jjjjohnson
commented
5 years ago Could you run it another time? I didnot encounter the error after the change
Could be den =0 in self.mu[ix, :] = num / den
Closed WuZhuoran closed 5 years ago
jjjjohnson
commented
5 years ago Could you run it another time? I didnot encounter the error after the change
Could be den =0 in self.mu[ix, :] = num / den
WuZhuoran
commented
5 years ago Now I get another error:
Traceback (most recent call last):
File "/Users/zhuoran/Documents/git/numpy-ml/gmm/tests.py", line 110, in <module>
plot()
File "/Users/zhuoran/Documents/git/numpy-ml/gmm/tests.py", line 89, in plot
ret = _G.fit(max_iter=75, verbose=False)
File "/Users/zhuoran/Documents/git/numpy-ml/gmm/gmm.py", line 59, in fit
self._E_step()
File "/Users/zhuoran/Documents/git/numpy-ml/gmm/gmm.py", line 102, in _E_step
assert_allclose(np.sum(q_i), 1, err_msg="{}".format(np.sum(q_i)))
File "/usr/local/lib/python3.7/site-packages/numpy/testing/nose_tools/utils.py", line 1398, in assert_allclose
verbose=verbose, header=header, equal_nan=equal_nan)
File "/usr/local/lib/python3.7/site-packages/numpy/testing/nose_tools/utils.py", line 781, in assert_array_compare
raise AssertionError(msg)
AssertionError:
Not equal to tolerance rtol=1e-07, atol=0
2.0
(mismatch 100.0%)
x: array(2.)
y: array(1)
And then I ran again, it passed all test.
It seems that the test is not stable.
ddbourgin
commented
5 years ago @jjjjohnson - your suggestion in #16 is correct; I must have forgot to change the np.zeros(2, 2) after testing the model for the 2D case!
Anyway, it's pretty unreasonable for me to consider plots as any sort of test at all. I renamed the test.py file in the above commits to plot.py.
RE: the instability @WuZhuoran observes - this was occurring because we never seeded the rng for in the GMM, so different runs had different inits, and thus different results. I've fixed this in the above commits by adding a seed parameter to the GMM class and using it within plot.py to ensure reproducibility across runs.
I think the assertion errors you saw earlier were due to instances in which a mixture component collapsed and wasn't caught properly. I've added a few small checks in the code to guard against this - I think that should address it, though please check and let me know.
Finally, as I went over the code, I noticed that X was being passed during init rather than as a parameter to the fit method. I've updated the code to address this.
- What bug I fixed
According to @jjjjohnson in #16 , we can apply multi dimension covariance in gmm.
This pull request fixes #16.
- How I fixed it
changedimension from2toself.d.- How you can verify it
The tests did not pass because of the following: And tests did pass before changing. We can ask @jjjjohnson to take a look.