def test_NNET_train(self):
struct = 'w2vec(/project/nakamura-lab01/Work/truong-dq/chainer/vidnn/exp/word2vec_truecase_200/vectors.bin):lstm(200-2):linear(2-2)'
# struct = 'embed(2-200):lstm(200-2):linear(2-2)'
nnet = NNET_Model.parse_structure(struct)
# Testing variable
test_var = chainer.Variable(np.asarray([1], dtype=np.int32))
output_before_train = nnet(test_var).data
inp = chainer.Variable(np.asarray([1], dtype=np.int32))
target = chainer.Variable(np.asarray([0], dtype=np.int32))
output = nnet(inp)
loss = F.softmax_cross_entropy(output, target)
optimizer = optimizers.SGD(lr=0.1)
optimizer.setup(nnet)
optimizer.zero_grads()
loss.backward()
optimizer.update()
inp = chainer.Variable(np.asarray([0], dtype=np.int32))
target = chainer.Variable(np.asarray([0], dtype=np.int32))
output = nnet(inp)
loss = F.softmax_cross_entropy(output, target)
optimizer = optimizers.SGD(lr=0.1)
optimizer.setup(nnet)
optimizer.zero_grads()
loss.backward()
optimizer.update()
nnet.save('test_output')
nnet_2 = NNET_Model.load('test_output')
nnet.forget_history()
nnet_2.forget_history()
np.testing.assert_equal(nnet[0][1].W.data, nnet_2[0][1].W.data)
np.testing.assert_equal(nnet[1][1].upward.W.data, nnet_2[1][1].upward.W.data)
np.testing.assert_equal(nnet[1][1].lateral.W.data, nnet_2[1][1].lateral.W.data)
np.testing.assert_equal(nnet[1][1].upward.b.data, nnet_2[1][1].upward.b.data)
output_after_train = nnet(test_var).data
output_after_load = nnet_2(test_var).data
after_first_layer_nnet = nnet[0][1](test_var)
after_first_layer_nnet_2 = nnet_2[0][1](test_var)
np.testing.assert_equal(after_first_layer_nnet.data, after_first_layer_nnet_2.data)
after_first_layer_nnet.volatile = False
after_first_layer_nnet_2.volatile = False
after_second_layer_nnet = nnet[1][1](after_first_layer_nnet)
after_second_layer_nnet_2 = nnet_2[1][1](after_first_layer_nnet_2)
np.testing.assert_equal(after_second_layer_nnet.data, after_second_layer_nnet_2.data)
assert (output_before_train != output_after_train).any()
assert (output_before_train != output_after_load).any()
np.testing.assert_equal(output_after_train, output_after_load)
def decode_mode():
nnet_model = NNET_Model.load(opts.fname_in_model)
for row, words in zip(data, text_data):
row_result = []
for w in row:
x = Variable(xp.asarray([w], dtype=np.int32))
row_result.append(nnet_model(x).data.argmax(axis=1)[0])
if opts.forget_on_new_utt:
nnet_model.forget_history()
nice_print(row_result, words)
def test_NNET_save(self):
struct = 'embed(2-2):lstm(2-2)'
nnet = NNET_Model.parse_structure(struct)
nnet.save('test_output')
nnet2 = NNET_Model.load('test_output')
inp = chainer.Variable(np.asarray([1], dtype=np.int32))
np.testing.assert_equal(nnet[0][1].W.data, nnet2[0][1].W.data)
np.testing.assert_equal(nnet[0][1](inp).data, nnet2[0][1](inp).data)
np.testing.assert_equal(nnet(inp).data, nnet2(inp).data)
struct = 'embed(2-2):lstm(2-3):linear(3-2)'
nnet = NNET_Model.parse_structure(struct)
nnet.save('test_output')
nnet2 = NNET_Model.load('test_output')
inp = chainer.Variable(np.asarray([1], dtype=np.int32))
np.testing.assert_equal(nnet(inp).data, nnet2(inp).data)
def test_NNET_load(self):
struct = 'linear(2-4):sigmoid:tanh:dropout:lstm(10-10)'
nnet = NNET_Model.parse_structure(struct)
nnet.save('test_save_model')
another = NNET_Model.load('test_save_model')
def test_NNET_parse_struct(self):
struct = 'linear(2-4):sigmoid:tanh:dropout:lstm(10-10)'
nnet = NNET_Model.parse_structure(struct)
self.assertIsInstance(nnet['linear(2-4)-c0'], L.Linear)
self.assertIsInstance(nnet['lstm(10-10)-c4'], L.LSTM)
def setUp(self):
np.random.seed(0)
if not os.path.exists('test_save_model'):
os.makedirs('test_save_model')
self.nnet_model = NNET_Model()
pass
class Test_NNET_Model(unittest.TestCase):
def setUp(self):
np.random.seed(0)
if not os.path.exists('test_save_model'):
os.makedirs('test_save_model')
self.nnet_model = NNET_Model()
pass
def test_NNET_add_comp(self):
self.nnet_model.add_component('linear(2-2)', L.Linear(2, 2))
