-
Notifications
You must be signed in to change notification settings - Fork 26
/
Copy pathtest_l1Fourier_lifted.py
239 lines (209 loc) · 9.84 KB
/
test_l1Fourier_lifted.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import os
# We do not use GPUs for SigPy
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID";
os.environ["CUDA_VISIBLE_DEVICES"] = "";
# Set threading for multiple libraries to prevent resource hogging
num_threads = 1
os.environ["OMP_NUM_THREADS"] = str(num_threads)
os.environ["OMP_DYNAMIC"] = "false"
os.environ["OPENBLAS_NUM_THREADS"] = str(num_threads)
os.environ["MKL_NUM_THREADS"] = str(num_threads)
os.environ["VECLIB_MAXIMUM_THREADS"] = str(num_threads)
os.environ["NUMEXPR_NUM_THREADS"] = str(num_threads)
import numpy as np
import sigpy as sp
import torch, sys, itertools, copy, argparse
sys.path.append('./')
from tqdm import tqdm as tqdm
from loaders import Channels
from torch.utils.data import DataLoader
torch.set_num_threads(num_threads)
from scipy.fft import ifft
from dotmap import DotMap
from matplotlib import pyplot as plt
# Args
parser = argparse.ArgumentParser()
parser.add_argument('--train', type=str, default='CDL-C')
parser.add_argument('--test', type=str, default='CDL-C')
parser.add_argument('--antennas', nargs='+', type=int, default=[16, 64])
parser.add_argument('--array', type=str, default='ULA')
parser.add_argument('--spacing', type=float, default=0.5)
parser.add_argument('--alpha', nargs='+', type=float, default=[0.6])
parser.add_argument('--lmbda', nargs='+', type=float, default=[0.3])
parser.add_argument('--lifting', type=int, default=4)
parser.add_argument('--steps', type=int, default=1000)
parser.add_argument('--lr', nargs='+', type=float, default=[3e-3])
args = parser.parse_args()
# Create and populate minimal configuration
config = DotMap()
# For CS methods, train/test affects normalization
config.data.channel = args.train
config.data.array = args.array
config.data.image_size = [args.antennas[0], args.antennas[1]]
config.data.num_pilots = args.antennas[1]
config.data.spacing_list = [args.spacing]
config.data.noise_std = 1 # Dummy value
config.data.mixed_channels = False
# Seeds for train and test datasets
train_seed, val_seed = 1234, 4321
# Load training dataset
dataset = Channels(train_seed, config, norm='global')
# Range of SNR, test channels and hyper-parameters
snr_range = np.asarray(np.arange(-10, 35, 5))
spacing_range = np.asarray([args.spacing]) # Antenna spacing
alpha_range = np.asarray(args.alpha) # Fraction of pilots
lmbda_range = np.asarray(args.lmbda) # L1 regularization strength
lr_range = np.asarray(args.lr) # Step size
lifting = int(args.lifting)
# SNR is defined as Nr / noise_power
# assuming average unit-power entries in MIMO channel matrix
noise_range = 10 ** (-snr_range / 10.) * args.antennas[1]
gd_iter = args.steps # Number of optimization steps
# Limit number of samples for faster results
kept_samples = 50
# Global results
nmse_log = np.zeros((len(spacing_range), len(alpha_range),
len(lmbda_range), len(lr_range),
len(snr_range), kept_samples)) # Should match data
complete_log = np.zeros((len(spacing_range), len(alpha_range),
len(lmbda_range), len(lr_range),
len(snr_range), gd_iter, kept_samples))
result_dir = './results/l1CS_lifted%d/train-%s_test-%s' % (
lifting, args.train, args.test)
os.makedirs(result_dir, exist_ok=True)
# Wrap sparsity, steps and spacings
meta_params = itertools.product(spacing_range, alpha_range,
lmbda_range, lr_range)
# For each hyper-combo
for meta_idx, (spacing, alpha, lmbda, lr) in tqdm(enumerate(meta_params)):
# Unwrap indices
spacing_idx, alpha_idx, lmbda_idx, lr_idx = np.unravel_index(
meta_idx, (len(spacing_range), len(alpha_range),
len(lmbda_range), len(lr_range)))
# Prepare validation configuration
val_config = copy.deepcopy(config)
val_config.data.channel = args.test
val_config.data.spacing_list = [spacing]
val_config.data.num_pilots = int(np.floor(args.antennas[1] * alpha))
# Normalize test data using training data statistics
val_dataset = Channels(val_seed, val_config, norm=[dataset.mean, dataset.std])
val_loader = DataLoader(val_dataset, batch_size=len(val_dataset),
shuffle=False, num_workers=0, drop_last=True)
val_iter = iter(val_loader)
# Get all validation data explicitly
val_sample = next(val_iter)
del val_iter, val_loader # Free up memory
val_P = val_sample['P']
val_P = torch.conj(torch.transpose(val_P, -1, -2))
val_H_herm = val_sample['H_herm']
val_H = val_H_herm[:, 0] + 1j * val_H_herm[:, 1]
# Convert to numpy
val_P = val_P.resolve_conj().numpy()
val_H = val_H.resolve_conj().numpy()
# Keep limited number of samples
val_P = val_P[:kept_samples, ...]
val_H = val_H[:kept_samples, ...]
