|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import time |
|
5
|
|
|
import os |
|
6
|
|
|
from collections import OrderedDict |
|
7
|
|
|
import numpy as np |
|
8
|
|
|
|
|
9
|
|
|
from deepy.trainers import GeneralNeuralTrainer |
|
10
|
|
|
|
|
11
|
|
|
import logging |
|
12
|
|
|
|
|
13
|
|
|
class MultiGPUTrainer(GeneralNeuralTrainer): |
|
14
|
|
|
""" |
|
15
|
|
|
General neural network trainer. |
|
16
|
|
|
""" |
|
17
|
|
|
|
|
18
|
|
|
|
|
19
|
|
|
def __init__(self, |
|
20
|
|
|
network, config=None, method='sgd', |
|
21
|
|
|
server_port=5567, |
|
22
|
|
|
start_halving_at=6, end_at=10, sync_freq=3, |
|
23
|
|
|
valid_freq=1500, learning_rate=None, halving_freq=1, |
|
24
|
|
|
using_easgd=True |
|
25
|
|
|
): |
|
26
|
|
|
super(MultiGPUTrainer, self).__init__(network, method, config) |
|
27
|
|
|
self._report_time = False |
|
28
|
|
|
self._port = server_port |
|
29
|
|
|
self.logger = logging.getLogger('MultiGPUTrainingWorker') |
|
30
|
|
|
self.epoch = 0 |
|
31
|
|
|
self._using_easgd = using_easgd |
|
32
|
|
|
if not learning_rate: |
|
33
|
|
|
learning_rate = float(self.config.learning_rate.get_value()) |
|
34
|
|
|
self._schedule_params = { |
|
35
|
|
|
'learning_rate': learning_rate, |
|
36
|
|
|
'start_halving_at': start_halving_at, |
|
37
|
|
|
'end_at': end_at, |
|
38
|
|
|
'sync_freq': sync_freq, |
|
39
|
|
|
'valid_freq': valid_freq, |
|
40
|
|
|
'halving_freq': halving_freq |
|
41
|
|
|
} |
|
42
|
|
|
|
|
43
|
|
|
def create_param_map(self): |
|
44
|
|
|
param_map = OrderedDict() |
|
45
|
|
|
for i, param in enumerate(self.training_params()): |
|
46
|
|
|
param_map["param_{}".format(i)] = param |
|
47
|
|
|
return param_map |
|
48
|
|
|
|
|
49
|
|
|
def sync_hyperparams(self, param_map): |
|
50
|
|
|
self.logger.info("(proc {}) sync hyperparameters".format(os.getpid())) |
|
51
|
|
|
if 'epoch' in param_map: |
|
52
|
|
|
self.epoch = param_map['epoch'] |
|
53
|
|
|
if 'learning_rate' in param_map: |
|
54
|
|
|
self.config.learning_rate.set_value(param_map['learning_rate']) |
|
55
|
|
|
|
|
56
|
|
|
def fix_costs(self): |
|
57
|
|
|
self.last_run_costs = [(a, float(b)) for (a,b) in self.last_run_costs] |
|
58
|
|
|
|
|
59
|
|
|
def train(self, train_set, valid_set=None, test_set=None, train_size=None): |
|
60
|
|
|
""" |
|
61
|
|
|
Train the model in multi-GPU environment. |
|
62
|
|
|
""" |
|
63
|
|
|
from platoon.channel import Worker |
|
64
|
|
|
from platoon.param_sync import EASGD, ASGD |
|
65
|
|
|
server_port = self._port |
|
66
|
|
|
param_map = self.create_param_map() |
|
67
|
|
|
# Initialize the worker |
|
68
|
|
|
worker = Worker(control_port=server_port) |
|
69
|
|
|
if self.config.learning_rate: |
|
70
|
|
|
worker.send_req({'init_schedule': self._schedule_params}) |
|
71
|
|
|
self.sync_hyperparams(worker.send_req('sync_hyperparams')['sync_hyperparams']) |
|
72
|
|
|
easgd_alpha = worker.send_req('get_easgd_alpha') |
|
73
|
|
|
if self._using_easgd: |
|
74
|
|
|
self.logger.info("using EASGD with alpha={}".format(easgd_alpha)) |
|
75
|
|
|
else: |
|
76
|
|
|
self.logger.info("using ASGD rule") |
|
77
|
|
|
rule = EASGD(easgd_alpha) if self._using_easgd else ASGD() |
|
78
|
|
|
worker.init_shared_params(param_map.values(), param_sync_rule=rule) |
|
79
|
|
|
worker.send_req({ |
|
80
|
|
|
"set_names": None, |
|
81
|
|
|
"training_names": self.training_names, |
