Full Code of JuewenPeng/BokehMe for AI

Repository: JuewenPeng/BokehMe
Branch: main
Commit: 8b3ed556dc14
Files: 9
Total size: 10.6 MB

Directory structure:
gitextract_mdifo9q2/

├── LICENSE
├── README.md
├── checkpoints/
│   ├── arnet.pth
│   └── iunet.pth
├── classical_renderer/
│   ├── scatter.py
│   └── scatter_ex.py
├── demo.py
├── neural_renderer.py
└── requirements.txt

================================================
FILE CONTENTS
================================================

================================================
FILE: LICENSE
================================================
                                 Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright [yyyy] [name of copyright owner]

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.


================================================
FILE: README.md
================================================
# BokehMe: When Neural Rendering Meets Classical Rendering (CVPR 2022 Oral)

[Juewen Peng](https://scholar.google.com/citations?hl=en&user=fYC6lCUAAAAJ)<sup>1</sup>,
[Zhiguo Cao](http://english.aia.hust.edu.cn/info/1085/1528.htm)<sup>1</sup>,
[Xianrui Luo](https://scholar.google.com/citations?hl=en&user=tUeWQ5AAAAAJ)<sup>1</sup>,
[Hao Lu](http://faculty.hust.edu.cn/LUHAO/en/index.htm)<sup>1</sup>,
[Ke Xian](https://sites.google.com/site/kexian1991/)<sup>1*</sup>,
[Jianming Zhang](https://jimmie33.github.io/)<sup>2</sup>

<sup>1</sup>Huazhong University of Science and Technology, <sup>2</sup>Adobe Research

<p align="center">
<img src=https://user-images.githubusercontent.com/38718148/171405815-b3cc8799-27cd-457e-89df-686695187554.jpg />
</p>

### [Project](https://juewenpeng.github.io/BokehMe/) | [Paper](https://github.com/JuewenPeng/BokehMe/blob/main/pdf/BokehMe.pdf) | [Supp](https://github.com/JuewenPeng/BokehMe/blob/main/pdf/BokehMe-supp.pdf) | [Poster](https://github.com/JuewenPeng/BokehMe/blob/main/pdf/BokehMe-poster.pdf) | [Video](https://www.youtube.com/watch?v=e-zr_wCxNc8) | [Data](#blb-dataset)

This repository is the official PyTorch implementation of the CVPR 2022 paper "BokehMe: When Neural Rendering Meets Classical Rendering".


**NOTE**: There is a citation mistake in the paper of the conference version. In section 4.1, the disparity maps of the EBB400 dataset are predicted by MiDaS [1] instead of DPT [2]. <!-- We have corrected it in the arXiv version. We apologize for this oversight and for any confusion that it may have caused.  --><br/>
> [1] Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer <br/>
> [2] Vision Transformers for Dense Prediction



## Installation
```
git clone https://github.com/JuewenPeng/BokehMe.git
cd BokehMe
pip install -r requirements.txt
```


## Usage
```
python demo.py --image_path 'inputs/21.jpg' --disp_path 'inputs/21.png' --save_dir 'outputs' --K 60 --disp_focus 90/255 --gamma 4 --highlight
```
- `image_path`:  path of the input all-in-focus image
- `disp_path`: path of the input disparity map (predicted by [DPT](https://github.com/isl-org/DPT) in this example)
- `save_dir`: directory to save the results
- `K`: blur parameter
- `disp_focus`: refocused disparity (range from 0 to 1)
- `gamma`: gamma value (range from 1 to 5)
- `highlight`: enhance RGB values of highlights before rendering for stunning bokeh balls

See `demo.py` for more details.




## BLB Dataset
The BLB dataset is synthesized by Blender 2.93. It contains 10 scenes, each consisting of an all-in-focus image, a disparity map, a stack of bokeh images with 5 blur amounts and 10 refocused disparities, and a parameter file. We additionally provide 15 corrupted disparity maps (through gaussian blur, dilation, erosion) for each scene. Our BLB dataset can be downloaded from [Google Drive](https://drive.google.com/drive/folders/1URpab6AXQsNTqcBcighF73w5pFlvM0Ej?usp=sharing) or [Baidu Netdisk](https://pan.baidu.com/s/1U0XlFM_84-vVgnXGYz0ncQ?pwd=re8q).

**Instructions**: 
- EXR images can be loaded by `image = cv2.imread(IMAGE_PATH, -1)[..., :3].astype(np.float32) ** (1/2.2)` . The loaded images are in BGR, so you can convert them to RGB by `image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)` if necessary.
- EXR depth maps can be loaded by `depth = cv2.imread(DEPTH_PATH, -1)[..., 0].astype(np.float32)`. You can convert them to disparity maps by `disp = 1 / depth`. Note that it is **unnecesary** to normalize the disparity maps since we have pre-processed them to ensure that the signed defocus maps calculated by `K * (disp - disp_focus)` are in line with the experimental settings of the paper.
- NOTE: Some pixel values of images may be larger than 1 for highlights (but mostly smaller than 1). Considering the fact that some rendering methods can only output values between 0 and 1, we clip the numerical ranges of the predicted bokeh images and the real ones to [0, 1] before evaluation. The main reason for this phenomenon (image values exceeding 1) is that the EXR images exported from Blender are in linear space, and we only process them with gamma 2.2 correction without tone mapping. We will improve it in the future.

