-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathnmnist.py
184 lines (160 loc) · 8.31 KB
/
nmnist.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
import torch
import matplotlib.pyplot as plt
import os
import sys
import time
from pathlib import Path
import argparse
import copy
from antlr import *
from trainer import *
import utils
import nmnist_dataset
parser = argparse.ArgumentParser(description='ANTLR NMNIST Example Arguments', fromfile_prefix_chars='@')
parser.add_argument('-t', '--tag', type=str, default='', metavar='S',
help='tag for the run')
parser.add_argument('-s', '--random-seed', type=int, metavar='N',
help='random seed. if not specified, new random seed is generated.')
parser.add_argument('--eval-model', type=str, metavar='S',
help='enable evaluation mode')
# Training/Model configuration.
parser.add_argument('--optim-name', type=str, default='adam', metavar='S',
help='a type of optimizer (e.g. adam (default), sgd)')
parser.add_argument('--learning-rate', type=float, default=0.001, metavar='F',
help='learning rate (default: 0.001)')
parser.add_argument('--momentum', type=float, default=0, metavar='F',
help='SGD momentum (default: 0))')
parser.add_argument('--weight-decay', type=float, default=0, metavar='F',
help='weight decay (default: 0))')
parser.add_argument('--max_target-spikes', type=int, default=1, metavar='N',
help='number of spikes for true target (default: 1)')
parser.add_argument('--min_target-spikes', type=int, default=0, metavar='N',
help='number of spikes for false target (default: 0)')
parser.add_argument('--resume', action='store_true', default=False,
help="Whether to resume at the last point.")
parser.add_argument('--config', default='configs/nmnist.json')
parser.add_argument('--num-workers', type=int, default=1, metavar='N',
help='(default: 1)')
parser.add_argument('--inf-speed-test', type=int, default=0, metavar='N',
help='(default: 0)')
model_args = parser.add_argument_group('model parameters')
parser.add_argument('--time-length', type=int, default=300, metavar='F',
help='simulation time length (default: 300))')
model_args.add_argument('-b', '--batch-size', type=int, default=16, metavar='N',
help='input batch size for training (default: 16)')
model_args.add_argument('--test-batch-size', type=int, default=256, metavar='N',
help='input batch size for testing (default: 256)')
model_args.add_argument('--alpha-i', type=float, default=0.99, metavar='F',
help='alpha_i used for decaying current\
(default: 0.99)')
model_args.add_argument('--alpha-v', type=float, default=0.99, metavar='F',
help='alpha_v used for decaying voltage\
(default: 0.99)')
model_args.add_argument('--beta-i', type=float, default=1.0, metavar='F',
help='beta_i used for scaling current (default: 1.0)')
model_args.add_argument('--beta-v', type=float, default=1.0, metavar='F',
help='beta_v used for scaling voltage (default: 1.0)')
model_args.add_argument('--beta-bias', type=float, default=1.0, metavar='F',
help='beta_bias used for scaling voltage bias (default: 1.0)')
model_args.add_argument('--surr-alpha', type=float, default=1.0, metavar='F',
help='surr_alpha used for surrogate derivative \
(default: 1.0)')
model_args.add_argument('--surr-beta', type=float, default=3.0, metavar='F',
help='surr_beta used for surrogate derivative \
(default: 3.0)')
model_args.add_argument('-l', '--lrule', type=str, default='ANTLR', metavar='S',
help='learning rule type')
model_args.add_argument('--target-type', type=str, default='latency', metavar='S',
help='type of target values (e.g. \'count\' (default), \'train\', \'latency\')')
# Depricated.
model_args.add_argument('--lambda-nospike', type=float, default=0.1, metavar='F',
help='lambda for no_spike loss or latency target (default: 0.1)')
model_args.add_argument('--timing-penalty', type=float, default=100.0, metavar='F',
help='(default: 100.0)')
model_args.add_argument('--grad-clip', type=str, default='1.0', metavar='F',
help='(default: 1.0)')
model_args.add_argument('--multi-model', type=int, default=1, metavar='N',
help='Default=0')
model_args.add_argument('--num-models', type=int, default=15, metavar='N',
help='Default=1')
model_args.add_argument('--init-bias-center', type=int, default=0, metavar='N',
help='0')
model_args.add_argument('--beta-auto', type=int, default=1, metavar='N',
help='1')
# Beta in softmax function.
model_args.add_argument('--softmax-beta', type=float, default=0.166667, metavar='N',
help='Beta (1/temperature) in softmax')
apargs = parser.parse_args()
apargs.grad_clip = [float(item) for item in apargs.grad_clip.split(',')]
if apargs.eval_model is not None:
apargs.config = Path(f"./logs/{apargs.eval_model}/config.json")
def main():
# Load configurations from json file.
config_dict = utils.read_json(apargs.config)
config = utils.Config(config_dict)
if apargs.eval_model is None:
config.__dict__.update({key: getattr(apargs, key) for key in vars(apargs)})
else:
config.eval_model = True
# When reusing the parameters trained with multi-model settings,
# the multi-model switch should be turned off before loading.
if config.eval_model :
config.num_models = 1
config.multi_model = 0
logger = utils.Logger(config.tag, config.resume, task='nmnist')
if logger.resume:
config.random_seed = logger.config_resume.random_seed
elif config.random_seed is None:
config.random_seed = np.random.randint(1, 1000000)
torch.manual_seed(config.random_seed)
logger.save_config(config)
cuda_enabled = config.gpu and torch.cuda.is_available()
if cuda_enabled:
try:
torch.multiprocessing.set_start_method('spawn', force=True )
except:
pass
torch.cuda.manual_seed(config.random_seed)
torch.set_default_tensor_type(torch.cuda.FloatTensor)
data_loaders = nmnist_dataset.load_loader(config=config, num_workers=config.num_workers,
batch_size=config.batch_size, test_batch_size=config.test_batch_size,
time_length=config.time_length)
trainer = Trainer(config, data_loaders=data_loaders, logger=logger, gpu=cuda_enabled, task='nmnist')
# seed again to ensure the model parameters are sampled right after the
# seed setting.
torch.manual_seed(config.random_seed)
if cuda_enabled:
torch.cuda.manual_seed(config.random_seed)
import pdb; pdb.set_trace()
trainer.make_model(config)
if apargs.eval_model is not None:
param_path = Path(f"./logs/{apargs.eval_model}/m0_best_model.pt")
if cuda_enabled:
trainer.load_model(torch.load(param_path))
else:
trainer.load_model(torch.load(param_path, map_location=torch.device('cpu')))
trainer.test()
else:
if logger.resume:
param_path_model = logger.log_dir / "last_model.pt"
param_path_optim = logger.log_dir / "last_optim.pt"
try:
if cuda_enabled:
trainer.load_model(torch.load(param_path_model))
trainer.load_optim(torch.load(param_path_optim))
else:
trainer.load_model(torch.load(param_path_model, map_location=torch.device('cpu')))
trainer.load_optim(torch.load(param_path_optim, map_location=torch.device('cpu')))
# for randomness in resuming
seed_keep = np.random.randint(1, 100000)
torch.manual_seed(seed_keep)
if cuda_enabled:
torch.cuda.manual_seed(seed_keep)
except:
logger.resume = False
else:
trainer.save_model('init')
trainer.make_scheduler()
trainer.run(config)
if __name__ == "__main__":
main()