def test_build_model():
m = Model("test model")
l_in = m.add_layer(xnn.layers.InputLayer(shape=(10, 200)), name="l_in")
l_h1 = m.add_layer(xnn.layers.DenseLayer(l_in, 100), name="l_h1")
l_out = m.add_layer(xnn.layers.DenseLayer(l_h1, 200), name="l_out")
m.bind_input(l_in, "pixels")
m.bind_output(l_h1, xnn.objectives.categorical_crossentropy, "emotions", "label", "mean")
m.bind_output(l_out, xnn.objectives.squared_error, "l_in", "recon", "mean")
def _build_lr_net(batch_size, in_size, out_size):
np.random.seed(100)
m = Model()
l_in = m.add_layer(InputLayer(shape=(batch_size,in_size)))
l_out = m.add_layer(DenseLayer(l_in, out_size, nonlinearity=softmax))
m.bind_input(l_in, "inputs")
m.bind_output(l_out, squared_error, "labels", "label", "mean")
return m, l_in, l_out
def test_trained_model_serialization():
batch_size = 2
img_size = 10
num_hid = 10
m = _build_model(batch_size, img_size, num_hid)
global_update_settings = ParamUpdateSettings(
update=lasagne.updates.nesterov_momentum,
learning_rate=0.1,
momentum=0.2)
trainer = Trainer(m, global_update_settings)
pixels = np.random.rand(batch_size, img_size).astype(theano.config.floatX)
emotions = np.random.rand(batch_size, num_hid).astype(theano.config.floatX)
batch_dict = dict(
# learning_rate_default=0.1,
# momentum_default=0.5,
pixels=pixels,
emotions=emotions)
outs = trainer.train_step(batch_dict)
preds = m.predict(batch_dict)
dirpath = tempfile.mkdtemp()
filepath = os.path.join(dirpath, 'testtrainerout')
m.save_model(filepath)
m2 = Model('load')
m2.load_model(filepath)
shutil.rmtree(dirpath)
global_update_settings = ParamUpdateSettings(
update=lasagne.updates.nesterov_momentum,
learning_rate=0.1,
momentum=0.2)
trainer2 = Trainer(m2, global_update_settings)
preds2 = m.predict(batch_dict)
outs = trainer.train_step(batch_dict)
outs2 = trainer2.train_step(batch_dict)
assert m.outputs.keys() == m2.outputs.keys()
for p, p2 in zip(preds.values(), preds2.values()):
assert np.allclose(p, p2)
for o, o2 in zip(outs, outs2):
assert np.allclose(o, o2)
def _build_lr_net(batch_size, in_size, out_size):
np.random.seed(100)
m = Model()
l_in = m.add_layer(InputLayer(shape=(batch_size, in_size)))
l_out = m.add_layer(DenseLayer(l_in, out_size, nonlinearity=softmax))
m.bind_input(l_in, "inputs")
m.bind_output(l_out, squared_error, "labels", "label", "mean")
return m, l_in, l_out
def test_trained_model_serialization():
batch_size = 2
img_size = 10
num_hid = 10
m = _build_model(batch_size,img_size,num_hid)
global_update_settings = ParamUpdateSettings(update=lasagne.updates.nesterov_momentum,learning_rate=0.1, momentum=0.2)
trainer = Trainer(m,global_update_settings)
pixels = np.random.rand(batch_size,img_size).astype(theano.config.floatX)
emotions = np.random.rand(batch_size,num_hid).astype(theano.config.floatX)
batch_dict = dict(
# learning_rate_default=0.1,
# momentum_default=0.5,
pixels=pixels,
emotions=emotions
)
outs = trainer.train_step(batch_dict)
preds = m.predict(batch_dict)
dirpath = tempfile.mkdtemp()
filepath = os.path.join(dirpath, 'testtrainerout')
m.save_model(filepath)
m2 = Model('load')
m2.load_model(filepath)