## Citation
If you find our work useful in your research, please cite our paper.

```
@inproceedings{Peng2022BokehMe,
  title = {BokehMe: When Neural Rendering Meets Classical Rendering},
  author = {Peng, Juewen and Cao, Zhiguo and Luo, Xianrui and Lu, Hao and Xian, Ke and Zhang, Jianming},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2022}
}
```


================================================
FILE: checkpoints/arnet.pth
================================================
[File too large to display: 10.6 MB]

================================================
FILE: classical_renderer/scatter.py
================================================
#!/user/bin/env python3
# -*- coding: utf-8 -*-

import torch
import torch.nn as nn
import torch.nn.functional as F
import cupy
import re

kernel_Render_updateOutput = '''

    extern "C" __global__ void kernel_Render_updateOutput(
        const int n,
        const float* image,          // original image
        const float* defocus,        // signed defocus map
        int* defocusDilate,          // signed defocus map after dilating
        float* bokehCum,             // cumulative bokeh image
        float* weightCum             // cumulative weight map
    )
    {
        for (int intIndex = (blockIdx.x * blockDim.x) + threadIdx.x; intIndex < n; intIndex += blockDim.x * gridDim.x) {
            const int intN = ( intIndex / SIZE_3(weightCum) / SIZE_2(weightCum) / SIZE_1(weightCum) ) % SIZE_0(weightCum);
            // const int intC = ( intIndex / SIZE_3(weightCum) / SIZE_2(weightCum)                     ) % SIZE_1(weightCum);
            const int intY = ( intIndex / SIZE_3(weightCum)                                         ) % SIZE_2(weightCum);
            const int intX = ( intIndex                                                             ) % SIZE_3(weightCum);

            float fltDefocus = VALUE_4(defocus, intN, 0, intY, intX);
            float fltRadius = fabsf(fltDefocus);

            for (int intDeltaY = -(int)(fltRadius)-1; intDeltaY <= (int)(fltRadius)+1; ++intDeltaY) {
                for (int intDeltaX = -(int)(fltRadius)-1; intDeltaX <= (int)(fltRadius)+1; ++intDeltaX) {

                    int intNeighborY = intY + intDeltaY;
                    int intNeighborX = intX + intDeltaX;

                    if ((intNeighborY >= 0) && (intNeighborY < SIZE_2(bokehCum)) && (intNeighborX >= 0) && (intNeighborX < SIZE_3(bokehCum))) {
                        float fltDist = sqrtf((float)(intDeltaY)*(float)(intDeltaY) + (float)(intDeltaX)*(float)(intDeltaX));
                        float fltWeight = (0.5 + 0.5 * tanhf(4 * (fltRadius - fltDist))) / (fltRadius * fltRadius + 0.2);
                        if (fltRadius >= fltDist) {
                            atomicMax(&defocusDilate[OFFSET_4(defocusDilate, intN, 0, intNeighborY, intNeighborX)], int(fltDefocus));
                        }
                        atomicAdd(&weightCum[OFFSET_4(weightCum, intN, 0, intNeighborY, intNeighborX)], fltWeight);
                        atomicAdd(&bokehCum[OFFSET_4(bokehCum, intN, 0, intNeighborY, intNeighborX)], fltWeight * VALUE_4(image, intN, 0, intY, intX));
                        atomicAdd(&bokehCum[OFFSET_4(bokehCum, intN, 1, intNeighborY, intNeighborX)], fltWeight * VALUE_4(image, intN, 1, intY, intX));
                        atomicAdd(&bokehCum[OFFSET_4(bokehCum, intN, 2, intNeighborY, intNeighborX)], fltWeight * VALUE_4(image, intN, 2, intY, intX));
                    }
                }
            }
        }
    }

'''


def cupy_kernel(strFunction, objVariables):
    strKernel = globals()[strFunction]

    while True:
        objMatch = re.search('(SIZE_)([0-4])(\()([^\)]*)(\))', strKernel)

        if objMatch is None:
            break
        # end

        intArg = int(objMatch.group(2))

        strTensor = objMatch.group(4)
        intSizes = objVariables[strTensor].size()

        strKernel = strKernel.replace(objMatch.group(), str(intSizes[intArg]))
    # end

    while True:
        objMatch = re.search('(OFFSET_)([0-4])(\()([^\)]+)(\))', strKernel)

        if objMatch is None:
            break
        # end

        intArgs = int(objMatch.group(2))
        strArgs = objMatch.group(4).split(',')

        strTensor = strArgs[0]
        intStrides = objVariables[strTensor].stride()
        strIndex = ['((' + strArgs[intArg + 1].replace('{', '(').replace('}', ')').strip() + ')*' + str(
            intStrides[intArg]) + ')' for intArg in range(intArgs)]

        strKernel = strKernel.replace(objMatch.group(0), '(' + str.join('+', strIndex) + ')')
    # end

    while True:
        objMatch = re.search('(VALUE_)([0-4])(\()([^\)]+)(\))', strKernel)

        if objMatch is None:
            break
        # end

        intArgs = int(objMatch.group(2))
        strArgs = objMatch.group(4).split(',')

