer(regex_pattern, log_line):
rank = int(match.groupdict()["worker_rank"])
epoch = int(match.groupdict()["epoch"])
loss = match.groupdict()["loss"]
metric = match.groupdict()["metric"]
samples = int(match.groupdict()["samples"])
loss_perepoch_perworker[rank].update({epoch:(loss, metric, samples)})
return loss_perepoch_perworker
MB_SOURCES = ["ctf_frame"]
@pytest.mark.parametrize("mb_source", MB_SOURCES)
def test_single_data_parallel_learner_vs_two_data_parallel_learners(tmpdir, device_id, mb_source):
if platform.system() == 'Linux':
pytest.skip('test only runs on Windows due to mpiexec -l option')
launch_args = []
launch_args += ["--outputdir", str(tmpdir)]
launch_args += ["--mb_source", mb_source]
launch_args += ["--trainer_type", "single"]
num_workers = 1 # use a single worker.
ret_str = mpiexec_execute(__file__, ['-n', str(num_workers), '-l'], launch_args)
print(ret_str)
loss_perepoch_perworker = get_loss_perepoch_perworker(ret_str, num_workers)
loss_per_worker = loss_perepoch_perworker.values()
single_learner_loss_per_worker_epochsort = []
for epoch_losses in loss_per_worker:
single_learner_loss_per_worker_epochsort.append([epoch_losses[i] for i in sorted(epoch_losses)])
launch_args = []
launch_args += ["--outputdir", str(tmpdir)]
launch_args += ["--mb_source", mb_source]
launch_args += ["--trainer_type", "two"]
num_workers = 2 # now run in distributed workers.
ret_str = mpiexec_execute(__file__, ['-n', str(num_workers), '-l'], launch_args)
print(ret_str)
loss_perepoch_perworker = get_loss_perepoch_perworker(ret_str, num_workers)
loss_per_worker = loss_perepoch_perworker.values()
multi_learner_loss_per_worker_epochsort = []
for epoch_losses in loss_per_worker:
multi_learner_loss_per_worker_epochsort.append([epoch_losses[i] for i in sorted(epoch_losses)])
assert all([single_learner_loss_per_worker_epochsort[0] == i for i in multi_learner_loss_per_worker_epochsort])
MB_SOURCES = ["ctf_frame"]
@pytest.mark.parametrize("mb_source", MB_SOURCES)
def test_multi_learner_bmuf_correct_metrics_averaging(tmpdir, device_id, mb_source):
if platform.system() == 'Linux':
pytest.skip('test only runs on Windows due to mpiexec -l option')
num_workers = 2
# check whether trainer can be initialized or not
bmuf = MultiLearnerTrainer()
if not bmuf.trainer:
pytest.skip('BMUF not available on this build')
launch_args = []
launch_args += ["--outputdir", str(tmpdir)]
launch_args += ["--mb_source", mb_source]
launch_args += ["--trainer_type", "multi"]
ret_str = mpiexec_execute(__file__, ['-n', str(num_workers), '-l'], launch_args)
print(ret_str)
loss_perepoch_perworker = get_loss_perepoch_perworker(ret_str, num_workers)
num_epochs_per_worker = list(map(len,loss_perepoch_perworker.values()))
#assert that data exists
assert len(num_epochs_per_worker) != 0
#assert that number of epochs isn't zero for 1st worker.
assert num_epochs_per_worker[0] != 0
# assert all workers have same number of epochs
assert min(num_epochs_per_worker) == max(num_epochs_per_worker)
# assert all workers have same loss and metric values
loss_per_worker = loss_perepoch_perworker.values()
loss_per_worker_epochsort = []
for epoch_losses in loss_per_worker:
loss_per_worker_epochsort.append([epoch_losses[i] for i in sorted(epoch_losses)])
assert all([loss_per_worker_epochsort[0] == i for i in loss_per_worker_epochsort])
# Do the same test with checkpoint and compare the results.
launch_args += ["--checkpointdir", str(tmpdir.join('checkpoint'))]
ret_str = mpiexec_execute(__file__, ['-n', str(num_workers), '-l'], launch_args)
print(ret_str)
loss_perepoch_perworker = get_loss_perepoch_perworker(ret_str, num_workers)
num_epochs_per_worker = list(map(len,loss_perepoch_perworker.values()))
#assert that data exists
assert len(num_epochs_per_worker) != 0
#assert that number of epochs isn't zero for 1st worker.
assert num_epochs_per_worker[0] != 0
# assert all workers have same number of epochs
assert min(num_epochs_per_worker) == max(num_epochs_per_worker)
# assert all workers have same loss and metric values
loss_per_worker = loss_perepoch_perworker.values()
multi_learner_loss_per_worker_epochsort = []
for epoch_losses in loss_per_worker:
multi_learner_loss_per_worker_epochsort.append([epoch_losses[i] for i in sorted(epoch_losses)])
# Compare no checkpoint loss, matric, and num samples, to checkpoint loss values.
for i in multi_learner_loss_per_worker_epochsort:
for n in range(3):
for m in range(3):
assert np.allclose(float(loss_per_worker_epochsort[0][n][m]), float(i[n][m]))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-outputdir', '--outputdir')
parser.add_argument('-checkpointdir', '--checkpointdir')
parser.add_argument('-mb_source', '--mb_source')
parser.add_argument("-trainer_type","--trainer_type")
args = vars(parser.parse_args())
frame_mode = (args["mb_source"] == "ctf_frame")
if args["trainer_type"] == "multi":
trainer = MultiLearnerTrainer(frame_mode)
if args["checkpointdir"]:
mpi_worker_multi_learner(trainer, args["outputdir"], args["checkpointdir"], args["mb_source"])
else:
mpi_worker_multi_learner(trainer, args["outputdir"], "", args["mb_source"])
elif args["trainer_type"] == "two":
trainer = TwoDataParallelTrainer(frame_mode)
mpi_worker_multi_learner(trainer, args["outputdir"], "", args["mb_source"])
elif args["trainer_type"] == "single":
print("Coming to a single learner")
trainer = SingleDataParallelTrainer(frame_mode)
mpi_worker_multi_learner(trainer, args["outputdir"], "", args["mb_source"])
cntk.distributed.Communicator.finalize()