-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
344 lines (304 loc) · 11.9 KB
/
trainer.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
# coding: utf-8
#!/usr/bin/env python
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed.rpc as rpc
import torch.optim as optim
from torch.distributed.rpc import RRef, rpc_async, remote
from time import time
import argparse
import sys
import json
import threading
import csv
import os.path
# from Garfield.pytorch_impl.applications.Aggregathor.garfieldpp.datasets import TINYIMAGENET
import garfieldpp
from garfieldpp.worker import Worker
from garfieldpp.byzWorker import ByzWorker
from garfieldpp.server import Server
from garfieldpp.tools import get_bytes_com,convert_to_gbit, adjust_learning_rate
from datetime import datetime
import aggregators
import numpy as np
CIFAR_NUM_SAMPLES = 50000
TINYIMAGENET_NUM_SAMPLES = 100000
MNIST_NUM_SAMPLES = 60000
#First, parse the inputs
parser = argparse.ArgumentParser(description="AggregaThor implementation using Garfield++ library", formatter_class=argparse.RawTextHelpFormatter)
parser.add_argument("--master",
type=str,
default="",
help="Master node in the deployment. This node takes rank 0, usually the first PS.")
parser.add_argument("--rank",
type=int,
default=0,
help="Rank of a process in a distributed setup.")
parser.add_argument("--dataset",
type=str,
default="cifar10",
help="Dataset to be used, e.g., mnist, cifar10, tinyimagenet, ...")
parser.add_argument("--batch",
type=int,
default=128,
help="Minibatch size to be employed by each worker.")
parser.add_argument("--num_ps",
type=int,
default=1,
help="Number of parameter servers in the deployment (Vanilla AggregaThor uses 1 ps).")
parser.add_argument("--num_workers",
type=int,
default=1,
help="Number of workers in the deployment.")
parser.add_argument("--fw",
type=int,
default=0,
help="Number of declared Byzantine workers.")
parser.add_argument("--fps",
type=int,
default=0,
help="Number of declared Byzantine parameter servers (Vanilla AggregaThor does not assume Byzantine servers).")
parser.add_argument("--model",
type=str,
default='convnet',
help="Model to be trained, e.g., convnet, resnet18, resnet50,...")
parser.add_argument("--loss",
type=str,
default='nll',
help="Loss function to optimize against.")
parser.add_argument("--optimizer",
type=str,
default='sgd',
help="Optimizer to use.")
parser.add_argument("--opt_args",
type=json.loads,
default={'lr':'0.1'},
help="Optimizer arguments; passed in dict format, e.g., '{\"lr\":\"0.1\"}'")
parser.add_argument("--num_iter",
type=int,
default=5000,
help="Number of training iterations to execute.")
parser.add_argument("--gar",
type=str,
default='average',
help="Aggregation rule for aggregating gradients.")
parser.add_argument('--acc_freq',
type=int,
default=100,
help="The frequency of computing accuracy while training.")
parser.add_argument('--bench',
type=bool,
default=False,
help="If True, time elapsed in each step is printed.")
parser.add_argument('--log',
type=bool,
default=False,
help="If True, accumulated loss at each iteration is printed.")
parser.add_argument('--lr_update_freq',
type=int,
default=10,
help="Every what number of epochs should adjust learning rate.")
parser.add_argument('--port',
type=int,
default=29500,
help="Port to be used for RPC communication.")
parser.add_argument('--attack',
type=str,
default='random',
help="Attack to be used by Byzantine workers, e.g., random, drop, ...")
parser.add_argument('--r_col',
type=int,
default=1,
help="FlagMedian parameter for r-dimensional subspace of R^n, [Y∗] ∈ Gr(r,n), that is in some sense the center of points")
parser.add_argument('--lambda_',
type=float,
default=0,
help="Flag aggergator regularization hyperparameter serving as a coefficient for pairwise terms in the objective function")
parser.add_argument('--augmenteddataset',
type=str,
default='none',
help="Which augmented dataset to use, cifar10noisy, mnistnoisy tinyimagenetnoisy")
parser.add_argument('--augmentedfolder',
type=str,
default='none',
help="Which augmented folder to use, lv, half, onethird, train")
parser.add_argument("--seed",
type=int,
default=1001,
help="seed for reproducibility.")
parser.add_argument('--savedir',
type=str,
default='',
help="Which folder should the results go to.")