def test_NNET_parse_struct(self):
struct = 'linear(2-4):sigmoid:tanh:dropout:lstm(10-10)'
nnet = NNET_Model.parse_structure(struct)
self.assertIsInstance(nnet['linear(2-4)-c0'], L.Linear)
self.assertIsInstance(nnet['lstm(10-10)-c4'], L.LSTM)
def test_NNET_load(self):
struct = 'linear(2-4):sigmoid:tanh:dropout:lstm(10-10)'
nnet = NNET_Model.parse_structure(struct)
nnet.save('test_save_model')
another = NNET_Model.load('test_save_model')
def test_NNET_save(self):
struct = 'embed(2-2):lstm(2-2)'
nnet = NNET_Model.parse_structure(struct)
nnet.save('test_output')
nnet2 = NNET_Model.load('test_output')
inp = chainer.Variable(np.asarray([1], dtype=np.int32))
np.testing.assert_equal(nnet[0][1].W.data, nnet2[0][1].W.data)
np.testing.assert_equal(nnet[0][1](inp).data, nnet2[0][1](inp).data)
np.testing.assert_equal(nnet(inp).data, nnet2(inp).data)
struct = 'embed(2-2):lstm(2-3):linear(3-2)'
nnet = NNET_Model.parse_structure(struct)
nnet.save('test_output')
nnet2 = NNET_Model.load('test_output')
inp = chainer.Variable(np.asarray([1], dtype=np.int32))
np.testing.assert_equal(nnet(inp).data, nnet2(inp).data)
def test_NNET_train(self):
xp = cuda.cupy
# struct = 'w2vec(/project/nakamura-lab01/Work/truong-dq/chainer/vidnn/exp/word2vec_truecase_200/vectors.bin):lstm(200-2):linear(2-2)'
struct = 'embed(2-200):lstm(200-2):linear(2-2)'
nnet = NNET_Model.parse_structure(struct)
nnet.to_gpu()
# Testing variable
test_var = chainer.Variable(xp.asarray([1], dtype=np.int32))
output_before_train = cuda.to_cpu(nnet(test_var).data)
inp = chainer.Variable(xp.asarray([1], dtype=np.int32))
target = chainer.Variable(xp.asarray([0], dtype=np.int32))
output = nnet(inp)
loss = F.softmax_cross_entropy(output, target)
optimizer = optimizers.SGD(lr=0.1)
optimizer.setup(nnet)
optimizer.zero_grads()
loss.backward()
optimizer.update()
inp = chainer.Variable(xp.asarray([0], dtype=np.int32))
target = chainer.Variable(xp.asarray([0], dtype=np.int32))
output = nnet(inp)
loss = F.softmax_cross_entropy(output, target)
optimizer = optimizers.SGD(lr=0.1)
optimizer.setup(nnet)
optimizer.zero_grads()
loss.backward()
optimizer.update()
nnet.save('test_output')
nnet.to_gpu()
nnet_2 = NNET_Model.load('test_output')
nnet_2.to_gpu()
nnet.forget_history()
nnet_2.forget_history()
np.testing.assert_equal(cuda.to_cpu(nnet[0][1].W.data), cuda.to_cpu(nnet_2[0][1].W.data))
np.testing.assert_equal(cuda.to_cpu(nnet[1][1].upward.W.data), cuda.to_cpu(nnet_2[1][1].upward.W.data))
np.testing.assert_equal(cuda.to_cpu(nnet[1][1].lateral.W.data), cuda.to_cpu(nnet_2[1][1].lateral.W.data))
np.testing.assert_equal(cuda.to_cpu(nnet[1][1].upward.b.data), cuda.to_cpu(nnet_2[1][1].upward.b.data))
output_after_train = cuda.to_cpu(nnet(test_var).data)
output_after_load = cuda.to_cpu(nnet_2(test_var).data)
after_first_layer_nnet = nnet[0][1](test_var)
after_first_layer_nnet_2 = nnet_2[0][1](test_var)
np.testing.assert_equal(cuda.to_cpu(after_first_layer_nnet.data), cuda.to_cpu(after_first_layer_nnet_2.data))
after_first_layer_nnet.volatile = False
after_first_layer_nnet_2.volatile = False
after_second_layer_nnet = nnet[1][1](after_first_layer_nnet)
after_second_layer_nnet_2 = nnet_2[1][1](after_first_layer_nnet_2)
np.testing.assert_equal(cuda.to_cpu(after_second_layer_nnet.data), cuda.to_cpu(after_second_layer_nnet_2.data))
assert (output_before_train != output_after_train).any()
assert (output_before_train != output_after_load).any()
np.testing.assert_equal(output_after_train, output_after_load)
# def test_NNET_W2vec(self):
# np.random.seed(0)
# struct = 'w2vec(/project/nakamura-lab01/Work/truong-dq/chainer/vidnn/exp/word2vec_truecase_200/vectors.bin):lstm(200-3):linear(3-2)'
# nnet = NNET_Model.parse_structure(struct)
# nnet.save('test_save_model_w2vec')
# another = NNET_Model.load('test_save_model_w2vec')
# inp = chainer.Variable(np.asarray([1], dtype=np.int32))
# np.testing.assert_equal(nnet(inp).data, another(inp).data)
def tearDown(self):
pass
