Added random_seed parameter to make LsiModel reproducible

[parashardhapola]

The results from multiple runs of LsiModel on same input data produces different results.
This is primarily due to the random gaussian matrix creation step in stochastic_svd function.
Hence, it is hard to setup a reproducible pipeline without resorting to changing global randomness controlling methods.

What was done in this PR :

• A local random state (numpy.random.RandomState) was created in the stochastic_svd function. This random state can be seeded by a user provided parameter (random_seed). The gaussian matrices are drawn using this random state.
• The random_seed parameter was exposed in calls to stochastic_svd from LsiModel.add_documents and Projection.__init__
• 'random_seed' parameter was added to 'init' of both LsiModel and Projection classes.
• Calls to Projection from LsiModel methods (__init__ and add_documents) now expose the random_seed parameter
• The default value of random_seed in Projection, LsiModel and stochastic_svd was set to None. This means the current behavior is persevered by default.

Overall, the above changes allows relaying a 'random_seed' from LsiModel class to downstream calls. The changes suggested in this PR are highly localized and are not expected to have any global impact.

from gensim.test.utils import common_dictionary, common_corpus
from gensim.models import LsiModel
import numpy as np

model1 = LsiModel(common_corpus, id2word=common_dictionary)
model2 = LsiModel(common_corpus, id2word=common_dictionary)

print (np.all(model1.get_topics() - model2.get_topics() < 1e-10))
# Prints 'False' for both the current version and this PR.

With random_seed parameter introduced in this PR:

model1 = LsiModel(common_corpus, id2word=common_dictionary, random_seed=1)
model2 = LsiModel(common_corpus, id2word=common_dictionary, random_seed=1)

print (np.all(model1.get_topics() - model2.get_topics() < 1e-10))
# Prints 'True'

[piskvorky]

Good feature and nicely crafted PR.

I left some code style nitpicks in the comments. I realize some of the existing code may not follow this code style (LSI is the oldest part of Gensim) but please try to follow / update the code style where you can. Thanks!

[piskvorky]

Looks good to me. Let's wait for @mpenkov 's blessing & merge.

[parashardhapola]

@mpenkov and @piskvorky : A gentle reminder

[piskvorky]

Approved by me; but @mpenkov is busy, thanks for your patience.

[mpenkov]

Suggested change

	extra_dims=self.extra_samples, dtype=dtype, random_seed=self.random_seed
	extra_dims=self.extra_samples, dtype=dtype, random_seed=self.random_seed,

[piskvorky]

@mpenkov was good merge for 4.1.0, seems we have missed this PR: https://github.com/RaRe-Technologies/gensim/milestone/5

Do we merge now or anything missing?

[mpenkov]

Yeah, looks like we missed this PR. I'll merge it in now. Do we want to do 4.1.3 (or 4.2.0) or can that wait till later?

[piskvorky]

I think the version decision can wait – 4.1.3devX is fine for now. Thanks!

[mpenkov]

Finally merged. Thank you @parashardhapola !!