|
82
|
|
|
"evaluation_names": self.evaluation_names |
|
83
|
|
|
}) |
|
84
|
|
|
# Load all training batches, consume vast memory here |
|
85
|
|
|
self.logger.info("started process {}".format(os.getpid())) |
|
86
|
|
|
self.logger.info("(proc {}) load training data".format(os.getpid())) |
|
87
|
|
|
train_batches = list(train_set) |
|
88
|
|
|
network_callback = bool(self.network.training_callbacks) |
|
89
|
|
|
trainer_callback = bool(self._iter_controllers) |
|
90
|
|
|
# Start from valid, so the performance when a worked join can be known |
|
91
|
|
|
worker.copy_to_local() |
|
92
|
|
|
if valid_set: |
|
93
|
|
|
self._run_valid(self.epoch, valid_set, dry_run=True) |
|
94
|
|
|
self.fix_costs() |
|
95
|
|
|
worker.send_req({ |
|
96
|
|
|
"valid_done": None, |
|
97
|
|
|
"valid_costs": self.last_run_costs, |
|
98
|
|
|
"auto_save": self.config.auto_save |
|
99
|
|
|
}) |
|
100
|
|
|
worker.copy_to_local() |
|
101
|
|
|
# Begin the loop |
|
102
|
|
|
while True: |
|
103
|
|
|
resp = worker.send_req('next') |
|
104
|
|
|
if resp == 'stop': |
|
105
|
|
|
break |
|
106
|
|
|
elif resp == 'wait': |
|
107
|
|
|
time.sleep(1) |
|
108
|
|
|
elif resp == 'get_num_batches': |
|
109
|
|
|
worker.send_req({'get_num_batches_done': len(train_batches)}) |
|
110
|
|
|
elif 'eval' in resp: |
|
111
|
|
|
self.best_cost = resp['best_valid_cost'] |
|
112
|
|
|
worker.copy_to_local() |
|
113
|
|
|
valid_costs = None |
|
114
|
|
|
test_costs = None |
|
115
|
|
|
if valid_set: |
|
116
|
|
|
self._run_valid(self.epoch, valid_set) |
|
117
|
|
|
self.fix_costs() |
|
118
|
|
|
valid_costs = self.last_run_costs |
|
119
|
|
|
if test_set: |
|
120
|
|
|
self._run_test(self.epoch, test_set) |
|
121
|
|
|
self.fix_costs() |
|
122
|
|
|
test_costs = self.last_run_costs |
|
123
|
|
|
worker.send_req({ |
|
124
|
|
|
"eval_done": None, |
|
125
|
|
|
"valid_costs": valid_costs, |
|
126
|
|
|
"test_costs": test_costs, |
|
127
|
|
|
"auto_save": self.config.auto_save |
|
128
|
|
|
}) |
|
129
|
|
|
elif 'valid' in resp: |
|
130
|
|
|
self.best_cost = resp['best_valid_cost'] |
|
131
|
|
|
worker.copy_to_local() |
|
132
|
|
|
if valid_set: |
|
133
|
|
|
self._run_valid(self.epoch, valid_set, dry_run=True) |
|
134
|
|
|
self.fix_costs() |
|
135
|
|
|
worker.send_req({ |
|
136
|
|
|
"valid_done": None, |
|
137
|
|
|
"valid_costs": self.last_run_costs, |
|
138
|
|
|
"auto_save": self.config.auto_save |
|
139
|
|
|
}) |
|
140
|
|
|
elif 'train' in resp: |
|
141
|
|
|
batch_ids = resp['train'] |
|
142
|
|
|
batch_costs = [[] for _ in self.training_names] |
|
143
|
|
|
for batch_id in batch_ids: |
|
144
|
|
|
x = train_batches[batch_id] |
|
145
|
|
|
cost_x = self.learn(*x) |
|
146
|
|
|
for i, cost in enumerate(cost_x): |
|
147
|
|
|
batch_costs[i].append(cost) |
|
148
|
|
|
self.last_cost = cost_x[0] |
|
149
|
|
|
if network_callback: |
|
150
|
|
|
self.network.training_callback() |
|
151
|
|
|
if trainer_callback: |
|
152
|
|
|
for func in self._iter_controllers: |
|
153
|
|
|
func(self) |
|
154
|
|
|
worker.sync_params(synchronous=True) |
|
155
|
|
|
worker.send_req({'train_done': None, 'costs': [float(np.mean(c)) for c in batch_costs]}) |
|
156
|
|
|
elif 'sync_hyperparams' in resp: |
|
157
|
|
|
self.sync_hyperparams(resp['sync_hyperparams']) |
|
158
|
|
|
worker.close() |
|
159
|
|
|
return [] |
|
160
|
|
|
|
zomux /
deepy