shutil.rmtree(dirpath)
global_update_settings = ParamUpdateSettings(update=lasagne.updates.nesterov_momentum,learning_rate=0.1, momentum=0.2)
trainer2 = Trainer(m2,global_update_settings)
preds2 = m.predict(batch_dict)
outs = trainer.train_step(batch_dict)
outs2 = trainer2.train_step(batch_dict)
assert m.outputs.keys()==m2.outputs.keys()
for p,p2 in zip(preds.values(),preds2.values()):
assert np.allclose(p,p2)
for o,o2 in zip(outs,outs2):
assert np.allclose(o,o2)
def _build_model():
r = xnn.nonlinearities.rectify
s = xnn.nonlinearities.scale
m2 = Model("test convenience")
l_in = m2.make_bound_input_layer((10, 1, 20, 10), "pixels")
l_in2 = m2.make_bound_input_layer((10, 1, 20, 10), "pixels")
l_conv = m2.add_layer(xnn.layers.Conv2DLayer(l_in2, 3, 4), name="l_conv")
l_den = m2.make_dense_drop_stack(
l_in, [60, 3, 2], [0.6, 0.4, 0.3], drop_type_list=["gauss", "gauss", "standard"], nonlin_list=[r, s(2.0), r]
)
l_reshp = m2.add_layer(xnn.layers.ReshapeLayer(l_conv, (10, -1)))
l_mer = xnn.layers.ConcatLayer([l_reshp, l_den])
m2.add_layer(l_mer, name="merger")
l_bn = xnn.layers.BatchNormLayer(l_in2, nonlinearity=xnn.nonlinearities.sigmoid)
m2.add_layer(l_bn)
l_loc = xnn.layers.LocalLayer(l_bn, num_units=32, img_shape=(20, 10), local_filters=[(2, 1)], seed=121212)
np.random.seed(int(time.time() * 10000 % 10000))
m2.add_layer(l_loc)
l_out = m2.add_layer(xnn.layers.DenseLayer(l_loc, num_units=2, nonlinearity=xnn.nonlinearities.linear), name="out")
l_pr = m2.add_layer(xnn.layers.PReLULayer(l_out, xnn.init.Constant(0.25)))
m2.bind_output(l_pr, xnn.objectives.squared_error, "age", "label", "mean", is_eval_output=True, scale=2)
l_out2 = m2.add_layer(
xnn.layers.DenseLayer(l_loc, num_units=2, nonlinearity=xnn.nonlinearities.softmax), name="out2"
)
m2.bind_output(l_out2, xnn.objectives.hinge_loss(threshold=2), "age", "label", "mean", is_eval_output=False)
return m2
def test_serialization():
m2 = _build_model()
serialized_old = m2.to_dict()
print "original model"
pprint.pprint(serialized_old)
m3 = Model("test serialize")
m3.from_dict(serialized_old)
print "model loaded from dict"
pprint.pprint(m3.to_dict())
data = dict(pixels=np.random.rand(10, 1, 20, 10).astype(theano.config.floatX))
dirpath = tempfile.mkdtemp()
filepath = os.path.join(dirpath, "testmodelout")
m3.save_model(filepath)
out3 = m3.predict(data, ["out", "out2"])
m4 = Model("test convenience")
m4.load_model(filepath)
shutil.rmtree(dirpath)
assert serialized_old == m4.to_dict()
assert np.allclose(m4.layers["DenseLayer_2"].W.get_value(), m3.layers["DenseLayer_2"].W.get_value())
assert ~np.allclose(m4.layers["DenseLayer_2"].W.get_value(), m2.layers["DenseLayer_2"].W.get_value())
out4 = m4.predict(data, ["out", "out2"])
print "out pre serialization"
print out3["out"]
print "out post serialization"
print out4["out"]
print "out2 pre serialization"
print out3["out2"]
print "out2 post serialization"
print out4["out2"]
assert np.allclose(out4["out"], out3["out"])
assert np.allclose(out4["out2"], out3["out2"])
def _build_model(batch_size, img_size, num_hid):
m = Model('test model cpu')
l_in = m.add_layer(layers.InputLayer(shape=(batch_size, img_size)),
