forked from emilbjornson/power-allocation-cell-free
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Distributed_FFNN.py
186 lines (143 loc) · 7.71 KB
/
Distributed_FFNN.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Wed Apr 14 11:43:54 2021
@author: Mahmoud Zaher
"""
# This script is used for training the DDNN and DDNN-SI models for the different
# optimization objectives and precoding schemes considered.
import tensorflow as tf
from keras import backend as k
from keras.models import Sequential
from keras.layers.core import Dense
from keras import optimizers, regularizers
from keras.callbacks import EarlyStopping
import matplotlib.pyplot as plt
from os import makedirs
from sklearn import preprocessing
#import pandas as pd
import pathlib
from pickle import dump
base_path = pathlib.Path().absolute()
filename = str(base_path) + '/pAssign_storage/'
models_filename = str(base_path) + '/pAssignModels/DNNmodels/'
#reproducible results using Keras
sd = 42# Here sd means seed.
# 1. Set `PYTHONHASHSEED` environment variable at a fixed value
import os
os.environ['PYTHONHASHSEED']=str(sd)
# 2. Set `python` built-in pseudo-random generator at a fixed value
import random as rn
rn.seed(sd)
# 3. Set `numpy` pseudo-random generator at a fixed value
import numpy as np
np.random.seed(sd)
# 4. Set `tensorflow` pseudo-random generator at a fixed value
tf.compat.v1.set_random_seed(sd)
# 5. Configure a new global `tensorflow` session
config = tf.compat.v1.ConfigProto(intra_op_parallelism_threads=1,inter_op_parallelism_threads=1)
sess = tf.compat.v1.Session(graph=tf.compat.v1.get_default_graph(), config=config)
tf.compat.v1.keras.backend.set_session(sess)
# Loading input to the NN
mu_MR_sumSE_DNN = np.load(filename + 'dataset_mu_MR_WMMSE_ADMM.npy',allow_pickle=True)
mu_RZF_sumSE_DNN = np.load(filename + 'dataset_mu_RZF_WMMSE_ADMM.npy',allow_pickle=True)
mu_MR_PF_DNN = np.load(filename + 'dataset_mu_MR_WMMSE_PF_ADMM.npy',allow_pickle=True)
mu_RZF_PF_DNN = np.load(filename + 'dataset_mu_RZF_WMMSE_PF_ADMM.npy',allow_pickle=True)
betas_DNN = np.load(filename + 'dataset_betas.npy',allow_pickle=True)
#Maximum downlink transmit power per BS (mW)
Pmax = 1000
K = betas_DNN.shape[0]
L = betas_DNN.shape[1]
NoOfSetups = betas_DNN.shape[2]
# Make sure the sum over the K UEs gives Pmax for each AP in each setup- (might not be necessary)
for n in range(NoOfSetups):
mu_MR_sumSE_DNN[:,:,n] = mu_MR_sumSE_DNN[:,:,n] * np.sqrt( Pmax/(np.max (np.sum(np.power(mu_MR_sumSE_DNN[:,:,n],2), axis=0) )) )
mu_RZF_sumSE_DNN[:,:,n] = mu_RZF_sumSE_DNN[:,:,n] * np.sqrt( Pmax/(np.max (np.sum(np.power(mu_RZF_sumSE_DNN[:,:,n],2), axis=0) )) )
mu_MR_PF_DNN[:,:,n] = mu_MR_PF_DNN[:,:,n] * np.sqrt( Pmax/(np.max (np.sum(np.power(mu_MR_PF_DNN[:,:,n],2), axis=0) )) )
mu_RZF_PF_DNN[:,:,n] = mu_RZF_PF_DNN[:,:,n] * np.sqrt( Pmax/(np.max (np.sum(np.power(mu_RZF_PF_DNN[:,:,n],2), axis=0) )) )
# Maximum number of epochs
Num_epoch = 20
# Batch size
N_batch_size = 128
K_initializer = 'random_normal'
B_initializer = 'random_normal'
########################## MODEL FFNN ##########################
# Optimizer
adam = optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-07) #Decay = 0.1 not working
# Early stopping
early_stopping = EarlyStopping(monitor='val_loss', min_delta=0., patience=20, verbose=1, mode='auto')
callback = [early_stopping]
## Adjust the layers for DDNN or DDNN-SI model as per the tables in the paper
for l in range(0, L):
model = Sequential()
model.add(Dense(32, input_dim=K, activation="linear", name = 'layer1', kernel_initializer = K_initializer, bias_initializer=B_initializer))