FLAGS = parser.parse_args(sys.argv[1:])
master = FLAGS.master
assert len(master) > 0
rank = FLAGS.rank
assert rank >= 0
num_ps = FLAGS.num_ps
assert num_ps >= 1
num_workers = FLAGS.num_workers
assert num_workers >= 1
world_size = num_workers + num_ps
fw = FLAGS.fw
assert fw*2 < num_workers
fps = FLAGS.fps
assert fps*2 < num_ps
dataset = FLAGS.dataset
assert len(dataset) > 0
batch = FLAGS.batch
assert batch >= 1
model = FLAGS.model
assert len(model) > 0
loss = FLAGS.loss
assert len(loss) > 0
optimizer = FLAGS.optimizer
assert len(optimizer) > 0
opt_args = FLAGS.opt_args
for k in opt_args:
opt_args[k] = float(opt_args[k])
assert opt_args['lr']
num_iter = FLAGS.num_iter
assert num_iter > 0
gar = FLAGS.gar
assert len(gar) > 0
acc_freq = FLAGS.acc_freq
assert(acc_freq > 10)
bench = FLAGS.bench
if bench:
from timeit import timeit
else:
timeit = None
log = FLAGS.log
port = FLAGS.port
attack = FLAGS.attack
r_col = FLAGS.r_col
assert r_col > 0
lambda_ = FLAGS.lambda_
assert lambda_ >= 0
seed = FLAGS.seed
assert seed > 0
savedir = FLAGS.savedir
augmenteddataset = FLAGS.augmenteddataset
augmentedfolder = FLAGS.augmentedfolder
lr_update_freq = FLAGS.lr_update_freq
mean_loss = 0
#os.environ['CUDA_VISIBLE_DEVICES'] = str((rank%2))
num_samples = CIFAR_NUM_SAMPLES
if dataset == 'cifar10' or dataset == 'cifar10noisy':
num_samples = CIFAR_NUM_SAMPLES
elif dataset == 'tinyimagenet':
num_samples = TINYIMAGENET_NUM_SAMPLES
elif dataset == 'mnist' or dataset == 'mnistnoisy' or dataset =='fmnist' or dataset == 'fmnistnoisy':
num_samples = MNIST_NUM_SAMPLES
print("**** SETUP AT NODE {} ***".format(rank))
print("Number of workers: ", num_workers)
print("Number of servers: ", num_ps)
print("Number of declared Byzantine workers: ", fw)
print("Number of declared Byzantine parameter servers: ", fps)
print("GAR: ", gar)
print("Dataset: ", dataset)
print("Model: ", model)
print("Batch size: ", batch)
print("Loss function: ", loss)
print("Optimizer: ", optimizer)
print("Optimizer Args", opt_args)
print("Benchmarking? ", bench)
print("Logging loss at each iteration?", log)
print("port: ", port)
print("attack: ", attack)
print("r_col: ", r_col)
print("lambda: ", lambda_)
print("seed:", seed)
print("augmenteddataset", augmenteddataset)
print("augmenteddfolder", augmentedfolder)
print("------------------------------------")
sys.stdout.flush()
filepath = "/home/data/Garfield/pytorch_impl/applications/Aggregathor/"
now = datetime.now()
now = now.strftime("%d-%m-%Y_%H-%M-%S")
logfilename = gar + "_" + "weights" + "_n_" + str(num_workers) + "_f_" + str(fw) + "_r_" + str(r_col) + "_lambda_" + str(lambda_) + "_" + str(attack) + "_" + now + ".txt"
checkpointfilename_base = gar + "_model_" + "_dataset_" + str(dataset) + "_batch_" + str(batch) + "_n_" + str(num_workers) + "_f_" + str(fw) + "_r_" + str(r_col) + "_lambda_" + str(lambda_) + "_" + str(attack) + "_epoch_"
lr = opt_args['lr']
output_file = "/home/data/Garfield/pytorch_impl/applications/Aggregathor/NIPS/"+savedir+gar+"_"+str(num_workers)+"_"+str(fw)+"_"+attack+"_"+model+"_"+dataset+"_"+augmenteddataset+"_"+augmentedfolder+"_"+str(batch)+"_"+str(lr)+"_"+str(r_col)+"_"+str(lambda_)+"_"+str(seed)+".csv"
with open(output_file, mode='w') as results_file:
writer = csv.writer(results_file)
writer.writerow(['Epoch', 'Loss', 'Accuracy', 'Time'])
#initiating the GAR
gar = aggregators.gars.get(gar)
assert gar is not None
os.environ['MASTER_ADDR'] = master
os.environ['MASTER_PORT'] = str(port)
os.environ['GLOO_SOCKET_IFNAME'] = 'ens3'
os.environ['TP_SOCKET_IFNAME'] = 'ens3'
torch.manual_seed(seed) #For reproducibility