        strTensor = strArgs[0]
        intStrides = objVariables[strTensor].stride()
        strIndex = ['((' + strArgs[intArg + 1].replace('{', '(').replace('}', ')').strip() + ')*' + str(
            intStrides[intArg]) + ')' for intArg in range(intArgs)]

        strKernel = strKernel.replace(objMatch.group(0), strTensor + '[' + str.join('+', strIndex) + ']')
    # end

    return strKernel
# end


# @cupy.util.memoize(for_each_device=True)
@cupy.memoize(for_each_device=True)
def cupy_launch(strFunction, strKernel):
    return cupy.cuda.compile_with_cache(strKernel).get_function(strFunction)
# end


class _FunctionRender(torch.autograd.Function):
    @staticmethod
    def forward(self, image, defocus):
        # self.save_for_backward(image, defocus)

        defocus_dilate = defocus.int()
        bokeh_cum = torch.zeros_like(image)
        weight_cum = torch.zeros_like(defocus)

        if defocus.is_cuda == True:
            n = weight_cum.nelement()
            cupy_launch('kernel_Render_updateOutput', cupy_kernel('kernel_Render_updateOutput', {
                'image': image,
                'defocus': defocus,
                'defocusDilate': defocus_dilate,
                'bokehCum': bokeh_cum,
                'weightCum': weight_cum
            }))(
                grid=tuple([int((n + 512 - 1) / 512), 1, 1]),
                block=tuple([512, 1, 1]),
                args=[
                    cupy.int(n),
                    image.data_ptr(),
                    defocus.data_ptr(),
                    defocus_dilate.data_ptr(),
                    bokeh_cum.data_ptr(),
                    weight_cum.data_ptr()
                ]
            )

        elif defocus.is_cuda == False:
            raise NotImplementedError()

        # end

        return defocus_dilate.float(), bokeh_cum, weight_cum
    # end

    # @staticmethod
    # def backward(self, gradBokehCum, gradWeightCum):
    # end

# end


def FunctionRender(image, defocus):
    defocus_dilate, bokeh_cum, weight_cum = _FunctionRender.apply(image, defocus)

    return defocus_dilate, bokeh_cum, weight_cum
# end


class ModuleRenderScatter(torch.nn.Module):
    def __init__(self):
        super(ModuleRenderScatter, self).__init__()
    # end

    def forward(self, image, defocus):
        defocus_dilate, bokeh_cum, weight_cum = FunctionRender(image, defocus)
        bokeh = bokeh_cum / weight_cum
        return bokeh, defocus_dilate
    # end
# end


================================================
FILE: classical_renderer/scatter_ex.py
================================================
#!/user/bin/env python3
# -*- coding: utf-8 -*-

import torch
import cupy
import re

kernel_Render_updateOutput = '''

    extern "C" __global__ void kernel_Render_updateOutput(
        const int n,
        const int polySides,
        const float initAngle,
        const float* image,          // original image
        const float* defocus,        // signed defocus map
        int* defocusDilate,          // signed defocus map after dilating
        float* bokehCum,             // cumulative bokeh image
        float* weightCum             // cumulative weight map
    )
    {
        // int polySides = 6;
        float PI = 3.1415926536;
        float fltAngle1 = 2 * PI / (float)(polySides);
        float fltAngle2 = PI / 2 - PI / (float)(polySides);
        // float initAngle = PI / 2;
        float donutRatio = 0;  // (0 -> 0.5 : circle -> donut)

        for (int intIndex = (blockIdx.x * blockDim.x) + threadIdx.x; intIndex < n; intIndex += blockDim.x * gridDim.x) {
            const int intN = ( intIndex / SIZE_3(weightCum) / SIZE_2(weightCum) / SIZE_1(weightCum) ) % SIZE_0(weightCum);
            // const int intC = ( intIndex / SIZE_3(weightCum) / SIZE_2(weightCum)                     ) % SIZE_1(weightCum);
            const int intY = ( intIndex / SIZE_3(weightCum)                                         ) % SIZE_2(weightCum);
            const int intX = ( intIndex                                                             ) % SIZE_3(weightCum);

            float fltDefocus = VALUE_4(defocus, intN, 0, intY, intX);
            float fltRadius = fabsf(fltDefocus);
            float fltRadiusSquare = fltRadius * fltRadius;
            // float fltWeight = 1.0 / (fltRadiusSquare + 0.4);

            for (int intDeltaY = -(int)(fltRadius)-1; intDeltaY <= (int)(fltRadius)+1; intDeltaY++) {
                for (int intDeltaX = -(int)(fltRadius)-1; intDeltaX <= (int)(fltRadius)+1; intDeltaX++) {

                    int intNeighborY = intY + intDeltaY;
                    int intNeighborX = intX + intDeltaX;

                    float fltAngle = atan2f((float)(intDeltaY), (float)(intDeltaX));
                    fltAngle = fmodf(fabsf(fltAngle + initAngle), fltAngle1);

