├── .gitignore ├── LICENSE ├── README.md ├── configs ├── __init__.py ├── data_configs.py ├── paths_config.py └── transforms_config.py ├── criteria ├── .ipynb_checkpoints │ └── id_loss-checkpoint.py ├── __init__.py ├── id_loss.py ├── lpips │ ├── __init__.py │ ├── lpips.py │ ├── networks.py │ └── utils.py ├── moco_loss.py └── w_norm.py ├── demo.py ├── docs ├── custom-color-edits.png ├── prog-synthesis.png └── watercolor-synthesis.png ├── environment └── paint2pix_env.yml ├── input-images ├── 143.jpg ├── 20.jpg ├── 214.jpg ├── 262.jpg └── 823.jpg ├── licenses ├── LICENSE_S-aiueo32 ├── LICENSE_TreB1eN ├── LICENSE_alaluf ├── LICENSE_eladrich ├── LICENSE_lessw2020 ├── LICENSE_omertov └── LICENSE_rosinality ├── models ├── .ipynb_checkpoints │ └── e4e-checkpoint.py ├── __init__.py ├── e4e.py ├── e4e_modules │ ├── __init__.py │ ├── discriminator.py │ └── latent_codes_pool.py ├── encoders │ ├── .ipynb_checkpoints │ │ ├── map2style-checkpoint.py │ │ └── restyle_e4e_encoders-checkpoint.py │ ├── __init__.py │ ├── fpn_encoders.py │ ├── helpers.py │ ├── map2style.py │ ├── model_irse.py │ ├── restyle_e4e_encoders.py │ └── restyle_psp_encoders.py ├── mtcnn │ ├── __init__.py │ ├── mtcnn.py │ └── mtcnn_pytorch │ │ ├── __init__.py │ │ └── src │ │ ├── __init__.py │ │ ├── align_trans.py │ │ ├── box_utils.py │ │ ├── detector.py │ │ ├── first_stage.py │ │ ├── get_nets.py │ │ ├── matlab_cp2tform.py │ │ ├── visualization_utils.py │ │ └── weights │ │ ├── onet.npy │ │ ├── pnet.npy │ │ └── rnet.npy ├── psp.py └── stylegan2 │ ├── .ipynb_checkpoints │ └── model-checkpoint.py │ ├── __init__.py │ ├── model.py │ └── op │ ├── __init__.py │ ├── fused_act.py │ ├── fused_bias_act.cpp │ ├── fused_bias_act_kernel.cu │ ├── upfirdn2d.cpp │ ├── upfirdn2d.py │ └── upfirdn2d_kernel.cu ├── output ├── result_0.png └── result_mask_0.png ├── predict.py └── utils ├── .ipynb_checkpoints ├── id_utils-checkpoint.py └── inference_utils-checkpoint.py ├── __init__.py ├── common.py ├── data_utils.py ├── id_utils.py ├── inference_utils.py ├── model_utils.py └── train_utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | notebooks 2 | data 3 | face_seg 4 | intellipaint 5 | files 6 | __pycache__ 7 | outputs 8 | pretrained_models 9 | *.dat -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2022 Jaskirat Singh 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Paint2Pix: Interactive Painting based Progressive Image Synthesis and Editing (ECCV 2022) 2 | 3 | > Jaskirat Singh, Liang Zheng, Cameron Smith, Jose Echevarria 4 | > 5 | > Controllable image synthesis with user scribbles is a topic of keen interest in the computer vision community. In this paper, for the first time we study the problem of photorealistic image synthesis from incomplete and primitive human paintings. In particular, we propose a novel approach paint2pix, which learns to predict (and adapt) “what a user wants to draw” from rudimentary brushstroke inputs, by learning a mapping from the manifold of incomplete human paintings to their realistic renderings. When used in conjunction with recent works in autonomous painting agents, we show that paint2pix can be used for progressive image synthesis from scratch. During this process, paint2pix allows a novice user to progressively synthesize the desired image output, while requiring just few coarse user scribbles to accurately steer the trajectory of the synthesis process. Furthermore, we find that our approach also forms a surprisingly convenient approach for real image editing, and allows the user to perform a diverse range of custom fine-grained edits through the addition of only a few well-placed brushstrokes. 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 |

