Fix major bug in v0.3.1: negative counts

[sjfleming]

Closes #306
Closes #348

Something about the update from v0.3.0 to v0.3.1 induced a bug (I'm not yet sure of the scope, but potentially affecting all datasets) where negative counts could be produced in the output. It is assumed that these outputs are totally compromised. v0.3.1 has already been redacted, but the master branch is still compromised.

I think the bug probably lies in the updated posterior generation code, which was the target of memory-saving updates in v0.3.1.

Expand tests covering that code.

• unit tests for posterior generation
• integration tests ensuring no negative count matrix entries in the output (not explicitly tested currently)

[sjfleming]

After running simulated data through v0.3.0 and the master branch, I see that the posterior generated in each case is identical.

So despite what I initially thought, it seems like the bug must have been introduced in estimation.py, which is quite well-covered with tests...

(It is not a CPU / GPU issue, as a run without --cuda basically agrees with a GPU run.)

[sjfleming]

It came down to a small typo in an integer datatype. I had put np.uint8 in estimation.py line 241 here, but np.uint8 is not large enough (it needs to accommodate genes). This was probably causing integer overflows that totally messed up the output count matrix.

Changing to np.uint32 seems to fix the problem.

Before

And there are negative count matrix entries in the output.

After

And there are no longer negative count matrix entries in the output.

[mbabadi]

Nice job expanding the tests! Just a small comment to address. Also, while you're at it, you may want to search the rest of the codebase for uint8 and other explicit dtypes and refactor accordingly.

[mbabadi]

Either remove the print statement, or convert it to something more informative-- e.g. old matrix shape = ..., new matrix shape = ...

[sjfleming]

Alright @mbabadi , what do you think now? I added a warning to the README, and I added a command-line argument to enable people to re-run their old checkpointed v0.3.1 runs to produce a valid output.

[mbabadi]

Looks good!