                    if ((intNeighborY >= 0) & (intNeighborY < SIZE_2(bokehCum)) & (intNeighborX >= 0) & (intNeighborX < SIZE_3(bokehCum))) {
                        float fltDist = sqrtf((float)(intDeltaY)*(float)(intDeltaY) + (float)(intDeltaX)*(float)(intDeltaX));
                        float fltWeight = (0.5 + 0.5 * tanhf(4 * (fltRadius * sinf(fltAngle2)/sinf(fltAngle+fltAngle2) - fltDist))) * (1 - donutRatio + donutRatio * tanhf(0.2 * (1 + fltDist - fltRadius * sinf(fltAngle2)/sinf(fltAngle+fltAngle2)))) / (fltRadius * fltRadius + 0.2);
                        if (fltRadius >= fltDist) {
                            atomicMax(&defocusDilate[OFFSET_4(defocusDilate, intN, 0, intNeighborY, intNeighborX)], int(fltDefocus));
                        }
                        atomicAdd(&weightCum[OFFSET_4(weightCum, intN, 0, intNeighborY, intNeighborX)], fltWeight);
                        atomicAdd(&bokehCum[OFFSET_4(bokehCum, intN, 0, intNeighborY, intNeighborX)], fltWeight * VALUE_4(image, intN, 0, intY, intX));
                        atomicAdd(&bokehCum[OFFSET_4(bokehCum, intN, 1, intNeighborY, intNeighborX)], fltWeight * VALUE_4(image, intN, 1, intY, intX));
                        atomicAdd(&bokehCum[OFFSET_4(bokehCum, intN, 2, intNeighborY, intNeighborX)], fltWeight * VALUE_4(image, intN, 2, intY, intX));
                    }
                }
            }
        }
    }

'''


def cupy_kernel(strFunction, objVariables):
    strKernel = globals()[strFunction]

    while True:
        objMatch = re.search('(SIZE_)([0-4])(\()([^\)]*)(\))', strKernel)

        if objMatch is None:
            break
        # end

        intArg = int(objMatch.group(2))

        strTensor = objMatch.group(4)
        intSizes = objVariables[strTensor].size()

        strKernel = strKernel.replace(objMatch.group(), str(intSizes[intArg]))
    # end

    while True:
        objMatch = re.search('(OFFSET_)([0-4])(\()([^\)]+)(\))', strKernel)

        if objMatch is None:
            break
        # end

        intArgs = int(objMatch.group(2))
        strArgs = objMatch.group(4).split(',')

        strTensor = strArgs[0]
        intStrides = objVariables[strTensor].stride()
        strIndex = ['((' + strArgs[intArg + 1].replace('{', '(').replace('}', ')').strip() + ')*' + str(
            intStrides[intArg]) + ')' for intArg in range(intArgs)]

        strKernel = strKernel.replace(objMatch.group(0), '(' + str.join('+', strIndex) + ')')
    # end

    while True:
        objMatch = re.search('(VALUE_)([0-4])(\()([^\)]+)(\))', strKernel)

        if objMatch is None:
            break
        # end

        intArgs = int(objMatch.group(2))
        strArgs = objMatch.group(4).split(',')

        strTensor = strArgs[0]
        intStrides = objVariables[strTensor].stride()
        strIndex = ['((' + strArgs[intArg + 1].replace('{', '(').replace('}', ')').strip() + ')*' + str(
            intStrides[intArg]) + ')' for intArg in range(intArgs)]

        strKernel = strKernel.replace(objMatch.group(0), strTensor + '[' + str.join('+', strIndex) + ']')
    # end

    return strKernel


# end

# @cupy.util.memoize(for_each_device=True)
@cupy.memoize(for_each_device=True)
def cupy_launch(strFunction, strKernel):
    return cupy.cuda.compile_with_cache(strKernel).get_function(strFunction)


# end

class _FunctionRender(torch.autograd.Function):
    @staticmethod
    def forward(self, image, defocus, poly_sides, init_angle):
        # self.save_for_backward(image, signedDisp)

        defocus_dilate = defocus.int()
        bokeh_cum = torch.zeros_like(image)
        weight_cum = torch.zeros_like(defocus)

        if defocus.is_cuda == True:
            n = weight_cum.nelement()
            cupy_launch('kernel_Render_updateOutput', cupy_kernel('kernel_Render_updateOutput', {
                'poly_sides': poly_sides,
                'init_angle': init_angle,
                'image': image,
                'defocus': defocus,
                'defocusDilate': defocus_dilate,
                'bokehCum': bokeh_cum,
                'weightCum': weight_cum,
            }))(
                grid=tuple([int((n + 512 - 1) / 512), 1, 1]),
                block=tuple([512, 1, 1]),
                args=[
                    cupy.int(n),
                    cupy.int(poly_sides),
                    cupy.float32(init_angle),
                    image.data_ptr(),
                    defocus.data_ptr(),
                    defocus_dilate.data_ptr(),
                    bokeh_cum.data_ptr(),
                    weight_cum.data_ptr()
                ]
            )

        elif defocus.is_cuda == False:
            raise NotImplementedError()

