def test_np_reshape():
# TODO(junwu): Add more test cases
data = mx.sym.var('a').as_np_ndarray()
ret = data.reshape(shape=())
assert type(ret) == mx.sym.np._Symbol
data = np.ones((1, 1, 1))
ret = np.reshape(data, ())
assert ret.shape == ()
ret = np.reshape(ret, (1, 1, 1, 1))
assert ret.shape == (1, 1, 1, 1)
assert type(ret) == np.ndarray
def forward(self, x):
embed_x = self.embedding(x)
square_of_sum = np.sum(embed_x, axis=1)**2
sum_of_square = np.sum(embed_x**2, axis=1)
inputs = np.reshape(embed_x, (-1, self.embed_output_dim))
x = self.linear_layer(self.fc(x).sum(1)) \
+ 0.5 * (square_of_sum - sum_of_square).sum(1, keepdims=True) \
+ self.mlp(inputs)
x = npx.sigmoid(x)
return x
def __call__(self, scores, offset):
"""
Get the lowest k elements per sentence from a `scores` matrix.
:param scores: Vocabulary scores for the next beam step. (batch_size * beam_size, target_vocabulary_size)
:param offset: Array to add to the hypothesis indices for offsetting in batch decoding.
:return: The row indices, column indices and values of the k smallest items in matrix.
"""
batch_times_beam, vocab_size = scores.shape
batch_size = int(batch_times_beam / self.k)
# Shape: (batch size, beam_size * vocab_size)
batchwise_scores = np.reshape(scores, (batch_size, self.k * vocab_size))
indices, values = super().__call__(batchwise_scores)
best_hyp_indices, best_word_indices = np.unravel_index(indices, shape=(batch_size * self.k, vocab_size))
if batch_size > 1:
# Offsetting the indices to match the shape of the scores matrix
best_hyp_indices = best_hyp_indices + offset
return best_hyp_indices, best_word_indices, values
def test_np_reshape():
class TestReshape(HybridBlock):
def __init__(self, newshape):
super(TestReshape, self).__init__()
self._newshape = newshape
def hybrid_forward(self, F, a):
return F.np.reshape(a, self._newshape)
shape_pairs = [((2, 6), (6, 2)), ((2, 6), (3, 4)), ((1, 0), (0, )),
((0, 0), (0, )), ((), (1, 1, 1))]
for hybridize in [True, False]:
for shape_pair in shape_pairs:
shape1, shape2 = shape_pair
print(shape1, shape2)
test_reshape = TestReshape(shape2)
if hybridize:
test_reshape.hybridize()
x = rand_ndarray(shape1).as_np_ndarray()
x.attach_grad()
np_out = _np.reshape(x.asnumpy(), shape2)
with mx.autograd.record():
mx_out = test_reshape(x)
assert mx_out.shape == np_out.shape
assert_almost_equal(mx_out.asnumpy(),
np_out,
rtol=1e-3,
atol=1e-5,
use_broadcast=False)
mx_out.backward()
np_backward = _np.ones(shape1)
assert_almost_equal(x.grad.asnumpy(),
np_backward,
rtol=1e-3,
atol=1e-5,
use_broadcast=False)
mx_out = np.reshape(x, shape2)
np_out = _np.reshape(x.asnumpy(), shape2)
assert_almost_equal(mx_out.asnumpy(),
np_out,
rtol=1e-3,
atol=1e-5,
use_broadcast=False)
def forward(self, source_encoded: np.ndarray,
source_encoded_length: np.ndarray) -> np.ndarray:
"""
Transformation to the length ratio. Returns a vector.
:param source_encoded: Encoder representation for n elements. Shape: (n, source_encoded_length, hidden_size).
:param source_encoded_length: A vector of encoded sequence lengths. Shape: (n,).
:return: Predictions of the ratio length(hypothesis)/length(reference). Shape(n, 1).
"""
# source_masked: (n, source_encoded_length, hidden_size)
source_masked = npx.sequence_mask(
source_encoded,
axis=1,
sequence_length=source_encoded_length,
use_sequence_length=True,
value=0.)
# calculate the proper means of encoded sources
# data: (n, hidden_size)
data = np.sum(source_masked, axis=1, keepdims=False) / np.reshape(
source_encoded_length, (-1, 1))
# MLP. Shape: (n, 1)
data = self.layers(data)
# Shape: (n,)
return np.squeeze(data)
def forward(self, data, indices):
mask = indices < 3
data = npx.reshape(data, (-1, -2), reverse=True)
mask = np.reshape(mask, (-1, ))
sel = nd.np._internal.boolean_mask(data, mask)
return sel
def forward(self, scores):
values, indices = npx.topk(scores, axis=1, k=self.k, ret_typ='both', is_ascend=True, dtype='int32')
# Project indices back into original shape (which is different for t==1 and t>1)
values, indices = np.reshape(values, (-1, 1)), np.reshape(indices, (-1,))
return indices, values