Add nthread arg to XGBoostRegressor , XGBoostCount and XGBoostClassifier

[dmahasen]

Hi, I have added nthread arg to XGBoostRegressor , XGBoostCount and XGBoostClassifier. So, number of threads used can be limited.

[ablaom]

@dmahasen Thanks for this contribution!

Looks good to me and will merge now.

However, I wasn't actually able to notice a speed-up with the number of threads on my machine. What have your observations been?

julia> mach.model.nthread
12

julia> @time fit!(mach, force=true, verbosity=-1)
  3.513411 seconds (104.47 k allocations: 6.427 MiB, 0.27% gc time)
Machine{XGBoostClassifier,…} @287 trained 7 times; caches data
  args: 
    1:  Source @985 ⏎ `Table{AbstractVector{Continuous}}`
    2:  Source @642 ⏎ `AbstractVector{Multiclass{3}}`


julia> mach.model.nthread = 1
1

julia> @time fit!(mach, force=true, verbosity=-1)
  3.522144 seconds (104.47 k allocations: 6.427 MiB)
Machine{XGBoostClassifier,…} @287 trained 8 times; caches data
  args: 
    1:  Source @985 ⏎ `Table{AbstractVector{Continuous}}`
    2:  Source @642 ⏎ `AbstractVector{Multiclass{3}}`

[dmahasen]

@ablaom, It is my pleasure to make this tiny contribution.

I have experienced a time difference when I change the number of threads.

Here, X has 1000 observations with 100 values for each.

julia> xgb_thread_2 = xgbReg(num_round=1000,nthread= 2);

julia> m_xgb_thread_2 = machine(xgb_thread_2, X, y);

julia> @time fit!(m_xgb_thread_2,verbosity=-1 );
2.038276 seconds (39.83 k allocations: 4.895 MiB)

julia> xgb_thread_6 = xgbReg(num_round=1000,nthread= 6);

julia> m_xgb_thread_6 = machine(xgb_thread_6, X, y);

julia> @time fit!(m_xgb_thread_6,verbosity=-1 );
0.881676 seconds (39.83 k allocations: 4.895 MiB)

The main requirement I had was to limit the number of cores in fitting the model. Otherwise, it consumes all the cores on my machine.

[ablaom]

Ah, yes of course! In tree methods the parallelism occurs across features not observations! I had a lot of observations but few features in my test.

Thanks again!