17 | 18 |
19 | We propose paint2pix which helps the user directly express his/her ideas in visual form by learning to predict user-intention from a few rudimentary brushstrokes. The proposed approach can be used for (a) synthesizing a desired image output directly from scratch 20 | wherein it allows the user to control the overall synthesis trajectory using just few coarse brushstrokes (blue arrows) at key points, 21 | or, (b) performing a diverse range of custom edits directly on real image inputs. 22 |

23 | 24 | ## Description 25 | Official implementation of our Paint2pix paper with streamlit demo. By using autonomous painting agents as a proxy for the human painting process, Paint2pix learns to predict *user-intention* ("what a user wants to draw") from fairly rudimentary paintings and user-scribbles. 26 | 27 | ## Updates 28 | 29 | * **(19/08/22)** Our [project demo](http://exposition.cecs.anu.edu.au:6009/) is online. Try generating amazing artwork or realistic image media right from your browser! 30 | 31 | https://user-images.githubusercontent.com/25987491/185323657-a71c239c-892c-4202-b753-a84c0bf19a30.mp4 32 | 33 | 34 | 35 | 36 | 37 | ## Table of Contents 38 | - [Getting Started](#getting-started) 39 | * [Prerequisites](#prerequisites) 40 | * [Installation](#installation) 41 | - [Pretrained Models](#pretrained-models) 42 | * [Paint2pix models](#paint2pix-models) 43 | - [Using the Demo](#using-the-demo) 44 | - [Example Results](#example-results) 45 | * [Progressive Image Synthesis](#progressive-image-synthesis) 46 | * [Real Image Editing](#real-image-editing) 47 | * [Artistic Content Generation](#artistic-content-generation) 48 | - [Acknowledgments](#acknowledgments) 49 | - [Citation](#citation) 50 | 51 | Table of contents generated with markdown-toc 52 | 53 | 54 | 55 | ## Getting Started 56 | ### Prerequisites 57 | - Linux or macOS 58 | - NVIDIA GPU + CUDA CuDNN (CPU may be possible with some modifications, but is not inherently supported) 59 | - Python 3 60 | - Tested on Ubuntu 20.04, Nvidia RTX 3090 and CUDA 11.5 61 | 62 | ### Installation 63 | - Dependencies: 64 | We recommend running this repository using [Anaconda](https://docs.anaconda.com/anaconda/install/). 65 | All dependencies for defining the environment are provided in `environment/paint2pix_env.yaml`. 66 | 67 | ## Pretrained Models 68 | Please download the following pretrained models essential for running the provided demo. 69 | 70 | ### Paint2pix models 71 | | Path | Description 72 | | :--- | :---------- 73 | |[Canvas Encoder - ReStyle](https://drive.google.com/file/d/1ufKEtDXEG6o96KjLh-i6EL7Ir9TlwPcs/view?usp=sharing) | Paint2pix Canvas Encoder trained with a ReStyle architecture. 74 | |[Identity Encoder - ReStyle](https://drive.google.com/file/d/1KT3YmSHgMJM3b7Ox9zciyo3FSELtJsyS/view?usp=sharing) | Paint2pix Identity Encoder trained with a ReStyle architecture. 75 | |[StyleGAN - Watercolor Painting](https://drive.google.com/file/d/1WW_a589lv7R9-PNvKlVkVxITIZnW7xlv/view?usp=sharing) | StyleGAN decoder network trained to generate watercolor paintings. Used for artistic content generation with paint2pix. 76 | |[IR-SE50 Model](https://drive.google.com/file/d/1U4q_o20uGMozSetOkMGddUcAWf_ons2-/view?usp=sharing) | Pretrained IR-SE50 model taken from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch) for use in ID loss and id-encoder training. 