def general_SVC_exclude_test_set(spec_svm_file=None):
''' Train the general SVM according to regimes on all data excluding the test data.
Compute power at all (s,t)'s.
'''
global st2pow
pfile_text = []
########## read (s,t)-specific SVMs from file ##########
if(spec_svm_file != None):
# read specific SVM dict from pickled file
sys.stdout.write("Reading in cPickled SVM-dict from: %s... " % spec_svm_file)
with open("../" + spec_svm_file, mode='rb') as svm_fh:
svm_dict = pickle.load(svm_fh)
print "loaded.\n"
############## train & test on all (s,t) ###############
for s,t in [(s,t) for s in p.selection for t in p.times]:
# training data, EXCLUDING test data
print "training (excluded) on (%g,%i)" % (s,t)
train_mix = data_mix_for_SVM_general_exclude(s,t)
svm_weights, svm_e_terms = {}, {}
for i in range(len(train_mix)/2): svm_weights[i], svm_e_terms[i] = 1.0, 0.1
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
# test
test_data, test_states = datas[s,t], states[s,t] # test data
st2pow[s,t], svm2frac, frac_s_str = learn.metaSVM_pow(test_data, test_states, meta_svm, s, t, svm_dict)
# report
fracs = ""
for i, frac in sorted(svm2frac.iteritems()): fracs += "%g\t" % frac
if(frac_s_str == None):
to_write = "%g\t%i\t%g\t%s" % (s, t, st2pow[s,t], fracs)
else:
to_write = "%g\t%i\t%g\t%s\t%s" % (s, t, st2pow[s,t], fracs, frac_s_str)
print to_write
pfile_text.append(to_write)
write_to_file('power_SVM_general.txt', pfile_text)
########### make final meta-SVM and pickle ###########
sys.stdout.write('Training final meta-SVM... ')
train_mix = data_mix_for_SVM_general(p.first_sim, p.last_sim)
# even svm-weights & error terms
svm_weights, svm_e_terms = {}, {}
for i in range(len(train_mix)/2):
svm_weights[i], svm_e_terms[i] = 1.0, 0.1
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
print 'done.\n'
with open(gen_svm_file_to_write, mode='wb') as gen_svm_fh:
pickle.dump(meta_svm, gen_svm_fh)
def general_SVC_partition_data(gen_svm_file=None, spec_svm_file=None):
''' Train general SVM according to regimes on partition 1, test on partition 2.
Compute power at all (s,t)'s.
'''
global st2pow
partition_at = 300 # 0 test all data, 300 train & test, 500 train all data
pfile_text = []
svm_dict = None
if (gen_svm_file is None):
############ train meta-SVM on training-set #############
sys.stdout.write('Training meta-SVM... ')
# get training data
train_mix = data_mix_for_SVM_general(p.first_sim, partition_at)
# set equal regime-svm weights and error terms
regimes_ind = [i for i in range(len(train_mix) / 2)]
svm_weights = dict(zip(regimes_ind, [1.0] * len(regimes_ind)))
svm_e_terms = dict(zip(regimes_ind, [0.1] * len(regimes_ind)))
# train
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
print 'done.\n'
# if this was only for training, pickle and exit
if (partition_at == p.last_sim):
print 'Pickling SVM, then exiting...'
