eseraygun / python-alignment Goto Github PK

Since backtraceFrom is implemented by recursion (instead of iteration), calling the aligner on "long" sequences (more than 1000 items) results in a RecursionError with Python defaults. Extending stack depth limit may cause other serious issues.

I am not sure whether there even exists a metric for this: if there is a segment where both sequences deviate a lot in their length, then the current algorithm will do as many substitutions to the left as possible and fill up the rest by insertions/deletions. Wouldn't it be fairer if it tried to distribute those more evenly across the segment's symbol mass?

Any thoughts or future plans on incorporating multiple sequence alignment algorithms?

P.S. really found your library useful 😄

Does anyone know whether there is any documentation available on the full functionalities? Maybe someone has something local that he/she is willing to share? So far, I only found the example code snippets that help you to get started.

Especially interested in a short explanation on the differences between:

• GlobalSequenceAlignment
• StrictGlobalSequenceAlignment
• LocalSequenceAlignment

Hi,

Thank you for providing the module.
I was wondering whether you could provide information on the algorithms used?
e.g. Smith Watermann?
(While your code is well structured I am getting a bit lost with some of the variables)

I am considering using / modifying it, but have currently the following blockers:

• It doesn't give me back the indices or the original sequence elements (because I am meant to encode them)
• Recursion is in general good but in this case it leads to a too deep stack level with long sequences

Because of using such constructions as print 'v', v this library can not be used with python 3.x.

When globally aligning sequences that deviate much, combinatory explosion can quickly leed to excessive runtime memory consumption in the current implementation. And it is not always easy to detect those cases by score heuristics in a prior backtrace=False pass.

I believe these should be added:

1. a package-exposed variable with a default limit (perhaps relative to the sequences' length)
2. an optional parameter with an override limit to be able to control the quality-performance trade-off.

(The limit could be based on stack depth or number of alternatives, for example.)

Example: I am trying to align OCRed images of German Fraktur script with their corresponding ground truth text. Sometimes the OCR fails miserably like so:
Mitreden andrer 274. Günſtiger Eindruck der Staatsrathsſitzungen 274. (original line)
*0obe-ondrer '? '-änſiger Eindrue der Torerotheflgg,, (OCR result)
In this case, using StrictGlobalSequenceAligner tries to take more than 20 GB RSS (at which point I quit).