77 | 78 | 79 | Please download and save the above models to the directory `pretrained_models`. 80 | 81 | 82 | ## Using the Demo 83 | 84 | We provide a streamlit-drawble canvas based demo for trying out different features of the Paint2pix model. To start the demo use, 85 | 86 | ``` 87 | CUDA_VISIBLE_DEVICES=2 streamlit run demo.py --server.port 6009 88 | ``` 89 | 90 | The demo can then be accessed on the local machine or ssh client via [localhost](http://localhost:6009). 91 | 92 | The demo has been divided into 3 convenient sections: 93 | 94 | 1. **Real Image Editing**: Allows the user to edit real images using coarse user scribbles 95 | 2. **Progressive Image Synthesis**: Start from an empty canvas and design your desired image output using just coarse scribbles. 96 | 3. **Artistic Content Generation**: Unleash your inner artist! create highly artistic portraits using just coarse scribbles. 97 | 98 | 99 | ## Example Results 100 | 101 | ### Progressive Image Synthesis 102 | 103 |

104 | 105 |
106 | Paint2pix for progressive image synthesis 107 |

108 | 109 | ### Real Image Editing 110 | 111 |

112 | 113 |
114 | Paint2pix for achieving diverse custom real-image edits 115 |

116 | 117 | ### Artistic Content Generation 118 | 119 | 120 |

121 | 122 |
123 | Paint2pix for generating highly artistic content using coarse scribbles 124 |

125 | 126 | ## Acknowledgments 127 | This code borrows heavily from [pixel2style2pixel](https://github.com/eladrich/pixel2style2pixel), 128 | [encoder4editing](https://github.com/omertov/encoder4editing) and [restyle-encoder](https://github.com/yuval-alaluf/restyle-encoder). 129 | 130 | ## Citation 131 | If you use this code for your research, please cite the following works: 132 | ``` 133 | @inproceedings{singh2022paint2pix, 134 | title={Paint2Pix: Interactive Painting based Progressive 135 | Image Synthesis and Editing}, 136 | author={Singh, Jaskirat and Zheng, Liang and Smith, Cameron and Echevarria, Jose}, 137 | booktitle={European conference on computer vision}, 138 | year={2022}, 139 | organization={Springer} 140 | } 141 | ``` 142 | ``` 143 | @inproceedings{singh2022intelli, 144 | title={Intelli-Paint: Towards Developing Human-like Painting Agents}, 145 | author={Singh, Jaskirat and Smith, Cameron and Echevarria, Jose and Zheng, Liang}, 146 | booktitle={European conference on computer vision}, 147 | year={2022}, 148 | organization={Springer} 149 | } 150 | ``` 151 | 152 | -------------------------------------------------------------------------------- /configs/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/configs/__init__.py -------------------------------------------------------------------------------- /configs/data_configs.py: -------------------------------------------------------------------------------- 1 | from configs import transforms_config 2 | from configs.paths_config import dataset_paths 3 | 4 | 5 | DATASETS = { 6 | 'ffhq_encode': { 7 | 'transforms': transforms_config.EncodeTransforms, 8 | 'train_source_root': dataset_paths['ffhq'], 9 | 'train_target_root': dataset_paths['ffhq'], 10 | 'test_source_root': dataset_paths['celeba_test'], 11 | 'test_target_root': dataset_paths['celeba_test'] 12 | }, 13 | "cars_encode_prev": { 14 | 'transforms': transforms_config.CarsEncodeTransforms, 15 | 'train_source_root': dataset_paths['cars_train'], 16 | 'train_target_root': dataset_paths['cars_train'], 17 | 'test_source_root': dataset_paths['cars_test'], 18 | 'test_target_root': dataset_paths['cars_test'] 19 | }, 20 | "church_encode": { 