        # end

        return defocus_dilate.float(), bokeh_cum, weight_cum
    # end

    # @staticmethod
    # def backward(self, gradBokehCum, gradWeightCum):
    # end

# end

def FunctionRender(image, defocus, poly_sides, init_angle):
    defocus_dilate, bokeh_cum, weight_cum = _FunctionRender.apply(image, defocus, poly_sides, init_angle)

    return defocus_dilate, bokeh_cum, weight_cum
# end

class ModuleRenderScatterEX(torch.nn.Module):
    def __init__(self):
        super(ModuleRenderScatterEX, self).__init__()
    # end

    def forward(self, image, defocus, poly_sides=10000, init_angle=3.1415926536/2):
        defocus_dilate, bokeh_cum, weight_cum = FunctionRender(image, defocus, poly_sides, init_angle)
        bokeh = bokeh_cum / weight_cum
        return bokeh, defocus_dilate
    # end
# end


================================================
FILE: demo.py
================================================
#!/usr/bin/env python
# encoding: utf-8

import os

# os.environ['CUDA_VISIBLE_DEVICES'] = '7'

import matplotlib.pyplot as plt
import numpy as np
import cv2
import argparse

import torch
import torch.nn.functional as F

from neural_renderer import ARNet, IUNet

from classical_renderer.scatter import ModuleRenderScatter  # circular aperture
from classical_renderer.scatter_ex import ModuleRenderScatterEX  # adjustable aperture shape


def gaussian_blur(x, r, sigma=None):
    r = int(round(r))
    if sigma is None:
        sigma = 0.3 * (r - 1) + 0.8
    x_grid, y_grid = torch.meshgrid(torch.arange(-int(r), int(r) + 1), torch.arange(-int(r), int(r) + 1))
    kernel = torch.exp(-(x_grid ** 2 + y_grid ** 2) / 2 / sigma ** 2)
    kernel = kernel.float() / kernel.sum()
    kernel = kernel.expand(1, 1, 2*r+1, 2*r+1).to(x.device)
    x = F.pad(x, pad=(r, r, r, r), mode='replicate')
    x = F.conv2d(x, weight=kernel, padding=0)
    return x


def pipeline(classical_renderer, arnet, iunet, image, defocus, gamma, args):
    bokeh_classical, defocus_dilate = classical_renderer(image**gamma, defocus*args.defocus_scale)
    # bokeh_classical, defocus_dilate = classical_renderer_ex(image**gamma, defocus*args.defocus_scale, poly_sides=6)

    bokeh_classical = bokeh_classical ** (1/gamma)
    defocus_dilate = defocus_dilate / args.defocus_scale
    gamma = (gamma - args.gamma_min) / (args.gamma_max - args.gamma_min)
    adapt_scale = max(defocus.abs().max().item(), 1)

    image_re = F.interpolate(image, scale_factor=1/adapt_scale, mode='bilinear', align_corners=True)
    defocus_re = 1 / adapt_scale * F.interpolate(defocus, scale_factor=1/adapt_scale, mode='bilinear', align_corners=True)
    bokeh_neural, error_map = arnet(image_re, defocus_re, gamma)
    error_map = F.interpolate(error_map, size=(image.shape[2], image.shape[3]), mode='bilinear', align_corners=True)
    bokeh_neural.clamp_(0, 1e5)

    if args.save_intermediate:
        cv2.imwrite(os.path.join(save_root, 'bokeh_neural_s0.jpg'), bokeh_neural[0].cpu().permute(1, 2, 0).numpy()[..., ::-1] * 255)

    scale = -1
    for scale in range(int(np.log2(adapt_scale))):
        ratio = 2**(scale+1) / adapt_scale
        h_re, w_re = int(ratio * image.shape[2]), int(ratio * image.shape[3])
        image_re = F.interpolate(image, size=(h_re, w_re), mode='bilinear', align_corners=True)
        defocus_re = ratio * F.interpolate(defocus, size=(h_re, w_re), mode='bilinear', align_corners=True)
        defocus_dilate_re = ratio * F.interpolate(defocus_dilate, size=(h_re, w_re), mode='bilinear', align_corners=True)
        bokeh_neural_refine = iunet(image_re, defocus_re.clamp(-1, 1), bokeh_neural, gamma).clamp(0, 1e5)
        mask = gaussian_blur(((defocus_dilate_re < 1) * (defocus_dilate_re > -1)).float(), 0.005 * (defocus_dilate_re.shape[2] + defocus_dilate_re.shape[3]))
        bokeh_neural = mask * bokeh_neural_refine + (1 - mask) * F.interpolate(bokeh_neural, size=(h_re, w_re), mode='bilinear', align_corners=True)
        if args.save_intermediate:
            cv2.imwrite(os.path.join(save_root, f'bokeh_neural_s{scale+1}.jpg'), bokeh_neural[0].cpu().permute(1, 2, 0).numpy()[..., ::-1] * 255)
            cv2.imwrite(os.path.join(save_root, f'fmask_neural_s{scale+1}.jpg'), mask[0][0].cpu().numpy() * 255)

    bokeh_neural_refine = iunet(image, defocus.clamp(-1, 1), bokeh_neural, gamma).clamp(0, 1e5)
    mask = gaussian_blur(((defocus_dilate < 1) * (defocus_dilate > -1)).float(), 0.005 * (defocus_dilate.shape[2] + defocus_dilate.shape[3]))
    bokeh_neural = mask * bokeh_neural_refine + (1 - mask) * F.interpolate(bokeh_neural, size=(image.shape[2], image.shape[3]), mode='bilinear', align_corners=True)
    if args.save_intermediate:
        cv2.imwrite(os.path.join(save_root, f'bokeh_neural_s{scale+2}.jpg'), bokeh_neural[0].cpu().permute(1, 2, 0).numpy()[..., ::-1] * 255)
        cv2.imwrite(os.path.join(save_root, f'fmask_neural_s{scale+2}.jpg'), mask[0][0].cpu().numpy() * 255)