with open(gen_svm_file_to_write, mode='wb') as gen_svm_fh:
pickle.dump(meta_svm, gen_svm_fh)
sys.exit(0)
else:
############# read SVM from pickled file ##############
sys.stdout.write("\nReading in cPickled SVM from: %s... " %
gen_svm_file)
with open("../" + gen_svm_file, mode='rb') as svm_fh:
meta_svm = pickle.load(svm_fh)
print "loaded.\n"
if (spec_svm_file != None):
# read (s,t)-specific SVM dict from pickled file
sys.stdout.write("\nReading in cPickled SVM-dict from: %s... " %
spec_svm_file)
with open("../" + spec_svm_file, mode='rb') as svm_fh:
svm_dict = pickle.load(svm_fh)
print "loaded.\n"
#dotprod_and_decison_func_of(meta_svm, 0.02,700 , regimes.regime_of(0.02,700) )
#dotprod_and_decison_func_of(meta_svm, 0.02,1400, regimes.regime_of(0.02,1400))
################# apply SVM to test data ##################
for s, t in [(s, t) for s in p.selection for t in p.times]:
# evaluate power on test set
test_data, test_states = data_and_states_for(p.case_type, p.cont_type,
[s], [t], partition_at,
p.last_sim) # test data
print "test-data: %i, test-states: %i" % (len(test_data),
len(test_states))
st2pow[s, t], svm2frac, frac_s_str = learn.metaSVM_pow(
test_data, test_states, meta_svm, s, t, svm_dict)
# report
fracs = ""
for i, frac in sorted(svm2frac.iteritems()):
fracs += "%g\t" % frac
if (frac_s_str == None):
to_write = "%g\t%i\t%g\t%s" % (s, t, st2pow[s, t], fracs)
else:
to_write = "%g\t%i\t%g\t%s\t%s" % (s, t, st2pow[s, t], fracs,
frac_s_str)
print to_write
pfile_text.append(to_write)
# write power stats for file
write_to_file('power_SVM_general.txt', pfile_text)
######## if none input, make & pickle the final meta-SVM ########
if (gen_svm_file is None):
# train meta-SVM on training-set
sys.stdout.write('Training final meta-SVM... ')
train_mix = data_mix_for_SVM_general(p.first_sim, p.last_sim)
# set equal regime-svm weights and error terms
regimes_ind = [i for i in range(len(train_mix) / 2)]
svm_weights = dict(zip(regimes_ind, [1.0] * len(regimes_ind)))
svm_e_terms = dict(zip(regimes_ind, [0.1] * len(regimes_ind)))
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
print 'done.\n'
with open(gen_svm_file_to_write, mode='wb') as gen_svm_fh:
pickle.dump(meta_svm, gen_svm_fh)
def general_SVC_exclude_test_set(spec_svm_file=None):
''' Train the general SVM according to regimes on all data excluding the test data.
Compute power at all (s,t)'s.
'''
global st2pow
pfile_text = []
########## read (s,t)-specific SVMs from file ##########
if (spec_svm_file != None):
# read specific SVM dict from pickled file
sys.stdout.write("Reading in cPickled SVM-dict from: %s... " %
spec_svm_file)
with open("../" + spec_svm_file, mode='rb') as svm_fh:
svm_dict = pickle.load(svm_fh)
print "loaded.\n"
############## train & test on all (s,t) ###############
for s, t in [(s, t) for s in p.selection for t in p.times]:
# training data, EXCLUDING test data
print "training (excluded) on (%g,%i)" % (s, t)
train_mix = data_mix_for_SVM_general_exclude(s, t)
svm_weights, svm_e_terms = {}, {}
for i in range(len(train_mix) / 2):
svm_weights[i], svm_e_terms[i] = 1.0, 0.1
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
# test
test_data, test_states = datas[s, t], states[s, t] # test data
st2pow[s, t], svm2frac, frac_s_str = learn.metaSVM_pow(
test_data, test_states, meta_svm, s, t, svm_dict)
# report
fracs = ""
for i, frac in sorted(svm2frac.iteritems()):
fracs += "%g\t" % frac
if (frac_s_str == None):
to_write = "%g\t%i\t%g\t%s" % (s, t, st2pow[s, t], fracs)
else:
to_write = "%g\t%i\t%g\t%s\t%s" % (s, t, st2pow[s, t], fracs,
frac_s_str)
print to_write
pfile_text.append(to_write)
write_to_file('power_SVM_general.txt', pfile_text)
########### make final meta-SVM and pickle ###########
sys.stdout.write('Training final meta-SVM... ')
train_mix = data_mix_for_SVM_general(p.first_sim, p.last_sim)
# even svm-weights & error terms
svm_weights, svm_e_terms = {}, {}
for i in range(len(train_mix) / 2):
svm_weights[i], svm_e_terms[i] = 1.0, 0.1
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
print 'done.\n'
with open(gen_svm_file_to_write, mode='wb') as gen_svm_fh:
pickle.dump(meta_svm, gen_svm_fh)
def general_SVC_partition_data(gen_svm_file=None, spec_svm_file=None):
''' Train general SVM according to regimes on partition 1, test on partition 2.