name="l_in")
l_loc = m.add_layer(
layers.LocalLayer(l_in,
num_units=3,
img_shape=(2, 5),
local_filters=[(2, 1)]))
l_h1 = m.add_layer(layers.DenseLayer(l_loc, num_hid), name="l_h1")
l_out = m.add_layer(layers.DenseLayer(l_h1, img_size), name="l_out")
m.bind_input(l_in, "pixels")
m.bind_output(l_h1, xnn.objectives.kl_divergence, "emotions", "label",
"mean")
m.bind_output(l_out, xnn.objectives.squared_error, "l_in", "recon", "mean")
return m
def test_aggregation():
batch_size = 128
img_size = 10
num_out = 3
m = Model('test model cpu')
l_in = m.add_layer(layers.InputLayer(shape=(batch_size, img_size)),
name="l_in")
l_out_mean = m.add_layer(layers.DenseLayer(l_in, num_out),
name="l_out_mean")
l_out_sum = m.add_layer(layers.DenseLayer(l_in, num_out), name="l_out_sum")
m.bind_input(l_in, "pixels")
m.bind_output(l_out_mean, lasagne.objectives.squared_error, "emotions",
"label", "mean")
m.bind_output(l_out_sum, lasagne.objectives.squared_error, "emotions",
"label", "sum")
from pprint import pprint
pprint(m.to_dict())
global_update_settings = ParamUpdateSettings(
update=lasagne.updates.nesterov_momentum,
learning_rate=0.1,
momentum=0.5)
trainer = Trainer(m, global_update_settings)
pixels = np.random.rand(batch_size, img_size).astype(theano.config.floatX)
emotions = np.random.rand(batch_size, num_out).astype(theano.config.floatX)
batch_dict = dict(
# learning_rate_default=0.1,
# momentum_default=0.5,
pixels=pixels,
emotions=emotions,
exmotions=emotions.copy())
outs = trainer.train_step(batch_dict)
print "Aggregation test succeeded"
def _build_model():
r = xnn.nonlinearities.rectify
s = xnn.nonlinearities.scale
m2 = Model('test convenience')
l_in = m2.make_bound_input_layer((10, 1, 20, 10), 'pixels')
l_in2 = m2.make_bound_input_layer((10, 1, 20, 10), 'pixels')
l_conv = m2.add_layer(xnn.layers.Conv2DLayer(l_in2, 3, 4), name='l_conv')
l_den = m2.make_dense_drop_stack(
l_in, [60, 3, 2], [.6, .4, .3],
drop_type_list=['gauss', 'gauss', 'standard'],
nonlin_list=[r, s(2.), r])
l_reshp = m2.add_layer(xnn.layers.ReshapeLayer(l_conv, (10, -1)))
l_mer = xnn.layers.ConcatLayer([l_reshp, l_den])
m2.add_layer(l_mer, name='merger')
l_bn = xnn.layers.BatchNormLayer(l_in2,
nonlinearity=xnn.nonlinearities.sigmoid)
m2.add_layer(l_bn)
l_loc = xnn.layers.LocalLayer(l_bn,
num_units=32,
img_shape=(20, 10),
local_filters=[(2, 1)],
seed=121212)
np.random.seed(int(time.time() * 10000 % 10000))
m2.add_layer(l_loc)
l_out = m2.add_layer(xnn.layers.DenseLayer(
l_loc, num_units=2, nonlinearity=xnn.nonlinearities.linear),
name='out')
l_pr = m2.add_layer(xnn.layers.PReLULayer(l_out, xnn.init.Constant(.25)))
m2.bind_output(l_pr,
xnn.objectives.squared_error,
'age',
'label',
'mean',
is_eval_output=True,
scale=2)
l_out2 = m2.add_layer(xnn.layers.DenseLayer(
l_loc, num_units=2, nonlinearity=xnn.nonlinearities.softmax),
name='out2')
m2.bind_output(l_out2,
xnn.objectives.hinge_loss(threshold=2),
'age',
'label',
'mean',
is_eval_output=False)
return m2
def test_serialization():
m2 = _build_model()
serialized_old = m2.to_dict()
print "original model"
pprint.pprint(serialized_old)