    bokeh_pred = bokeh_classical * (1 - error_map) + bokeh_neural * error_map

    return bokeh_pred.clamp(0, 1), bokeh_classical.clamp(0, 1), bokeh_neural.clamp(0, 1), error_map




device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

parser = argparse.ArgumentParser(description='Bokeh Rendering', fromfile_prefix_chars='@')


parser.add_argument('--defocus_scale',               type=float, default=10.)
parser.add_argument('--gamma_min',                   type=float, default=1.)
parser.add_argument('--gamma_max',                   type=float, default=5.)

# Model 1
parser.add_argument('--arnet_shuffle_rate',          type=int,   default=2)
parser.add_argument('--arnet_in_channels',           type=int,   default=5)
parser.add_argument('--arnet_out_channels',          type=int,   default=4)
parser.add_argument('--arnet_middle_channels',       type=int,   default=128)
parser.add_argument('--arnet_num_block',             type=int,   default=3)
parser.add_argument('--arnet_share_weight',                      action='store_true')
parser.add_argument('--arnet_connect_mode',          type=str,   default='distinct_source')
parser.add_argument('--arnet_use_bn',                            action='store_true')
parser.add_argument('--arnet_activation',            type=str,   default='elu')

# Model 2
parser.add_argument('--iunet_shuffle_rate',          type=int,   default=2)
parser.add_argument('--iunet_in_channels',           type=int,   default=8)
parser.add_argument('--iunet_out_channels',          type=int,   default=3)
parser.add_argument('--iunet_middle_channels',       type=int,   default=64)
parser.add_argument('--iunet_num_block',             type=int,   default=3)
parser.add_argument('--iunet_share_weight',                      action='store_true')
parser.add_argument('--iunet_connect_mode',          type=str,   default='distinct_source')
parser.add_argument('--iunet_use_bn',                            action='store_true')
parser.add_argument('--iunet_activation',            type=str,   default='elu')

# Checkpoint
parser.add_argument('--arnet_checkpoint_path',       type=str,   default='./checkpoints/arnet.pth')
parser.add_argument('--iunet_checkpoint_path',       type=str,   default='./checkpoints/iunet.pth')

# Input
parser.add_argument('--image_path',                  type=str,   default='./inputs/21.jpg')
parser.add_argument('--disp_path',                   type=str,   default='./inputs/21.png')
parser.add_argument('--save_dir',                    type=str,   default='./outputs')
parser.add_argument('--K',                           type=float, default=60,          help='blur parameter')
parser.add_argument('--disp_focus',                  type=float, default=90/255,      help='refocused disparity (0~1)')
parser.add_argument('--gamma',                       type=float, default=4,           help='gamma value (1~5)')

parser.add_argument('--highlight',                               action='store_true', help='forcibly enchance RGB values of highlights')
parser.add_argument('--highlight_RGB_threshold',     type=float, default=220/255)
parser.add_argument('--highlight_enhance_ratio',     type=float, default=0.4)

parser.add_argument('--save_intermediate',                       action='store_true', help='save intermediate results')

args = parser.parse_args()

arnet_checkpoint_path = args.arnet_checkpoint_path
iunet_checkpoint_path = args.iunet_checkpoint_path

classical_renderer = ModuleRenderScatter().to(device)
# classical_renderer_ex = ModuleRenderScatterEX().to(device)

arnet = ARNet(args.arnet_shuffle_rate, args.arnet_in_channels, args.arnet_out_channels, args.arnet_middle_channels,
              args.arnet_num_block, args.arnet_share_weight, args.arnet_connect_mode, args.arnet_use_bn, args.arnet_activation)
iunet = IUNet(args.iunet_shuffle_rate, args.iunet_in_channels, args.iunet_out_channels, args.iunet_middle_channels,
              args.iunet_num_block, args.iunet_share_weight, args.iunet_connect_mode, args.iunet_use_bn, args.iunet_activation)

arnet.cuda()
iunet.cuda()

checkpoint = torch.load(arnet_checkpoint_path)
arnet.load_state_dict(checkpoint['model'])
checkpoint = torch.load(iunet_checkpoint_path)
iunet.load_state_dict(checkpoint['model'])

arnet.eval()
iunet.eval()

save_root = os.path.join(args.save_dir, os.path.splitext(os.path.basename(args.image_path))[0])
os.makedirs(save_root, exist_ok=True)

K = args.K                     # blur parameter
disp_focus = args.disp_focus   # 0~1
gamma = args.gamma             # 1~5

image = cv2.imread(args.image_path).astype(np.float32) / 255.0
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image_ori = image.copy()

disp = np.float32(cv2.imread(args.disp_path, cv2.IMREAD_GRAYSCALE))
disp = (disp - disp.min()) / (disp.max() - disp.min())