Compute power at all (s,t)'s.
'''
global st2pow
partition_at = 300 # 0 test all data, 300 train & test, 500 train all data
pfile_text = []
svm_dict = None
if(gen_svm_file is None):
############ train meta-SVM on training-set #############
sys.stdout.write('Training meta-SVM... ')
# get training data
train_mix = data_mix_for_SVM_general(p.first_sim, partition_at)
# set equal regime-svm weights and error terms
regimes_ind = [i for i in range(len(train_mix)/2)]
svm_weights = dict( zip(regimes_ind, [1.0]*len(regimes_ind)) )
svm_e_terms = dict( zip(regimes_ind, [0.1]*len(regimes_ind)) )
# train
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
print 'done.\n'
# if this was only for training, pickle and exit
if(partition_at == p.last_sim):
print 'Pickling SVM, then exiting...'
with open(gen_svm_file_to_write, mode='wb') as gen_svm_fh:
pickle.dump(meta_svm, gen_svm_fh)
sys.exit(0)
else:
############# read SVM from pickled file ##############
sys.stdout.write("\nReading in cPickled SVM from: %s... " % gen_svm_file)
with open("../" + gen_svm_file, mode='rb') as svm_fh:
meta_svm = pickle.load(svm_fh)
print "loaded.\n"
if(spec_svm_file != None):
# read (s,t)-specific SVM dict from pickled file
sys.stdout.write("\nReading in cPickled SVM-dict from: %s... " % spec_svm_file)
with open("../" + spec_svm_file, mode='rb') as svm_fh:
svm_dict = pickle.load(svm_fh)
print "loaded.\n"
#dotprod_and_decison_func_of(meta_svm, 0.02,700 , regimes.regime_of(0.02,700) )
#dotprod_and_decison_func_of(meta_svm, 0.02,1400, regimes.regime_of(0.02,1400))
################# apply SVM to test data ##################
for s,t in [(s,t) for s in p.selection for t in p.times]:
# evaluate power on test set
test_data, test_states = data_and_states_for(p.case_type, p.cont_type, [s], [t], partition_at, p.last_sim) # test data
print "test-data: %i, test-states: %i" % (len(test_data), len(test_states))
st2pow[s,t], svm2frac, frac_s_str = learn.metaSVM_pow(test_data, test_states, meta_svm, s, t, svm_dict)
# report
fracs = ""
for i, frac in sorted(svm2frac.iteritems()): fracs += "%g\t" % frac
if(frac_s_str == None):
to_write = "%g\t%i\t%g\t%s" % (s, t, st2pow[s,t], fracs)
else:
to_write = "%g\t%i\t%g\t%s\t%s" % (s, t, st2pow[s,t], fracs, frac_s_str)
print to_write
pfile_text.append(to_write)
# write power stats for file
write_to_file('power_SVM_general.txt', pfile_text)
######## if none input, make & pickle the final meta-SVM ########
if(gen_svm_file is None):
# train meta-SVM on training-set
sys.stdout.write('Training final meta-SVM... ')
train_mix = data_mix_for_SVM_general(p.first_sim, p.last_sim)
# set equal regime-svm weights and error terms
regimes_ind = [i for i in range(len(train_mix)/2)]
svm_weights = dict( zip(regimes_ind, [1.0]*len(regimes_ind)) )
svm_e_terms = dict( zip(regimes_ind, [0.1]*len(regimes_ind)) )
meta_svm = learn.train_metaSVM(train_mix, svm_weights, svm_e_terms)
print 'done.\n'
with open(gen_svm_file_to_write, mode='wb') as gen_svm_fh:
pickle.dump(meta_svm, gen_svm_fh)