def load_avg_convergence(d, prefix = ''):
c = []
for d_ in d:
et.globaldata.directory = prefix + d_
context = et.load('context.pkl')
c.append(np.array(et.load('acc_hist.pkl')))
return np.mean(c, axis = 0), context
def plot_mac():
et.globaldata.directory = '/homes/eneftci/work/code/C/auryn_rbp/build/release/experiments/Results/133__17-01-2017/'
context = et.load('context.pkl')
convergence_perbp = np.array(et.load('acc_hist.pkl'))
nh = context['nh']
nc = context['nc']
nv = context['nv']
spk_cnt = et.load('spkcnt.pkl')
acc_hist = np.array(et.load('acc_hist.pkl'))
nerr2 = extract_spk_cnt(spk_cnt, 0)
nerr1 = extract_spk_cnt(spk_cnt, 1)
nvis = extract_spk_cnt(spk_cnt, 2)
nout = extract_spk_cnt(spk_cnt, 3)
nhid = extract_spk_cnt(spk_cnt, 4)
import pickle
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_100.pkl','r')
cgpubp = np.array(pickle.load(fh)['test'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_100.pkl','r')
tbp = np.array(pickle.load(fh)['epoch'])
ss = np.cumsum((nv*nh + nh*nc + nv*nh + nh*nc)*np.ones(2000)*50000)
#plot(acc_hist[:,1],np.cumsum(nvis*nh + nhid*nc + 2*nout*nc + (nerr1+nerr2)*(nc+nh)).astype('float'),'r.-', linewidth=3)
plot(1-cgpubp/100,ss[tbp-1].astype('float'),'g.-', linewidth=3)
ylabel('MACs')
xlabel('MNIST Accuracy')
def plot_synop_mac():
et.globaldata.directory = '/homes/eneftci/work/code/C/auryn_rbp/build/release/experiments/Results/135__27-01-2017/'
context = et.load('context.pkl')
convergence_perbp = np.array(et.load('acc_hist.pkl'))
nh = context['nh']
nc = context['nc']
nv = context['nv']
spk_cnt = et.load('spkcnt.pkl')
acc_hist = np.array(et.load('acc_hist.pkl'))
nerr2 = extract_spk_cnt(spk_cnt, 0)
nerr1 = extract_spk_cnt(spk_cnt, 1)
nvis = extract_spk_cnt(spk_cnt, 2)
nout = extract_spk_cnt(spk_cnt, 3)
nhid = extract_spk_cnt(spk_cnt, 4)
import pickle
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_200_200.pkl','r')
cgpubp = np.array(pickle.load(fh)['test'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_200_200.pkl','r')
tbp = np.array(pickle.load(fh)['epoch'])
ss = np.cumsum((nv*nh + nh*nc + nv*nh + nh*nc)*np.ones(2000)*50000)
figure(figsize=(6,4))
semilogy(acc_hist[:,1],np.cumsum(nvis*nh + nhid*nc + 2*nout*nc + (nerr1+nerr2)*(nc+nh)).astype('float'),'r.-', linewidth=3, label='peRBP (Spiking)')
semilogy(1-cgpubp/100,ss[tbp-1].astype('float'),'k.-', linewidth=3, label='BP (GPU)', alpha=.6)
ylabel('MACs / SynOps')
xlabel('Accuracy')
legend(loc=2)
xticks([.75,.85,.90,.95,1.0])
tight_layout()
savefig('Results/synop_mac.png', format='png', dpi=1200)
def plot_convergence_200_200():
#stim_show(M['vh'].todense()[:,:100].T)
#et.savefig('features_100.png')
pylab.ion()
fh = file('/home/eneftci/Projects/code/python/dtp/Results/rbp_200_200.pkl','r')
import pickle
cgpu = np.array(pickle.load(fh)['test'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/rbp_200_200.pkl','r')
trbp = np.array(pickle.load(fh)['epoch'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_200_200.pkl','r')
cgpubp = np.array(pickle.load(fh)['test'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_200_200.pkl','r')
tbp = np.array(pickle.load(fh)['epoch'])
et.globaldata.directory = '/homes/eneftci/work/code/C/auryn_rbp/build/release/experiments/Results/135__27-01-2017/'
context = et.load('context.pkl')
convergence_perbp = np.array(et.load('acc_hist.pkl'))
et.globaldata.directory = '/homes/eneftci/work/code/C/auryn_rbp/build/release/experiments/Results/173__31-03-2017/'
context = et.load('context.pkl')
convergence_erbp = np.array(et.load('acc_hist.pkl'))
