-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate.py
189 lines (152 loc) · 8.02 KB
/
evaluate.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
import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
import time
from collections import defaultdict, OrderedDict
import pickle
from tqdm import tqdm
import random
import math
import argparse
import json
import random
import numpy as np
from timm.scheduler import *
from model.vgg_cifar10 import *
from model.vgg_cifar100 import*
from model.lenet_fashion import *
from optimizer.gossip_optimizer import *
from optimizer.d2_optimizer import *
from data.loader_dirichlet import *
from ring import *
from exponential_graph import *
from one_peer_exponential_graph import *
from base_graph import *
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
def run(rank, size, datasets, config):
# initialize the model parameters with same seed value.
torch.manual_seed(config["seed"])
random.seed(config["seed"])
np.random.seed(config["seed"])
torch.set_num_threads(1)
if config["model"] == "vgg":
if config["dataset"] == "cifar10":
net = VggCifar10(device=config["device"][rank]).to(config["device"][rank])
elif config["dataset"] == "cifar100":
net = VggCifar100(device=config["device"][rank]).to(config["device"][rank])
if config["model"] == "lenet":
net = LeNetFashion(device=config["device"][rank]).to(config["device"][rank])
net.to(config["device"][rank])
loaders = datasets_to_loaders(datasets, config["batch"])
if config["optimizer"] == "gossip":
optimizer = GossipOptimizer(params=net.parameters(), node_id=rank, graph=config["graph"], local_step=config["local_step"], lr=config["lr"], beta=config["beta"], device=config["device"][rank])
elif config["optimizer"] == "d2":
optimizer = D2Optimizer(params=net.parameters(), node_id=rank, graph=config["graph"], local_step=config["local_step"], lr=config["lr"], beta=config["beta"], device=config["device"][rank])
scheduler = CosineLRScheduler(optimizer, t_initial=config["epochs"], lr_min=1e-4, warmup_t=10, warmup_lr_init=5e-5, warmup_prefix=True)
history = {"train_loss": [], "train_acc": [], "val_loss": [], "val_acc": [], "test_loss": [], "test_acc": [], "diff_param": []}
count_epoch = 0
with tqdm(range(config["epochs"]), desc=("node "+str(rank)), position=rank) as pbar:
for epoch in pbar:
train_loss, train_acc = net.run(loaders, optimizer)
if (count_epoch % 10 == 0) or (count_epoch == config["epochs"] -1):
val_loss, val_acc = net.run_val(loaders)
test_loss, test_acc = net.run_test(loaders)
# save loss and accuracy
history["train_loss"] += [train_loss]
history["test_loss"] += [test_loss]
history["val_loss"] += [val_loss]
history["train_acc"] += [train_acc]
history["test_acc"] += [test_acc]
history["val_acc"] += [val_acc]
pbar.set_postfix(OrderedDict(loss=(round(train_loss, 2), round(test_loss, 2)), acc=(round(train_acc, 2), round(test_acc, 2))))
count_epoch += 1
scheduler.step(count_epoch)
pickle.dump(history, open(config["log_path"] + "node" + str(rank) + ".pk", "wb"))
def init_process(rank, size, datasets, config, fn, backend='gloo'):
os.environ['MASTER_ADDR'] = config["config"]["master_address"] #'127.0.0.1'
os.environ['MASTER_PORT'] = config["config"]["port"]
dist.init_process_group(backend, rank=rank, world_size=size)
fn(rank, size, datasets, config)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('log', default="./results", type=str, help="A directory where the log is stored.")
parser.add_argument('--dataset', default="cifar10", type=str, help="Dataset. cifar10 and cifar100 are available.")
parser.add_argument('--optimizer', default="gossip", type=str, help="Optimization method.")
parser.add_argument('--batch', default=32, type=int, help="Batch size.")
parser.add_argument('--seed', default=0, type=int, help="Seed value.")