m3 = Model('test serialize')
m3.from_dict(serialized_old)
print "model loaded from dict"
pprint.pprint(m3.to_dict())
data = dict(
pixels=np.random.rand(10, 1, 20, 10).astype(theano.config.floatX))
dirpath = tempfile.mkdtemp()
filepath = os.path.join(dirpath, 'testmodelout')
m3.save_model(filepath)
out3 = m3.predict(data, ['out', 'out2'])
m4 = Model('test convenience')
m4.load_model(filepath)
shutil.rmtree(dirpath)
assert serialized_old == m4.to_dict()
assert np.allclose(m4.layers['DenseLayer_2'].W.get_value(),
m3.layers['DenseLayer_2'].W.get_value())
assert ~np.allclose(m4.layers['DenseLayer_2'].W.get_value(),
m2.layers['DenseLayer_2'].W.get_value())
out4 = m4.predict(data, ['out', 'out2'])
print 'out pre serialization'
print out3['out']
print 'out post serialization'
print out4['out']
print 'out2 pre serialization'
print out3['out2']
print 'out2 post serialization'
print out4['out2']
assert np.allclose(out4['out'], out3['out'])
assert np.allclose(out4['out2'], out3['out2'])
def test_build_model():
m = Model('test model')
l_in = m.add_layer(xnn.layers.InputLayer(shape=(10, 200)), name="l_in")
l_h1 = m.add_layer(xnn.layers.DenseLayer(l_in, 100), name="l_h1")
l_out = m.add_layer(xnn.layers.DenseLayer(l_h1, 200), name="l_out")
m.bind_input(l_in, "pixels")
m.bind_output(l_h1, xnn.objectives.categorical_crossentropy, "emotions",
"label", "mean")
m.bind_output(l_out, xnn.objectives.squared_error, "l_in", "recon", "mean")
def _build_model(batch_size,img_size,num_hid):
m = Model('test model cpu')
l_in = m.add_layer(layers.InputLayer(shape=(batch_size,img_size)), name="l_in")
l_loc = m.add_layer(layers.LocalLayer(l_in,num_units=3,img_shape=(2,5),local_filters=[(2,1)]))
l_h1 = m.add_layer(layers.DenseLayer(l_loc, num_hid), name="l_h1")
l_out = m.add_layer(layers.DenseLayer(l_h1, img_size), name="l_out")
m.bind_input(l_in, "pixels")
m.bind_output(l_h1, xnn.objectives.kl_divergence, "emotions", "label", "mean")
m.bind_output(l_out, xnn.objectives.squared_error, "l_in", "recon", "mean")
return m
def test_aggregation():
batch_size = 128
img_size = 10
num_out = 3
m = Model('test model cpu')
l_in = m.add_layer(layers.InputLayer(shape=(batch_size,img_size)), name="l_in")
l_out_mean = m.add_layer(layers.DenseLayer(l_in, num_out), name="l_out_mean")
l_out_sum = m.add_layer(layers.DenseLayer(l_in, num_out), name="l_out_sum")
m.bind_input(l_in, "pixels")
m.bind_output(l_out_mean, lasagne.objectives.squared_error, "emotions", "label", "mean")
m.bind_output(l_out_sum, lasagne.objectives.squared_error, "emotions", "label", "sum")
from pprint import pprint
pprint(m.to_dict())
global_update_settings = ParamUpdateSettings(update=lasagne.updates.nesterov_momentum,learning_rate=0.1, momentum=0.5)
trainer = Trainer(m, global_update_settings)
pixels = np.random.rand(batch_size,img_size).astype(theano.config.floatX)
emotions = np.random.rand(batch_size,num_out).astype(theano.config.floatX)
batch_dict = dict(
# learning_rate_default=0.1,
# momentum_default=0.5,
pixels=pixels,
emotions=emotions,
exmotions=emotions.copy()
)
outs = trainer.train_step(batch_dict)
print "Aggregation test succeeded"