add pixel shuffle

[MARD1NO]

增加 Pixelshuffle 算子

PixelShufflev1是宽高都是相同的缩放因子（主要是为了对齐其他框架）
效果如下

import oneflow as flow
import oneflow.typing as tp
import numpy as np 


@flow.global_function()
def PixelShufflev2Job(input: tp.Numpy.Placeholder(shape=(1, 8, 2, 2), dtype=flow.float32))->tp.Numpy:
    out = flow.nn.PixelShufflev2(input, h_upscale_factor=2, w_upscale_factor=2)

    return out

input = np.random.uniform(size=(1, 8, 2, 2)).astype(np.float32)
out = PixelShuffleJob(input)

# out.shape (1, 2, 4, 4)

PixelShuffle2允许用户对宽高设定不同的缩放因子

import oneflow as flow
import oneflow.typing as tp
import numpy as np 


@flow.global_function()
def PixelShufflev2Job(input: tp.Numpy.Placeholder(shape=(1, 8, 2, 4), dtype=flow.float32))->tp.Numpy:
    out = flow.nn.PixelShufflev2(input, h_upscale_factor=2, w_upscale_factor=4)

    return out

input = np.random.uniform(size=(1, 8, 2, 4)).astype(np.float32)
out = PixelShuffleJob(input)

# out.shape (1, 1, 4, 16)

附上循环实现代码，方便后续重构为C++算子

import numpy as np
batchsize = 2
channel = 6
height = 2
width = 2
h_scale = 1
w_scale = 3
#
#
# input = np.ones(shape=(batchsize, 1, height, width))
input = np.random.random(size=(batchsize, channel, height, width)).astype(np.float32)
# for i in range(1, channel):
#     new_input = np.ones(shape=(batchsize, 1, height, width)) * (i + 1)
#     input = np.concatenate([input, new_input], axis=1)
out_h = height * h_scale
out_w = width * w_scale
out_c = channel // (h_scale * w_scale)
fp_result = np.zeros((batchsize, out_c, out_h, out_w)).astype(np.float32)
#
#
def getchannel(c_idx, h_idx, w_idx, h_scale, w_scale):
    inner_h_idx = h_idx - int(h_idx / h_scale) * h_scale
    inner_w_idx = w_idx - int(w_idx / w_scale) * w_scale
    # Compute the index, Because the index is left->right, up->down,
    # so the index of height should multiply `scale`
    print(h_idx)
    print("inner h idx: ", inner_h_idx)
    print("inner w idx: ", inner_w_idx)
    inner_1d_idx = inner_h_idx * w_scale + inner_w_idx
    # print(inner_1d_idx)
    return (c_idx * h_scale * w_scale + inner_1d_idx) % channel
#
#
# # ======forward======
for b in range(batchsize):
    for c in range(out_c):
        for h in range(out_h):
            for w in range(out_w):
                in_c = getchannel(c, h, w, h_scale, w_scale)
                in_h = int(h / h_scale)
                in_w = int(w / w_scale)
                fp_result[b, c, h, w] = input[b, in_c, in_h, in_w]
print(input)
print(fp_result)