How to use oneflow to achieve similar functions to torch.nn.PixelShuffle()

[416047850]

How to use oneflow to achieve similar functions to torch.nn.PixelShuffle()

[doombeaker]

Thanks for your feedback. This Op is not yet available, but will be added to the development plan. Anyone else who are intrested in helping us to implement this op, can reply here and we will give a hand as much as possible.

[Ldpe2G]

@416047850 I have implemented the PixelShuffle in oneflow which can get the same result as Pytorch's pixel_shuffle.

You can have a try :)

reference: apache/mxnet#13571

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

import torch

def PixelShuffle(input, h_factor, w_factor):
    b, c, h, w = input.shape

    assert c % (h_factor * w_factor) == 0

    new_c = int(c / (h_factor * w_factor))

    out = flow.reshape(input, [b, new_c, h_factor * w_factor, h, w])
    out = flow.reshape(out, [b * new_c, h_factor, w_factor, h, w])
    out = flow.transpose(out, [0, 3, 1, 4, 2])
    out = flow.reshape(out, [b, new_c, h * h_factor, w * w_factor])

    return out

input_shape = (3, 4, 2, 2)
h_factor = 2
w_factor = 2

func_config = flow.FunctionConfig()
func_config.default_data_type(flow.float)
@flow.global_function(type="predict", function_config=func_config)
def PixelShuffleJob(input: oft.Numpy.Placeholder(input_shape, dtype=flow.float32)):
    with flow.scope.placement("gpu", "0:0"):

        out = PixelShuffle(input, h_factor, w_factor)

        return out

# OneFlow
check_point = flow.train.CheckPoint()
check_point.init()

input = np.random.uniform(size=input_shape).astype(np.float32)
of_out = PixelShuffleJob(input).get()

# print(input)
# print(of_out.numpy())

arr = torch.tensor(input)
t_out = torch.pixel_shuffle(arr, upscale_factor=2)

# print(t_out.numpy())

print(np.allclose(of_out.numpy(), t_out.numpy()))

[416047850]

@416047850 I have implemented the PixelShuffle in oneflow which can get the same result as Pytorch's pixel_shuffle.

You can have a try :)

reference: apache/incubator-mxnet#13571

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

import torch

def PixelShuffle(input, h_factor, w_factor):
    b, c, h, w = input.shape

    assert c % (h_factor * w_factor) == 0

    new_c = int(c / (h_factor * w_factor))

    out = flow.reshape(input, [b, new_c, h_factor * w_factor, h, w])
    out = flow.reshape(out, [b * new_c, h_factor, w_factor, h, w])
    out = flow.transpose(out, [0, 3, 1, 4, 2])
    out = flow.reshape(out, [b, new_c, h * h_factor, w * w_factor])

    return out

input_shape = (3, 4, 2, 2)
h_factor = 2
w_factor = 2

func_config = flow.FunctionConfig()
func_config.default_data_type(flow.float)
@flow.global_function(type="predict", function_config=func_config)
def PixelShuffleJob(input: oft.Numpy.Placeholder(input_shape, dtype=flow.float32)):
    with flow.scope.placement("gpu", "0:0"):

        out = PixelShuffle(input, h_factor, w_factor)

        return out

# OneFlow
check_point = flow.train.CheckPoint()
check_point.init()

input = np.random.uniform(size=input_shape).astype(np.float32)
of_out = PixelShuffleJob(input).get()

# print(input)
# print(of_out.numpy())

arr = torch.tensor(input)
t_out = torch.pixel_shuffle(arr, upscale_factor=2)

# print(t_out.numpy())

print(np.allclose(of_out.numpy(), t_out.numpy()))

Thanks. Youre my God! : )