def test_conversion(self):
a = np.arange(10, dtype=np.float32).reshape((5, 2))
b = csr_matrix(a)
c = torch.from_numpy(a).to(utils.device)
for arr in (a, b, c):
for t in utils.OutputType:
for dtype in (
torch.float32,
torch.float64) if t is utils.OutputType.tensor else (
np.float32, np.float64):
utils.convert(arr, t, dtype) # These should all pass
with self.assertRaises(ValueError):
utils.convert("a", utils.OutputType.tensor) # This shouldn't work
def forward(self, x):
embeds = [embedding(x) for embedding in self.embeddings]
is_list = isinstance(embeds[0], list) # happens for word embeddings before pooling
types = set([OutputType.from_object(embed[0]) if is_list else OutputType.from_object(embed) for embed in embeds])
if OutputType.tensor in types: # need to convert everything to torch
embeds = convert(embeds, OutputType.tensor)
cat_fn = _Concat._tensor_concat
elif len(types) == 2: # both numpy and sparse
embeds = convert(embeds, OutputType.array)
cat_fn = _Concat._array_concat
elif OutputType.array in types: # only numpy
cat_fn = _Concat._array_concat
else: # only sparse
cat_fn = _Concat._sparse_concat
if is_list:
return [cat_fn(embed) for embed in zip(*embeds)]
return cat_fn(embeds)
def _convert_to_dtype(self, x, detach=False):
typ = OutputType.tensor if isinstance(
self.dtype, torch.dtype) else OutputType.array
return convert(x, typ, self.dtype, detach=detach)
def fit_transform(self, x, y=None):
needs_fit = self._init_vectorizer()
if needs_fit:
return self.vectorizer.fit_transform(x, convert(y, OutputType.array))
return self.vectorizer.transform(x)
def fit(self, x, y=None):
needs_fit = self._init_vectorizer()
if needs_fit:
self.vectorizer.fit(x, convert(y, OutputType.array))
def _fit_transform(self, x, y=None):
self.model = self.Model(**self.init_args)
return self.model.fit_transform(x, convert(
y, OutputType.array)).astype(np.float32)
def _fit(self, x, y=None):
self.model = self.Model(**self.init_args)
self.model.fit(x, convert(y, OutputType.array))
def _check_input(self, x):
if not isinstance(x, tuple(t.value for t in self.input_types)):
return convert(x, self.input_types[0])
return x