[python-package] do not copy column-major numpy arrays when predicting

[jmoralez]

Follow up to #6721. This prevents copying the data array when predicting if it's column-major.

[StrikerRUS]

LGTM! Thanks for continuing working on this!

[StrikerRUS]

There is one more place where _np2d_to_np1d() can simplify the code and save memory:

LightGBM/python-package/lightgbm/basic.py

Lines 2358 to 2361 in 53be01f

	if mat.dtype == np.float32 or mat.dtype == np.float64:
	mats[i] = np.asarray(mat.reshape(mat.size), dtype=mat.dtype)
	else: # change non-float data to float data, need to copy
	mats[i] = np.array(mat.reshape(mat.size), dtype=np.float32)

But it'll require to change from single flag int is_row_major to array of such flags int* is_row_major in LGBM_DatasetCreateFromMats

Could you please prepare one more PR for that?

[jameslamb]

Thanks!

Merging this as-is because @StrikerRUS 's suggestion in #6751 (comment) is for a follow-up PR.

[jmoralez]

The create from mats function takes a single int argument for the layout

LightGBM/include/LightGBM/c_api.h

Lines 431 to 436 in 480600b

	LIGHTGBM_C_EXPORT int LGBM_DatasetCreateFromMats(int32_t nmat,
	const void** data,
	int data_type,
	int32_t* nrow,
	int32_t ncol,
	int is_row_major,

so they should all have the same layout.

I can work on that, i.e. try not to copy if they all have the same layout and enforce row-major otherwise.

[StrikerRUS]

@jmoralez

The create from mats function takes a single int argument for the layout ...

Yeah, I saw this before writing my suggestion, and that's why I wrote:

But it'll require to change from single flag int is_row_major to array of such flags int* is_row_major in LGBM_DatasetCreateFromMats

I think that a small breaking change in public API worth it in exchange of saving memory.
This will allow to pass mats with different layouts.

[jmoralez]

I see, I had misunderstood the suggestion. Sounds good, I'll work on that.

[StrikerRUS]

@jmoralez Thanks a lot!