21 | 'transforms': transforms_config.EncodeTransforms, 22 | 'train_source_root': dataset_paths['church_train'], 23 | 'train_target_root': dataset_paths['church_train'], 24 | 'test_source_root': dataset_paths['church_test'], 25 | 'test_target_root': dataset_paths['church_test'] 26 | }, 27 | "horse_encode": { 28 | 'transforms': transforms_config.EncodeTransforms, 29 | 'train_source_root': dataset_paths['horse_train'], 30 | 'train_target_root': dataset_paths['horse_train'], 31 | 'test_source_root': dataset_paths['horse_test'], 32 | 'test_target_root': dataset_paths['horse_test'] 33 | }, 34 | "afhq_wild_encode": { 35 | 'transforms': transforms_config.EncodeTransforms, 36 | 'train_source_root': dataset_paths['afhq_wild_train'], 37 | 'train_target_root': dataset_paths['afhq_wild_train'], 38 | 'test_source_root': dataset_paths['afhq_wild_test'], 39 | 'test_target_root': dataset_paths['afhq_wild_test'] 40 | }, 41 | "toonify": { 42 | 'transforms': transforms_config.EncodeTransforms, 43 | 'train_source_root': dataset_paths['ffhq'], 44 | 'train_target_root': dataset_paths['ffhq'], 45 | 'test_source_root': dataset_paths['celeba_test'], 46 | 'test_target_root': dataset_paths['celeba_test'] 47 | }, 48 | 'paint_ffhq_encode': { 49 | 'transforms': transforms_config.EncodeTransforms, 50 | 'train_source_root': dataset_paths['paint_train_source_ffhq'], 51 | 'train_target_root': dataset_paths['paint_train_target_ffhq'], 52 | 'test_source_root': dataset_paths['paint_test_source_ffhq'], 53 | 'test_target_root': dataset_paths['paint_test_target_ffhq'], 54 | }, 55 | 'paint_ffhq_encode2': { 56 | 'transforms': transforms_config.EncodeTransforms, 57 | 'train_source_root': dataset_paths['paint_train_source_ffhq2'], 58 | 'train_target_root': dataset_paths['paint_train_target_ffhq2'], 59 | 'test_source_root': dataset_paths['paint_test_source_ffhq2'], 60 | 'test_target_root': dataset_paths['paint_test_target_ffhq2'], 61 | }, 62 | 'paint_ffhq_encode_id': { 63 | 'transforms': transforms_config.EncodeTransforms, 64 | 'train_source_root': dataset_paths['paint_train_target_ffhq2'], 65 | 'train_target_root': dataset_paths['paint_train_source_ffhq2'], 66 | 'test_source_root': dataset_paths['paint_test_target_ffhq2'], 67 | 'test_target_root': dataset_paths['paint_test_source_ffhq2'], 68 | }, 69 | 'cars_encode': { 70 | 'transforms': transforms_config.CarsEncodeTransforms, 71 | 'train_source_root': dataset_paths['paint_train_source'], 72 | 'train_target_root': dataset_paths['paint_train_target'], 73 | 'test_source_root': dataset_paths['paint_test_source'], 74 | 'test_target_root': dataset_paths['paint_test_target'], 75 | }, 76 | } -------------------------------------------------------------------------------- /configs/paths_config.py: -------------------------------------------------------------------------------- 1 | dataset_paths = { 2 | 'celeba': '', 3 | 'celeba_test': '', 4 | 5 | 'cars_train': '', 6 | 'cars_test': '', 7 | 8 | 'church_train': '', 9 | 'church_test': '', 10 | 11 | 'horse_train': '', 12 | 'horse_test': '', 13 | 14 | 'afhq_wild_train': '', 15 | 'afhq_wild_test': '', 16 | 17 | # Painting completion ffhq 18 | 'paint_train_source_ffhq': 'data/celeba/intelli-paint/train/painting', 19 | 'paint_train_target_ffhq': 'data/celeba/intelli-paint/train/target', 20 | 'paint_test_source_ffhq': 'data/celeba/intelli-paint/test/painting', 21 | 'paint_test_target_ffhq': 'data/celeba/intelli-paint/test/target', 22 | 23 | # Painting completion cars 24 | 'paint_train_source': 'data/intelli-paint/train/painting', 