def run_exp(param):
np.random.seed(param.seed)
net = nncfg.create_net(param)
print('network configure end, start loading data ...')
# load in data
train_images, train_labels = load(list(range(10)), 'training',
param.path_data)
test_images, test_labels = load(list(range(10)), 'testing',
param.path_data)
# create a batch data
# nbatch: batch size
# doshuffle: True, shuffle the data
# scale: 1.0/256 scale by this factor so all features are in [0,1]
train_xdata, train_ylabel = nncfg.create_batch(train_images, train_labels,
param.batch_size, True,
1.0 / 256.0)
test_xdata, test_ylabel = nncfg.create_batch(test_images, test_labels,
param.batch_size, True,
1.0 / 256.0)
# split validation set
ntrain = train_xdata.shape[0]
nvalid = 10000
assert nvalid % param.batch_size == 0
nvalid = int(nvalid / param.batch_size)
valid_xdata, valid_ylabel = train_xdata[0:nvalid], train_ylabel[0:nvalid]
train_xdata, train_ylabel = train_xdata[nvalid:ntrain], train_ylabel[
nvalid:ntrain]
# setup evaluator
evals = []
evals.append(
nnet.NNEvaluator(net, train_xdata, train_ylabel, param, 'train'))
evals.append(
nnet.NNEvaluator(net, valid_xdata, valid_ylabel, param, 'valid'))
evals.append(nnet.NNEvaluator(net, test_xdata, test_ylabel, param, 'test'))
# set parameters
param.num_train = train_ylabel.size
print('loading end,%d train,%d valid,%d test, start update ...' %
(train_ylabel.size, valid_ylabel.size, test_ylabel.size))
# By default, `net` has number of updaters equal to 2x the amount of network
# weights we store, e.g. in the paper they said A, B, a, b, so each has a
# separate SGHMC updater and also its own hyperparameter updater.
for it in range(param.num_round):
param.set_round(it)
net.update_all(train_xdata, train_ylabel)
sys.stderr.write('[%d]' % it)
for ev in evals:
ev.eval(it, sys.stderr)
sys.stderr.write('\n')
print('all update end')
def run_exp(param):
np.random.seed(param.seed)
net = nncfg.create_net(param)
print('network configure end, start loading data ...')
# load in data
train_images, train_labels = load(range(10), 'training', param.path_data)
test_images, test_labels = load(range(10), 'testing', param.path_data)
# create a batch data
# nbatch: batch size
# doshuffle: True, shuffle the data
# scale: 1.0/256 scale by this factor so all features are in [0,1]
train_xdata, train_ylabel = nncfg.create_batch(train_images, train_labels,
param.batch_size, True,
1.0 / 256.0)
test_xdata, test_ylabel = nncfg.create_batch(test_images, test_labels,
param.batch_size, True,
1.0 / 256.0)
# split validation set
ntrain = train_xdata.shape[0]
nvalid = 10000
assert nvalid % param.batch_size == 0
nvalid = nvalid // param.batch_size
valid_xdata, valid_ylabel = train_xdata[0:nvalid], train_ylabel[0:nvalid]
train_xdata, train_ylabel = train_xdata[nvalid:ntrain], train_ylabel[
nvalid:ntrain]
# setup evaluator
evals = []
evals.append(
nnet.NNEvaluator(net, train_xdata, train_ylabel, param, 'train'))
evals.append(
nnet.NNEvaluator(net, valid_xdata, valid_ylabel, param, 'valid'))
evals.append(nnet.NNEvaluator(net, test_xdata, test_ylabel, param, 'test'))
# set parameters
param.num_train = train_ylabel.size
print('loading end,%d train,%d valid,%d test, start update ...' %
(train_ylabel.size, valid_ylabel.size, test_ylabel.size))
for it in range(param.num_round):
param.set_round(it)
net.update_all(train_xdata, train_ylabel)
sys.stderr.write('[%d]' % it)
for ev in evals:
ev.eval(it, sys.stderr)
sys.stderr.write('\n')
print('all update end')
def run_exp(param):
np.random.seed(param.seed)
net = nncfg.mlp2layer(param)
print('network configure end, start loading data ...')
# load in data
train_images, train_labels = load(range(10), 'training', param.path_data)
test_images, test_labels = load(range(10), 'testing', param.path_data)
# create a batch data
# nbatch: batch size
# doshuffle: True, shuffle the data
# scale: 1.0/256 scale by this factor so all features are in [0,1]
train_xdata, train_ylabel = nncfg.create_batch(train_images, train_labels,
param.batch_size, True,
1.0 / 256.0)
test_xdata, test_ylabel = nncfg.create_batch(test_images, test_labels,
param.batch_size, True,
1.0 / 256.0)
# split validation set
ntrain = train_xdata.shape[0]
nvalid = 10000
assert nvalid % param.batch_size == 0
nvalid = int(nvalid / param.batch_size)
valid_xdata, valid_ylabel = train_xdata[0:nvalid], train_ylabel[0:nvalid]
train_xdata, train_ylabel = train_xdata[nvalid:ntrain], train_ylabel[
nvalid:ntrain]
# setup evaluator
evals = []
evals.append(
nnet.NNEvaluator(net, train_xdata, train_ylabel, param, 'train'))
evals.append(
nnet.NNEvaluator(net, valid_xdata, valid_ylabel, param, 'valid'))
evals.append(nnet.NNEvaluator(net, test_xdata, test_ylabel, param, 'test'))
# set parameters
param.num_train = train_ylabel.size
print('loading end,%d train,%d valid,%d test, start update ...' %
(train_ylabel.size, valid_ylabel.size, test_ylabel.size))
modelevals = []
for it in range(param.num_round):
param.set_round(it)
net.update_all(train_xdata, train_ylabel)
sys.stderr.write('[%d]' % it)
for ev in evals:
ev.eval(it, sys.stderr)
modelevals.append(ev.results)
sys.stderr.write('\n')
modelevals = np.array(modelevals).reshape(param.num_round, -1)
inds = np.arange(24, param.num_round, 10)
plt.plot(inds, modelevals[inds, 4], 'x-.', color='purple')
plt.ylim(0.015, 0.05)
plt.ylabel('test error')
plt.xlabel('iteration')
plt.savefig('test-err.png')
print('all update end')