# model.add(Dense(128, input_dim=64, activation="elu", name = 'layer2', kernel_initializer = K_initializer, bias_initializer=B_initializer))
model.add(Dense(64, input_dim=32, activation="tanh", name = 'layer3', kernel_initializer = K_initializer, bias_initializer=B_initializer))
model.add(Dense(32, input_dim=64, activation="tanh", name = 'layer4', kernel_initializer = K_initializer, bias_initializer=B_initializer))
model.add(Dense(K+1, input_dim=32, activation="relu", name = 'layer5', kernel_initializer = K_initializer, bias_initializer=B_initializer))
#Preparing inputs for NN
#beta vector preparation(removing outliers and scaling)
betas = betas_DNN[:,l,:].T # or use a function of betas
betas = 10*np.log10(betas*1000) # dB scale
big_values = []
for i in range(0, NoOfSetups):
if np.any(betas[i,:] > 34):
big_values = big_values + [i]
betas = np.delete(betas, big_values, axis=0)
NoOfSetups = betas.shape[0]
scaler = preprocessing.RobustScaler(
with_centering=False,
with_scaling=True,
quantile_range=(25.0, 75.0),
copy=True,)
betas = 10**(betas/10) # changing back to linear scale
## For calculation of the extra input used in the DDNN-SI model
# v = 0.6 # Fractional power allocation factor
# betas_to_all = np.delete(betas_DNN, big_values, axis = 2) * 1000 # linear scale (no kilo)
# extraInput = np.sqrt(Pmax) * ((betas ** v) / np.sum(betas_to_all ** v, axis=1 ).T)
# extraInput = 10*np.log10(extraInput) # dB scale
# The betas are changed to mus with fractional power allocatrion (gives better scaling)
v = 0.6 # Fractional power allocation factor
betas = np.sqrt(Pmax) * ((betas ** v) / np.sum( (betas ** v), axis=1 ).reshape(NoOfSetups,1))
betas = 10*np.log10(betas) # dB scale
betas = scaler.fit_transform(betas)
## The extra input used in the DDNN-SI model
# extraInput = scaler.transform(extraInput)
DNNinput = betas #np.concatenate((betas, extraInput), axis=1) # Second option is for DDNN-SI
x_train = DNNinput[0:NoOfSetups-100,:]
### Choose one of the following 4 options to train the model for
# mu_MR_sumSE_DNN, mu_RZF_sumSE_DNN, mu_MR_PF_DNN, mu_RZF_PF_DNN
## mu preparation
mu = np.abs(mu_MR_sumSE_DNN[:,l,:].T)
mu = np.delete(mu, big_values, axis=0)
temp = np.sqrt( np.reshape(sum((mu[0:NoOfSetups-100,:].T) ** 2), (NoOfSetups-100, 1)) / Pmax)
y_train = np.concatenate((mu[0:NoOfSetups-100,:], temp), axis=1)
small_values = []
small_val = 5 / np.sqrt(Pmax)
for i in range(0, NoOfSetups-100):
if y_train[i,K] < small_val:
small_values = small_values + [i]
y_train = np.delete(y_train, small_values, axis=0)
NoOfSetups = y_train.shape[0]
y_train[y_train < 0.001] = 0.001
y_train[:,0:K] = np.sqrt(K) * tf.keras.utils.normalize(y_train[:,0:K], axis=1)
#############################################################
x_train = np.delete(x_train, small_values, axis=0)
model.compile(optimizer = adam, loss = 'mean_squared_error', metrics=['accuracy'])
print(model.summary())
k.set_value(model.optimizer.lr, 0.001)
history = model.fit(x_train, y_train, epochs = Num_epoch, batch_size = N_batch_size, validation_split = 0.1, callbacks=callback)
k.set_value(model.optimizer.lr, 0.0001)
history2 = model.fit(x_train, y_train, epochs = 5, batch_size = N_batch_size, validation_split = 0.1, callbacks=callback)
x_test = DNNinput[NoOfSetups-100:NoOfSetups,:]
## Assign y_test based on the model choice above
y_test = mu[NoOfSetups-100:NoOfSetups,:]
y_test = np.sqrt(K) * tf.keras.utils.normalize(y_test, axis=1)
y_predictions = model.predict(x_test)
test_mse = np.mean((y_test - y_predictions[:,0:K])**2)
print('Test MSE:' + str(test_mse))
# Save models (You may adjust the name based on the chosen model)
model.save(models_filename + 'Trained_models_for_mu_MR_sumSE_DDNN_for_AP' + str(l+1))
print('Saved model %s' % models_filename + str(l+1))
# Save Scaler for betas (only required once)
# dump(scaler, open(models_filename + 'scaler.pkl', 'wb'))