def test__apply_activate_(self):
"""Test `_apply_activate` for tables with both continuous and categoricals.
Check every continuous column has all values between -1 and 1
(since they are normalized), and check every categorical column adds up to 1.
Setup:
- Mock `self._transformer.output_info_list`
Input:
- data = tensor of shape (N, data_dims)
Output:
- tensor = tensor of shape (N, data_dims)
"""
model = CTGANSynthesizer()
model._transformer = Mock()
model._transformer.output_info_list = [[SpanInfo(
3, 'softmax')], [SpanInfo(1, 'tanh'),
SpanInfo(2, 'softmax')]]
data = torch.randn(100, 6)
result = model._apply_activate(data)
assert result.shape == (100, 6)
_assert_is_between(result[:, 0:3], 0.0, 1.0)
_assert_is_between(result[:3], -1.0, 1.0)
_assert_is_between(result[:, 4:6], 0.0, 1.0)
def test__cond_loss(self):
"""Test `_cond_loss`.
Test that the loss is purely a function of the target categorical.
Setup:
- mock transformer.output_info_list
- create two categoricals, one continuous
- compute the conditional loss, conditioned on the 1st categorical
- compare the loss to the cross-entropy of the 1st categorical, manually computed
Input:
data - the synthetic data generated by the model
c - a tensor with the same shape as the data but with only a specific one-hot vector
corresponding to the target column filled in
m - binary mask used to select the categorical column to condition on
Output:
loss scalar; this should only be affected by the target column
Note:
- even though the implementation of this is probably right, I'm not sure if the idea
behind it is correct
"""
model = CTGANSynthesizer()
model._transformer = Mock()
model._transformer.output_info_list = [
[SpanInfo(1, 'tanh'), SpanInfo(2, 'softmax')],
[SpanInfo(3, 'softmax')
], # this is the categorical column we are conditioning on
[SpanInfo(2, 'softmax')
], # this is the categorical column we are bry jrbec on
]
data = torch.tensor([
# first 3 dims ignored, next 3 dims are the prediction, last 2 dims are ignored
[0.0, -1.0, 0.0, 0.05, 0.05, 0.9, 0.1, 0.4],
])
c = torch.tensor([
# first 3 dims are a one-hot for the categorical,
# next 2 are for a different categorical that we are not conditioning on
# (continuous values are not stored in this tensor)
[0.0, 0.0, 1.0, 0.0, 0.0],
])
# this indicates that we are conditioning on the first categorical
m = torch.tensor([[1, 0]])
result = model._cond_loss(data, c, m)
expected = torch.nn.functional.cross_entropy(
torch.tensor([
[0.05, 0.05, 0.9], # 3 categories, one hot
]),
torch.tensor([2]))
assert (result - expected).abs() < 1e-3