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed) #For reproducibility
if bench:
torch.backends.cudnn.benchmark=True
#convention: low ranks are reserved for parameter servers
if rank < num_ps:
rpc.init_rpc('ps:{}'.format(rank), rank=rank, world_size=world_size, rpc_backend_options=rpc.TensorPipeRpcBackendOptions(init_method='env://', _transports=["uv"],)) # rpc_backend_options=rpc.TensorPipeRpcBackendOptions(num_worker_threads=80,rpc_timeout=20)
#Initialize a parameter server and write the training loop
ps = Server(rank, world_size, num_workers,1, fw, fps, 'worker:', 'ps:', batch, model, dataset, augmentedfolder, optimizer, **opt_args)
scheduler = torch.optim.lr_scheduler.MultiStepLR(ps.optimizer, milestones=[150, 250, 350], gamma=0.1) #This line shows sophisticated stuff that can be done out of the Garfield++ library
start_time = time()
iter_per_epoch = num_samples//(num_workers * batch) #this value records how many iteration per sample
print("One EPOCH consists of {} iterations".format(iter_per_epoch))
sys.stdout.flush()
# ps.model.load_state_dict(torch.load("/home/data/Garfield/pytorch_impl/applications/Aggregathor/flag_median_model__dataset_cifar10_batch_200_n_15_f_3_r_10_lambda_0.0_random_epoch_40.pt"))
# ps.model.eval()
for i in range(num_iter):
if i%(iter_per_epoch*lr_update_freq) == 0 and i!=0: #One hack for better convergence with Cifar10
lr*=0.2
adjust_learning_rate(ps.optimizer, lr)
#training loop goes here
def train_step():
global mean_loss
if bench:
bytes_rec = get_bytes_com() #record number of bytes sent before the training step to work as a checkpoint
with torch.autograd.profiler.profile(enabled=bench) as prof:
gradients, losses = ps.get_gradients(i, num_workers) #get_gradients(iter_num, num_wait_wrk)
mean_loss = np.mean(losses)
aggr_grad = gar(gradients=gradients, f=fw, log_file=None, data_itr=i, r_col=r_col, lambda_=lambda_, filepath=filepath, iter=i) #aggr_grad = gar.aggregate(gradients)
ps.update_model(aggr_grad)
# if i % (5 * iter_per_epoch) == 0:
# torch.save(ps.model.state_dict(), filepath + checkpointfilename_base + str(int(i/iter_per_epoch)) + ".pt")
if bench:
print(prof.key_averages().table(sort_by="self_cpu_time_total"))
bytes_train = get_bytes_com()
print("Consumed bandwidth in this iteration: {} Gbits".format(convert_to_gbit(bytes_train-bytes_rec)))
# print("Memory allocated to GPU {} Memory cached on GPU {}".format(torch.cuda.memory_allocated(0), torch.cuda.memory_cached(0)))
sys.stdout.flush()
if timeit is not None:
res = timeit(train_step,number=1)
print("Training step {} takes {} seconds".format(i,res))
sys.stdout.flush()
else:
train_step()
if i%iter_per_epoch == 0:
def test_step():
acc = ps.compute_accuracy()
num_epochs = int(i/iter_per_epoch)
with open(output_file, mode='a') as results_file:
writer = csv.writer(results_file)
writer.writerow([num_epochs, "{:.4f}".format(mean_loss), acc, "{:.3f}".format(time()-start_time)])
sys.stdout.flush()
if timeit is not None:
res = timeit(test_step,number=1)
print("Test step takes {} seconds".format(res))
else:
threading.Thread(target=test_step).start()
else:
rpc.init_rpc('worker:{}'.format(rank-num_ps), rank=rank, world_size=world_size, rpc_backend_options=rpc.TensorPipeRpcBackendOptions(init_method='env://', _transports=["uv"],)) #
#initialize a worker here
if (rank > fw):
Worker(rank, world_size, num_workers, batch, model, dataset, augmentedfolder, loss)
else:
if(augmenteddataset != "none"):
Worker(rank, world_size, num_workers, batch, model, augmenteddataset, augmentedfolder, loss)
else:
ByzWorker(rank, world_size, num_workers, batch, model, dataset, augmentedfolder, loss, attack)
rpc.shutdown()