figure(figsize=(6,4))
title('784-200-200-10')
v = context['n_samples_train']/50000
semilogx(trbp,cgpu[:], 'b.-', linewidth=3, alpha=.6, label='RBP (GPU)')
semilogx(tbp,cgpubp[:], 'k.-', linewidth=3, alpha=.6, label='BP (GPU)')
semilogx(v+convergence_erbp[:-1,0]*v,100-100*convergence_erbp[:-1,1], 'g.-',linewidth=3, alpha=.6, label='eRBP (Spiking)')
semilogx(v+convergence_perbp[:-1,0]*v,100-100*convergence_perbp[:-1,1], 'r.-',linewidth=3, alpha=.6, label='peRBP (Spiking)')
print("200-200 perbp: {0:1.2f} erbp: {1:1.2f}".format(100*np.mean(1-convergence_perbp[-5:,1]),100*np.mean(1-convergence_erbp[-5:,1])))
xlim([0,2000])
xlabel('Epochs')
ylabel('Error %')
xticks([1,10,100,1000,2000],[1,10,100,1000,2000],rotation=45)
ylim([0,15])
#legend(frameon=False)
tight_layout()
savefig('Results/convergence_erbp_auryn_200_200.png', format='png', dpi=1200)
def plot_convergence_100():
#stim_show(M['vh'].todense()[:,:100].T)
#et.savefig('features_100.png')
pylab.ion()
fh = file('/home/eneftci/Projects/code/python/dtp/Results/rbp_100.pkl','r')
import pickle
cgpu = np.array(pickle.load(fh)['test'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/rbp_100.pkl','r')
trbp = np.array(pickle.load(fh)['epoch'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_100.pkl','r')
cgpubp = np.array(pickle.load(fh)['test'])
fh = file('/home/eneftci/Projects/code/python/dtp/Results/bp_100.pkl','r')
tbp = np.array(pickle.load(fh)['epoch'])
d = ['013__28-01-2017/', '014__28-01-2017/', '015__28-01-2017/', '016__28-01-2017/']
convergence_erbp, context = load_avg_convergence(d, prefix = '/home/eneftci/Projects/code/C/auryn_rbp/build/release/experiments/Results_scripts/')
assert context['sigma'] > 0
assert context['nh'] ==100
convergence_erbp = np.array(et.load('acc_hist.pkl'))
d = ['002__22-01-2017/', '003__22-01-2017/', '004__22-01-2017/', '001__22-01-2017/']
convergence_perbp, context = load_avg_convergence(d, prefix = '/home/eneftci/Projects/code/C/auryn_rbp/build/release/experiments/Results_scripts/')
assert context['sigma'] == 0
assert context['nh'] ==100
print("100 perbp: {0:1.2f} erbp: {1:1.2f}".format(100*np.mean(1-convergence_perbp[-5:,1]),100*np.mean(1-convergence_erbp[-5:,1])))
figure(figsize=(6,4))
title('784-100-10')
v = context['n_samples_train']/50000
semilogx(trbp,cgpu[:], 'b.-', linewidth=3, alpha=.6, label='RBP (GPU)')
semilogx(tbp,cgpubp[:], 'k.-', linewidth=3, alpha=.6, label='BP (GPU)')
semilogx(v+convergence_erbp[:-1,0]*v,100-100*convergence_erbp[:-1,1], 'g.-',linewidth=3, alpha=.6, label='eRBP (Spiking)')
semilogx(v+convergence_perbp[:-1,0]*v,100-100*convergence_perbp[:-1,1], 'r.-',linewidth=3, alpha=.6, label='peRBP (Spiking)')
xlim([0,2000])
xlabel('Epochs')
ylabel('Error %')
ylim([0,15])
xticks([1,10,100,1000,2000],[1,10,100,1000,2000],rotation=45)
legend(frameon=False, loc=0,prop={'size':16})
draw()
tight_layout()
savefig('Results/convergence_erbp_auryn_100.png', format='png', dpi=1200)
n_samples_test = context['n_samples_test']
n_epochs = context['n_epochs']
test_every = context['test_every'] #never test
if not hot_init and directory is None:
print 'Cold initialization..'
os.system('rm -rf inputs/{directory}/train/'.format(**context))
os.system('rm -rf inputs/{directory}/test/'.format(**context))
os.system('mkdir -p inputs/{directory}/train/' .format(**context))
os.system('mkdir -p inputs/{directory}/test/' .format(**context))
create_cnn_init(base_filename = 'inputs/{directory}/train/fwmat'.format(**context), **context)
elif directory is not None:
print 'Loading previous run...'
