Install deps from requirements.txt even in pip installation

Hi @ajtulloch

Thanks for your nice work! I have a question about complexity in your implementation.

In your implementation computing pca, you use

Q = mult(A, np.random.uniform(low=-1.0, high=1.0, size=(n, l)))

in line 1534 in fbpca.py, which leads to the computational complexity of calculating pca as O(mnk) if A is in shape m * n and we are computing first k singular values. However in their paper arXiv: 0909.4061 and another paper by same authors Liberty, Edo, et al., PNAS 104.51 (2007): 20167-20172., they improved their algorithm to O(k^2(m+n) + mn\log(k)) by bypassing the multiplication directly instead. May I ask if my understanding of your implementation is correct? And should we work on improving it further to see how it works?

Hello authors,

In the benchmarks you mention running SVD on 1e5 x 1e5 dense matrix in 120 seconds.
Can you specify the machines and parameters used for these benchmarks?
Also, a little guidance on runtime w.r.t. k (number of components).

Thanks for sharing this!

Was this work modeled after this article https://research.fb.com/fast-randomized-svd/ ? which mentions Spark, it doesn't look this code can run in Spark / can run distributely?

9.4 * 10⁷ × 4 * 10³ matrix with 1.6 * 10⁹ non-zeros, 60 seconds on single machine server vs 50 seconds on a 68 machine cluster with Spark.