parser.add_argument('--model', default="lenet", type=str, help="Neural network architecture. vgg is available.")
parser.add_argument('--nw', default="one_peer_base", type=str, help="An undelying network topology. one_peer_base, two_peer_base, etc. are available.")
parser.add_argument('--cuda', default=None, type=str)
parser.add_argument('--config', default="config/8_node.json", type=str, help="A configulation file.")
parser.add_argument('--node_list', nargs="*", type=int, help="A list of node id.")
parser.add_argument('--lr', default=1e-3, type=float, help="Learning rate.")
parser.add_argument('--epoch', default=1000, type=int, help="The number of epochs.")
parser.add_argument('--alpha', default=100, type=float, help="Hyperparameter of Dirichlet distribution.")
parser.add_argument('--beta', default=0.9, type=float, help="Momentum coefficient.")
parser.add_argument('--local_step', default=5, type=int, help="The number of local step. The local step is not available now. Please specify 1.")
args = parser.parse_args()
config = defaultdict(dict)
config["dataset"] = args.dataset
config["optimizer"] = args.optimizer
config["lr"] = args.lr
config["seed"] = args.seed
config["epochs"] = args.epoch
config["log_path"] = args.log
config["batch"] = args.batch
config["model"] = args.model
config["node_list"] = args.node_list
config["nw"] = args.nw
config["beta"] = args.beta
config["alpha"] = args.alpha
config["local_step"] = args.local_step
with open(args.config) as f:
config["config"] = json.load(f)
n_nodes = config["config"]["n_nodes"]
if config["nw"] == "ring":
config["graph"] = Ring(n_nodes)
elif config["nw"] == "exp":
config["graph"] = ExponentialGraph(n_nodes)
elif config["nw"] == "one_peer_exp":
config["graph"] = OnePeerExponentialGraph(n_nodes)
elif config["nw"] == "one_peer_base":
config["graph"] = BaseGraph(n_nodes, max_degree=1, seed=config["seed"])
elif config["nw"] == "two_peer_base":
config["graph"] = BaseGraph(n_nodes, max_degree=2, seed=config["seed"])
elif config["nw"] == "three_peer_base":
config["graph"] = BaseGraph(n_nodes, max_degree=3, seed=config["seed"])
elif config["nw"] == "four_peer_base":
config["graph"] = BaseGraph(n_nodes, max_degree=4, seed=config["seed"])
else:
print("ERROR: exp, one_peer_exp, {one,two,three,four}_peer_base are available", file=sys.stderr)
sys.exit(1)
if args.cuda is None:
config["device"] = {node_id : config["config"][f"node{node_id}"]["cuda"] for node_id in config["node_list"]}
else:
config["device"] = [args.cuda for _ in range(n_nodes)]
torch.manual_seed(config["seed"])
random.seed(config["seed"])
np.random.seed(config["seed"])
if config["dataset"] == "cifar10":
datasets = load_CIFAR10(n_nodes, batch=config["batch"], alpha=config["alpha"], val_rate=0.1, seed=config["seed"])
elif config["dataset"] == "cifar100":
datasets = load_CIFAR100(n_nodes, batch=config["batch"], alpha=config["alpha"], val_rate=0.1, seed=config["seed"])
elif config["dataset"] == "fashion_mnist":
datasets = load_FMNIST(n_nodes, alpha=config["alpha"], val_rate=0.1, seed=config["seed"])
else:
print('cifar10 or cifar100 are available in dataset', file=sys.stderr)
sys.exit(1)
processes = []
try:
mp.set_start_method("spawn")
except RuntimeError:
pass
for rank in config["node_list"]:
print(rank)
node_datasets = {"train": datasets["train"][rank], "val": datasets["val"], "test": datasets["test"]}
p = mp.Process(target=init_process, args=(rank, n_nodes, node_datasets, config, run))
p.start()
processes.append(p)
for p in processes:
p.join()