########## Highlights ##########
if args.highlight:
    mask1 = np.clip(np.tanh(200 * (np.abs(disp - disp_focus)**2 - 0.01)), 0, 1)[..., np.newaxis]  # out-of-focus areas
    # mask2 = (np.max(image, axis=2, keepdims=True) > args.highlight_RGB_threshold)  # highlight areas
    mask2 = np.clip(np.tanh(10*(image - args.highlight_RGB_threshold)), 0, 1)    # highlight areas
    mask = mask1 * mask2
    image = image * (1 + mask * args.highlight_enhance_ratio)
################################


defocus = K * (disp - disp_focus) / args.defocus_scale

with torch.no_grad():
    image = torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0)
    defocus = torch.from_numpy(defocus).unsqueeze(0).unsqueeze(0)
    image = image.cuda()
    defocus = defocus.cuda()

    bokeh_pred, bokeh_classical, bokeh_neural, error_map = pipeline(
        classical_renderer, arnet, iunet, image, defocus, gamma, args
    )


defocus = defocus[0][0].cpu().numpy()
error_map = error_map[0][0].cpu().numpy()
bokeh_classical = bokeh_classical[0].cpu().permute(1, 2, 0).numpy()
bokeh_neural = bokeh_neural[0].cpu().permute(1, 2, 0).detach().numpy()
bokeh_pred = bokeh_pred[0].cpu().permute(1, 2, 0).detach().numpy()

cv2.imwrite(os.path.join(save_root, 'image.jpg'), image_ori[..., ::-1] * 255)
plt.imsave(os.path.join(save_root, 'defocus.jpg'), defocus, cmap='coolwarm', vmin=-max(defocus.max(), -defocus.min()), vmax=max(defocus.max(), -defocus.min()))
cv2.imwrite(os.path.join(save_root, 'disparity.jpg'), disp * 255)
cv2.imwrite(os.path.join(save_root, 'error_map.jpg'), error_map * 255)
cv2.imwrite(os.path.join(save_root, 'bokeh_classical.jpg'), bokeh_classical[..., ::-1] * 255)
cv2.imwrite(os.path.join(save_root, 'bokeh_neural.jpg'), bokeh_neural[..., ::-1] * 255)
cv2.imwrite(os.path.join(save_root, 'bokeh_pred.jpg'), bokeh_pred[..., ::-1] * 255)


================================================
FILE: neural_renderer.py
================================================
#!/usr/bin/env python
# encoding: utf-8
import os

# os.environ['CUDA_VISIBLE_DEVICES'] = '5'

import torch
import torch.nn as nn
import torch.nn.functional as F


class Space2Depth(nn.Module):
    def __init__(self, down_factor):
        super(Space2Depth, self).__init__()
        self.down_factor = down_factor

    def forward(self, x):
        n, c, h, w = x.size()
        unfolded_x = torch.nn.functional.unfold(x, self.down_factor, stride=self.down_factor)
        return unfolded_x.view(n, c * self.down_factor ** 2, h // self.down_factor, w // self.down_factor)


def conv_bn_activation(in_channels, out_channels, kernel_size, stride, padding, use_bn, activation):
    module = nn.Sequential()
    # module.add_module('pad', nn.ReflectionPad2d(padding))
    module.add_module('conv', nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding))
    module.add_module('bn', nn.BatchNorm2d(out_channels)) if use_bn else None
    module.add_module('activation', activation) if activation else None

    return module


class BlockStack(nn.Module):
    def __init__(self, channels, num_block, share_weight, connect_mode, use_bn, activation):
        # connect_mode: refer to "Fast and Accurate Image Super-Resolution with Deep Laplacian Pyramid Networks"
        super(BlockStack, self).__init__()

        self.num_block = num_block
        self.connect_mode = connect_mode

        self.blocks = nn.ModuleList()

        if share_weight is True:
            block = nn.Sequential(
                conv_bn_activation(
                    in_channels=channels,
                    out_channels=channels,
                    kernel_size=3, stride=1, padding=1,
                    use_bn=use_bn, activation=activation
                ),
                conv_bn_activation(
                    in_channels=channels,
                    out_channels=channels,
                    kernel_size=3, stride=1, padding=1,
                    use_bn=use_bn, activation=activation
                )
            )
            for i in range(num_block):
                self.blocks.append(block)

        else:
            for i in range(num_block):
                block = nn.Sequential(
                    conv_bn_activation(
                        in_channels=channels,
                        out_channels=channels,
                        kernel_size=3, stride=1, padding=1,
                        use_bn=use_bn, activation=activation
                    ),
                    conv_bn_activation(
                        in_channels=channels,
                        out_channels=channels,
                        kernel_size=3, stride=1, padding=1,
                        use_bn=use_bn, activation=activation
                    )
                )
                self.blocks.append(block)

    def forward(self, x):
        if self.connect_mode == 'no':
            for i in range(self.num_block):
                x = self.blocks[i](x)
        elif self.connect_mode == 'distinct_source':
            for i in range(self.num_block):
                x = self.blocks[i](x) + x
        elif self.connect_mode == 'shared_source':
            x0 = x
            for i in range(self.num_block):
                x = self.blocks[i](x) + x0
        else:
            print('"connect_mode" error!')
            exit(0)
        return x