# Dictionary matrices for ULA/UPA array shapes
left_dict = np.conj(ifft(np.eye(val_H[0].shape[0]),
n=val_H[0].shape[0]*lifting, norm='ortho'))
right_dict = ifft(np.eye(val_H[0].shape[1]),
n=val_H[0].shape[1]*lifting, norm='ortho').T
# Lifted shape
lifted_shape = (val_H[0].shape[0]*lifting, val_H[0].shape[1]*lifting)
# Proximal op for sigpy
prox_op = sp.prox.L1Reg(lifted_shape, lmbda)
# Run CS for each SNR value
for snr_idx, local_noise in tqdm(enumerate(noise_range)):
val_Y = np.matmul(val_P, val_H)
val_Y = val_Y + np.sqrt(local_noise) / np.sqrt(2.) * \
(np.random.normal(size=val_Y.shape) + \
1j * np.random.normal(size=val_Y.shape))
# For each sample
for sample_idx in tqdm(range(val_Y.shape[0])):
# Create forward and regularization ops
array_op = sp.linop.Compose(
(sp.linop.MatMul((
lifted_shape[0], val_H[sample_idx].shape[1]), left_dict),
sp.linop.RightMatMul(lifted_shape, right_dict)))
fw_op = sp.linop.Compose(
(sp.linop.MatMul(val_H[sample_idx].shape, val_P[sample_idx]),
array_op))
# Gradient function in closed form
def gradf(x):
return fw_op.H * (fw_op * x - val_Y[sample_idx])
# Initial point and instantiate algorithm object
val_H_hat = np.zeros(lifted_shape, complex)
alg = sp.alg.GradientMethod(
gradf, val_H_hat, lr, proxg=prox_op, max_iter=gd_iter,
accelerate=True)
# For each optimization step
for step_idx in range(gd_iter):
# Run update step
alg.update()
# Convert current estimate from Fourier domain to spatial domain
est_H = array_op(val_H_hat)
# Log estimation errors
complete_log[spacing_idx, alpha_idx,
lmbda_idx, lr_idx, snr_idx, step_idx,
sample_idx] = \
(np.sum(np.square(
np.abs(est_H - val_H[sample_idx])), axis=(-1, -2)))/\
np.sum(np.square(
np.abs(val_H[sample_idx])), axis=(-1, -2))
# Convert final estimate to IFFT
est_H = array_op(val_H_hat)
# Save NMSE for each sample
nmse_log[spacing_idx, alpha_idx,
lmbda_idx, lr_idx, snr_idx, sample_idx] = \
(np.sum(np.square(
np.abs(est_H - val_H[sample_idx])), axis=(-1, -2)))/\
np.sum(np.square(
np.abs(val_H[sample_idx])), axis=(-1, -2))
# Use average estimation error to find best hyper-parameters
avg_nmse = np.mean(nmse_log, axis=-1)
best_nmse = np.zeros((len(alpha_range), len(snr_range)))
best_lmbda = np.zeros((len(alpha_range), len(snr_range)))
best_lr = np.zeros((len(alpha_range), len(snr_range)))
# For each alpha and SNR value
for alpha_idx, local_alpha in enumerate(alpha_range):
for snr_idx, local_snr in enumerate(snr_range):
local_nmse = avg_nmse[0, alpha_idx, ..., snr_idx]
local_nmse = local_nmse.flatten()
best_idx = np.argmin(local_nmse)
lmbda_idx, lr_idx = \
np.unravel_index(best_idx, (len(lmbda_range), len(lr_range)))
# Store and verbose
best_nmse[alpha_idx, snr_idx] = local_nmse[best_idx]
best_lmbda[alpha_idx, snr_idx] = lmbda_range[lmbda_idx]
best_lr[alpha_idx, snr_idx] = lr_range[lr_idx]
print('SNR = %.2f dB, NMSE = %.2f dB using lambda = %.1e and step size = %.1e' % (
local_snr, 10*np.log10(best_nmse[alpha_idx, snr_idx]),
best_lmbda[alpha_idx, snr_idx], best_lr[alpha_idx, snr_idx]))
# Plot results for all alpha values
plt.rcParams['font.size'] = 14
plt.figure(figsize=(10, 10))
for alpha_idx, local_alpha in enumerate(alpha_range):
plt.plot(snr_range, 10*np.log10(best_nmse[alpha_idx]),
linewidth=4, label='Alpha=%.2f' % local_alpha)
plt.grid(); plt.legend()
plt.title('Compressed Sensing fsAD, lifting = %d' % args.lifting)
plt.xlabel('SNR [dB]'); plt.ylabel('NMSE [dB]')
plt.tight_layout()
plt.savefig(os.path.join(result_dir, 'results.png'), dpi=300,
bbox_inches='tight')
plt.close()
# Save full results to file
torch.save({'complete_log': complete_log,
'nmse_log': nmse_log,
'best_nmse': best_nmse,
'best_lmbda': best_lmbda,
'best_lr': best_lr,
'snr_range': snr_range,
'spacing_range': spacing_range,
'alpha_range': alpha_range,
'lmbda_range': lmbda_range,
'lr_range': lr_range,
'config': config, 'args': args
}, os.path.join(result_dir, 'results.pt'))