def _get_one_dim(self, val, key, axis):
"""Slice along one axis, keeping the dimensionality of the input"""
if isinstance(key, slice):
if any(k is not None for k in [key.start, key.stop, key.step]):
ix = np.arange(*key.indices(val.shape[axis]))
return O.gather(val, ix, axis=axis)
else:
return val
elif isinstance(key, int):
key %= val.shape[axis]
return O.gather(val, [key], axis=axis)
else:
return O.gather(val, key, axis=axis)
def __call__(self, x):
"""Perform the forward pass"""
embs = [
O.gather(self.embeddings[i](), x[:, i])
for i in range(len(self.embeddings))
]
return O.cat(embs, -1)
def test_gather():
"""Tests gather"""
pf.set_backend('pytorch')
# Should lookup along 1st axis by default
vals = torch.Tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])
inds = torch.LongTensor([0, 1, 2, 1, 0])
output = ops.gather(vals, inds)
assert output.ndim == 2
assert output.shape[0] == 5
assert output.shape[1] == 2
assert output.numpy()[0, 0] == 1.0
assert output.numpy()[0, 1] == 2.0
assert output.numpy()[1, 0] == 3.0
assert output.numpy()[1, 1] == 4.0
assert output.numpy()[2, 0] == 5.0
assert output.numpy()[2, 1] == 6.0
assert output.numpy()[3, 0] == 3.0
assert output.numpy()[3, 1] == 4.0
assert output.numpy()[4, 0] == 1.0
assert output.numpy()[4, 1] == 2.0
# But can set axis
inds = torch.LongTensor([1, 0, 1, 0])
output = ops.gather(vals, inds, axis=1)
assert output.ndim == 2
assert output.shape[0] == 3
assert output.shape[1] == 4
assert output.numpy()[0, 0] == 2.0
assert output.numpy()[1, 0] == 4.0
assert output.numpy()[2, 0] == 6.0
assert output.numpy()[0, 1] == 1.0
assert output.numpy()[1, 1] == 3.0
assert output.numpy()[2, 1] == 5.0
assert output.numpy()[0, 2] == 2.0
assert output.numpy()[1, 2] == 4.0
assert output.numpy()[2, 2] == 6.0
assert output.numpy()[0, 3] == 1.0
assert output.numpy()[1, 3] == 3.0
assert output.numpy()[2, 3] == 5.0
def test_gather():
"""Tests gather"""
# Should lookup along 1st axis by default
vals = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])
inds = tf.constant([0, 1, 2, 1, 0], dtype=tf.int32)
output = ops.gather(vals, inds)
assert output.ndim == 2
assert output.shape[0] == 5
assert output.shape[1] == 2
assert output.numpy()[0, 0] == 1.0
assert output.numpy()[0, 1] == 2.0
assert output.numpy()[1, 0] == 3.0
assert output.numpy()[1, 1] == 4.0
assert output.numpy()[2, 0] == 5.0
assert output.numpy()[2, 1] == 6.0
assert output.numpy()[3, 0] == 3.0
assert output.numpy()[3, 1] == 4.0
assert output.numpy()[4, 0] == 1.0
assert output.numpy()[4, 1] == 2.0
# But can set axis
inds = tf.constant([1, 0, 1, 0])
output = ops.gather(vals, inds, axis=1)
assert output.ndim == 2
assert output.shape[0] == 3
assert output.shape[1] == 4
assert output.numpy()[0, 0] == 2.0
assert output.numpy()[1, 0] == 4.0
assert output.numpy()[2, 0] == 6.0
assert output.numpy()[0, 1] == 1.0
assert output.numpy()[1, 1] == 3.0
assert output.numpy()[2, 1] == 5.0
assert output.numpy()[0, 2] == 2.0
assert output.numpy()[1, 2] == 4.0
assert output.numpy()[2, 2] == 6.0
assert output.numpy()[0, 3] == 1.0
assert output.numpy()[1, 3] == 3.0
assert output.numpy()[2, 3] == 5.0
def __call__(self, x):
"""Perform the forward pass"""
return O.gather(self.embeddings(), x)