class ARNet(nn.Module):  # Adaptive Rendering Network
    def __init__(self, shuffle_rate=2, in_channels=5, out_channels=4, middle_channels=128, num_block=3, share_weight=False, connect_mode='distinct_source', use_bn=False, activation='elu'):
        super(ARNet, self).__init__()

        self.shuffle_rate = shuffle_rate
        self.connect_mode = connect_mode

        if activation == 'relu':
            activation = nn.ReLU(inplace=True)
        elif activation == 'leaky_relu':
            activation = nn.LeakyReLU(inplace=True)
        elif activation == 'elu':
            activation = nn.ELU(inplace=True)
        else:
            print('"activation" error!')
            exit(0)

        self.downsample = Space2Depth(shuffle_rate)
        self.conv0 = conv_bn_activation(
            in_channels=(in_channels - 1) * shuffle_rate ** 2 + 1,
            out_channels=middle_channels,
            kernel_size=3, stride=1, padding=1,
            use_bn=use_bn, activation=activation
        )
        self.block_stack = BlockStack(
            channels=middle_channels,
            num_block=num_block, share_weight=share_weight, connect_mode=connect_mode,
            use_bn=use_bn, activation=activation
        )
        self.conv1 = conv_bn_activation(
            in_channels=middle_channels,
            out_channels=out_channels * shuffle_rate ** 2,
            kernel_size=3, stride=1, padding=1,
            use_bn=False, activation=None
        )
        self.upsample = nn.PixelShuffle(shuffle_rate)

    def forward(self, image, defocus, gamma):
        _, _, h, w = image.shape
        h_re = int(h // self.shuffle_rate * self.shuffle_rate)
        w_re = int(w // self.shuffle_rate * self.shuffle_rate)
        x = torch.cat((image, defocus), dim=1)
        x = F.interpolate(x, size=(h_re, w_re), mode='bilinear', align_corners=True)
        x = self.downsample(x)
        gamma = torch.ones_like(x[:, :1]) * gamma
        x = torch.cat((x, gamma), dim=1)
        x = self.conv0(x)
        x = self.block_stack(x)
        x = self.conv1(x)
        x = self.upsample(x)
        x = F.interpolate(x, size=(h, w), mode='bilinear', align_corners=True)

        bokeh = x[:, :-1]
        mask = torch.sigmoid(x[:, -1:])

        return bokeh, mask


class IUNet(nn.Module):  # Iterative Upsampling Network
    def __init__(self, shuffle_rate=2, in_channels=8, out_channels=3, middle_channels=64, num_block=3, share_weight=False, connect_mode='distinct_source', use_bn=False, activation='elu'):
        super(IUNet, self).__init__()

        self.shuffle_rate = shuffle_rate
        self.connect_mode = connect_mode

        if activation == 'relu':
            activation = nn.ReLU(inplace=True)
        elif activation == 'leaky_relu':
            activation = nn.LeakyReLU(inplace=True)
        elif activation == 'elu':
            activation = nn.ELU(inplace=True)
        else:
            print('"activation" error!')
            exit(0)

        self.downsample = Space2Depth(shuffle_rate)
        self.conv0 = conv_bn_activation(
            in_channels=(in_channels - 4) * shuffle_rate ** 2 + 4,
            out_channels=middle_channels,
            kernel_size=3, stride=1, padding=1,
            use_bn=use_bn, activation=activation
        )
        self.block_stack = BlockStack(
            channels=middle_channels,
            num_block=num_block, share_weight=share_weight, connect_mode=connect_mode,
            use_bn=use_bn, activation=activation
        )
        self.conv1 = conv_bn_activation(
            in_channels=middle_channels,
            out_channels=out_channels * shuffle_rate ** 2,
            kernel_size=3, stride=1, padding=1,
            use_bn=False, activation=None
        )
        self.upsample = nn.PixelShuffle(shuffle_rate)

    def forward(self, image, defocus, bokeh_coarse, gamma):
        _, _, h, w = image.shape
        h_re = int(h // self.shuffle_rate * self.shuffle_rate)
        w_re = int(w // self.shuffle_rate * self.shuffle_rate)
        x = torch.cat((image, defocus), dim=1)
        x = F.interpolate(x, size=(h_re, w_re), mode='bilinear', align_corners=True)
        x = self.downsample(x)
        if bokeh_coarse.shape[2] != x.shape[2] or bokeh_coarse.shape[3] != x.shape[3]:
            bokeh_coarse = F.interpolate(bokeh_coarse, size=(x.shape[2], x.shape[3]), mode='bilinear', align_corners=False)
        gamma = torch.ones_like(x[:, :1]) * gamma
        x = torch.cat((x, bokeh_coarse, gamma), dim=1)
        x = self.conv0(x)
        x = self.block_stack(x)
        x = self.conv1(x)
        x = self.upsample(x)
        bokeh_refine = F.interpolate(x, size=(h, w), mode='bilinear', align_corners=True)

        return bokeh_refine


================================================
FILE: requirements.txt
================================================
cupy==10.5.0
cupy_cuda90==7.7.0
matplotlib==3.5.1
numpy==1.18.5
opencv_python==4.2.0.34
Pillow==9.1.1
torch==1.8.1
torchvision==0.9.1