Thinker BoardのOpenCL（２）

 前の投稿で、ビルドして動かすところまで、やってみましたが、今回はOpenCLを使うと、あまりに遅かったので、パラメータを調整してみます。

どうも、max_bin = 255となっていたので、64とかに減らして実行すると15がいいよというWarningがでました。

[LightGBM] [Info] Finished loading parameters

[LightGBM] [Info] Using column number 0 as label

[LightGBM] [Info] Loading weights...

[LightGBM] [Info] Construct bin mappers from text data time 0.03 seconds

[LightGBM] [Info] Loading weights...

[LightGBM] [Info] Finished loading data in 0.207653 seconds

[LightGBM] [Info] Number of positive: 3716, number of negative: 3284

[LightGBM] [Info] This is the GPU trainer!!

[LightGBM] [Info] Total Bins 0

[LightGBM] [Info] Number of data points in the train set: 7000, number of used features: 28

[LightGBM] [Warning] Setting max_bin to 15 is suggested for best performance

・・・

・・・

[LightGBM] [Info] 27.172875 seconds elapsed, finished iteration 99

[LightGBM] [Info] Iteration:100, training binary_logloss : 0.267676

[LightGBM] [Info] Iteration:100, training auc : 0.991391

[LightGBM] [Info] Iteration:100, valid_1 binary_logloss : 0.5031

[LightGBM] [Info] Iteration:100, valid_1 auc : 0.834687

[LightGBM] [Info] 27.380465 seconds elapsed, finished iteration 100

[LightGBM] [Info] Finished training

それで、15に設定すると。確かに早くなりましたが、それでもCPUより圧倒的に遅いです。

[LightGBM] [Info] Finished loading parameters
[LightGBM] [Info] Using column number 0 as label
[LightGBM] [Info] Loading weights...
[LightGBM] [Info] Construct bin mappers from text data time 0.03 seconds
[LightGBM] [Info] Loading weights...
[LightGBM] [Info] Finished loading data in 0.213896 seconds
[LightGBM] [Info] Number of positive: 3716, number of negative: 3284
[LightGBM] [Info] This is the GPU trainer!!
[LightGBM] [Info] Total Bins 0
[LightGBM] [Info] Number of data points in the train set: 7000, number of used features: 28
[LightGBM] [Info] Using GPU Device: Mali-T760, Vendor: ARM
[LightGBM] [Info] Compiling OpenCL Kernel with 16 bins...
[LightGBM] [Info] GPU programs have been built
[LightGBM] [Info] Size of histogram bin entry: 8
[LightGBM] [Info] 27 dense feature groups (0.11 MB) transferred to GPU in 0.011649 secs. 1 sparse feature groups
[LightGBM] [Info] Finished initializing training

・・・

・・・

[LightGBM] [Info] 17.915968 seconds elapsed, finished iteration 99

[LightGBM] [Info] Iteration:100, training binary_logloss : 0.346398

[LightGBM] [Info] Iteration:100, training auc : 0.96812

[LightGBM] [Info] Iteration:100, valid_1 binary_logloss : 0.566812

[LightGBM] [Info] Iteration:100, valid_1 auc : 0.773077

[LightGBM] [Info] 18.062390 seconds elapsed, finished iteration 100

[LightGBM] [Info] Finished training

同じパラメータでCPUで実行すると、CPUでもさらに早くなりました。２倍ぐらいでしょうか。

・・・

[LightGBM] [Info] Iteration:100, training binary_logloss : 0.268637

[LightGBM] [Info] Iteration:100, training auc : 0.991395

[LightGBM] [Info] Iteration:100, valid_1 binary_logloss : 0.518213

[LightGBM] [Info] Iteration:100, valid_1 auc : 0.816853

[LightGBM] [Info] 0.491855 seconds elapsed, finished iteration 100

[LightGBM] [Info] Finished training

他のパラメータもチューニングしていく必要がありそうです。ただ、GPUがそもそも遅いように思います。