search

jmschrei / pomegranate

Fast, flexible and easy to use probabilistic modelling in Python.
http://pomegranate.readthedocs.org/en/latest/
MIT License
3.35k stars 589 forks source link

conditional_categorical.py : RuntimeError: scatter(): Expected self.dtype to be equal to src.dtype[BUG] #1091

Closed frewenta closed 2 months ago

frewenta commented 5 months ago

I have a Bayesian Net model as a pomegranate object and just need to update the parameters so it seems like I should be able to use model.fit. As a sanity check, I pass 1000 samples I generate into the fit routine (I realize that should not change the parameters significantly since it samples from the model). The samples generated match expected output (based on node & edge structure I define) 

samples = model.sample(1000)
model.fit(samples)

This fails with: "RuntimeError: scatter(): Expected self.dtype to be equal to src.dtype' ( in conditional_categorical.py", line 168, in summarize: self._xw_sum[j].view(-1).scatteradd(0, X_, sample_weight[:,j]) ). When I check the dtypes in conditional_categorical.py I see:

self.dtype Out[1]: torch.float32 X.dtype Out[2]: torch.int32

I then tried :

samples = samples.type(torch.float32)
model.fit(samples)

but this fails with: ValueError: Parameter X dtype must be one of (torch.int32, torch.int64)

My conda env has pomegranate 1.0.3 (also torchegranate 0.5.0) . Not sure if this is more of a PyTorch issue; I have seen this "Pytorch issue: add dtype checks for scatter/gather family of functions" https://github.com/pytorch/pytorch/pull/38646

I am able to get the following simple/trivial test case to work (the only difference with the above example is that above I appear to have some torch.float64 types in the transition probabilities (based on a more complex computation to arrive at them) )

d0 = Categorical([[0.05, 0.05, 0.9]])
d1 = Categorical([[0.1, 0.8, 0.1]])
d2 = ConditionalCategorical([[[0.4, 0.6, 0], [0.3, 0.6, 0.1], [0.3, 0.6, 0.1]]])
d3 = Categorical([[0.1, 0.8, 0.1]])
d4 = Categorical([[0.1, 0.8, 0.1]])
d5 = Categorical([[0.1, 0.8, 0.1]])

model_test6D = BayesianNetwork([d0, d1, d2, d3, d4, d5], [(d1, d2)])
samples_test6D = model_test6D.sample(1000)

model_test6D.fit(samples_test6D)
frewenta commented 5 months ago

Resolved by ensuring the probabilities passed to ConditionalCategorical are of type float32 (float64 caused issues)