et.globaldata.directory = directory
M = et.load('M.pkl')
M = process_allparameters_rbp(context)
#save_parameters(M, context)
if test_every>0:
labels_test, SL_test = create_data_rbp(n_samples = n_samples_test,
output_directory = '{directory}/test'.format(**context),
data_url = context['test_data_url'],
labels_url = context['test_labels_url'],
randomize = False,
with_labels = False,
duration_data = context['sample_duration_test'],
duration_pause = context['sample_pause_test'],
generate_sl = False,
**context)
hot_init = True
if o == '-n': #change number of epochs
context['n_epochs'] = int(a)
if o == '-t':
context['test_every'] = int(a)
if o == '-d':
directory = a
if o == '-s':
save = True
if o == '-g':
generate = True
if o == '-c': #custom parameters passed through input arguments
context.update(eval(a))
et.globaldata.directory = directory
context = et.load('context.pkl')
context['n_samples_test'] = 10000
context['sample_duration_test'] = .5
context['sample_pause_test'] = .0
context['simtime_test'] = context['n_samples_test'] * (
context['sample_duration_test'] + context['sample_pause_test'])
if not context.has_key('prob_syn'):
context['prob_syn'] = 1.0
if not context.has_key('sigma'):
context['sigma'] = 50e-3
n_samples_train = context['n_samples_train']
n_samples_test = context['n_samples_test']
n_epochs = context['n_epochs']
test_every = context['test_every'] #never test
M = et.load('M.pkl')
import pickle
import numpy as np
import matplotlib.pylab as plt
import experimentTools as ext
from matplotlib import gridspec
from pyNSATlib import nsat_reader
# from experimentLib import *
if __name__ == '__main__':
# idir = "005__05-04-2017/"
idir = "Results/016__22-09-2017/"
cfg = ext.load(idir + 'cfg_train.pkl')
nsat_stats = np.cumsum(pickle.load(
file('Results/016__22-09-2017/stats_nsat.pkl', 'r')),
axis=0)
stats_h = nsat_stats[:, :100].sum(axis=1)
stats_o = nsat_stats[:, 100:110].sum(axis=1)
stats_e = nsat_stats[:, 110:].sum(axis=1)
weights = nsat_reader.read_synaptic_weights(cfg[0], idir + '/weights.pkl',
idir + '/ptr.pkl')
w0 = weights[:794, 794:894, 0].flatten()
w1 = weights[793:893, 894:(894 + 10), 0].flatten()
w2 = weights[904:, 794:894, 1].flatten()
W = [w0, w1, w2]
epo, err = np.array(pickle.load(file(idir + 'pip.pkl', 'r'))).T
print(epo)
# w = cfg.core_cfgs[0].W[:794, 794:894, 1]
# np.append(w, cfg.core_cfgs[0].W[:])
epo_tf, acc_tf, macs_tf = pickle.load(
# Licence : GPLv2
#-----------------------------------------------------------------------------
from experimentLib import *
import experimentTools as et
from pylab import *
import matplotlib, pickle
font = {'family': 'normal', 'weight': 'bold', 'size': 14}
matplotlib.rc('font', **font)
matplotlib.rc('savefig', dpi=600)
dnsat = '/home/eneftci/Projects/code/python/HiAER-NSAT/examples/erbp/Results/021__15-01-2017/'
et.globaldata.directory = '/homes/eneftci/work/code/C/auryn_rbp/build/release/experiments/Results/127__13-01-2017/'
convergence = np.array(et.load('acc_hist.pkl'))
context = et.load('context.pkl')
figure(figsize=(6, 4))
v = context['n_samples_train'] / 50000
semilogx(v + (convergence[:, 0]) * v,
100 - 100 * convergence[:, 1],
'o-',
linewidth=3,
alpha=.6,
label='peRBP (Spiking 64-bit)')
tt, nsat = np.array(pickle.load(file(dnsat + 'pip.pkl', 'r'))).T
semilogx(1 + tt,
100 - nsat,
'x-',
linewidth=3,
alpha=.6,
et.mksavedir()
et.globaldata.context = context
start_epoch = 0
acc_hist = []
snr_hist = []
weight_stats = {}
output_weights = []
spkcnt = [None for i in range(context['n_epochs'])]
if args.resume is not '':
if args.resume[-1] is not '/':
args.resume += '/'
print 'Loading previous run from {}'.format(args.resume)
et.globaldata.directory = args.resume
M = et.load('M.pkl')
old_context = et.load('context.pkl')
acc_hist = et.load('acc_hist.pkl')
spkcnt = et.load('spkcnt.pkl')
snr_hist = et.load('snr_hist.pkl')
bestM = et.load('bestM.pkl')
start_epoch = acc_hist[-1][0] + 1
context.update(old_context)
# take n_cores and test interval from current args
context['ncores'] = args.n_cores
context['test_every'] = args.testinterval
print('Restored context: {}'.format(context))
elib.write_allparameters_rbp(M, context)
max_samples_train = context['max_samples_train']
if __name__ == '__main__':
try:
last_perf = (0.0, 0.0)
init = True # initialize weights?
new_test_data = True # generate new test data ras files?
test = True # test before first training epoch?
save = True # save results to result folder?
# folder = '066__21-07-2018'
# directory = 'Results/{}/'.format(folder)
directory = None
if directory is not None:
print 'Loading previous run...'
et.globaldata.directory = directory
M = et.load('M.pkl')
old_context = et.load('context.pkl')
context.update(old_context)
elib.write_allparameters_rbp(M, context)
context['test_every'] = 2
context['n_epochs'] = 0
context['n_samples_train'] = 10
context['eta'] = 1e-7
context['n_samples_test'] = 1000
max_samples_train = context['max_samples_train']
max_samples_test = context['max_samples_test']
n_samples_train = context['n_samples_train']
n_samples_test = context['n_samples_test']
n_epochs = context['n_epochs']
test_every = context['test_every']