25 | 'paint_train_target': 'data/intelli-paint/train/target', 26 | 'paint_test_source': 'data/intelli-paint/test/painting', 27 | 'paint_test_target': 'data/intelli-paint/test/target', 28 | } 29 | 30 | model_paths = { 31 | 'ir_se50': 'pretrained_models/model_ir_se50.pth', 32 | 'resnet34': 'pretrained_models/resnet34-333f7ec4.pth', 33 | 'stylegan_celeba': 'pretrained_models/stylegan2-celeba-config-f.pt', 34 | 'stylegan_cars': 'pretrained_models/stylegan2-car-config-f.pt', 35 | 'stylegan_church': 'pretrained_models/stylegan2-church-config-f.pt', 36 | 'stylegan_horse': 'pretrained_models/stylegan2-horse-config-f.pt', 37 | 'stylegan_ada_wild': 'pretrained_models/afhqwild.pt', 38 | 'stylegan_toonify': 'pretrained_models/celeba_cartoon_blended.pt', 39 | 'shape_predictor': 'pretrained_models/shape_predictor_68_face_landmarks.dat', 40 | 'circular_face': 'pretrained_models/CurricularFace_Backbone.pth', 41 | 'mtcnn_pnet': 'pretrained_models/mtcnn/pnet.npy', 42 | 'mtcnn_rnet': 'pretrained_models/mtcnn/rnet.npy', 43 | 'mtcnn_onet': 'pretrained_models/mtcnn/onet.npy', 44 | 'moco': 'pretrained_models/moco_v2_800ep_pretrain.pth', 45 | 'id_encoder': 'pretrained_models/id-encoder.pt', 46 | 'canvas_encoder': 'pretrained_models/canvas-encoder.pt', 47 | 'stylegan_watercolor': 'pretrained_models/stylegan2-watercolor.pt', 48 | } 49 | -------------------------------------------------------------------------------- /configs/transforms_config.py: -------------------------------------------------------------------------------- 1 | from abc import abstractmethod 2 | import torchvision.transforms as transforms 3 | 4 | 5 | class TransformsConfig(object): 6 | 7 | def __init__(self, opts): 8 | self.opts = opts 9 | 10 | @abstractmethod 11 | def get_transforms(self): 12 | pass 13 | 14 | 15 | class EncodeTransforms(TransformsConfig): 16 | 17 | def __init__(self, opts): 18 | super(EncodeTransforms, self).__init__(opts) 19 | 20 | def get_transforms(self): 21 | transforms_dict = { 22 | 'transform_gt_train': transforms.Compose([ 23 | transforms.Resize((256, 256)), 24 | #transforms.RandomHorizontalFlip(0.5), 25 | transforms.ToTensor(), 26 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]), 27 | #'transform_source': None, 28 | 'transform_source': transforms.Compose([ 29 | transforms.Resize((256, 256)), 30 | #transforms.RandomHorizontalFlip(0.5), 31 | transforms.ToTensor(), 32 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]), 33 | 'transform_test': transforms.Compose([ 34 | transforms.Resize((256, 256)), 35 | transforms.ToTensor(), 36 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]), 37 | 'transform_inference': transforms.Compose([ 38 | transforms.Resize((256, 256)), 39 | transforms.ToTensor(), 40 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]) 41 | } 42 | return transforms_dict 43 | 44 | 45 | class CarsEncodeTransforms(TransformsConfig): 46 | 47 | def __init__(self, opts): 48 | super(CarsEncodeTransforms, self).__init__(opts) 49 | 50 | def get_transforms(self): 51 | transforms_dict = { 52 | 'transform_gt_train': transforms.Compose([ 53 | transforms.Resize((192, 256)), 54 | #transforms.RandomHorizontalFlip(0.5), 55 | transforms.ToTensor(), 56 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]), 57 | 'transform_source': transforms.Compose([ 58 | transforms.Resize((192, 256)), 59 | #transforms.RandomHorizontalFlip(0.5), 60 | transforms.ToTensor(), 61 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]), 62 | 'transform_test': transforms.Compose([ 63 | transforms.Resize((192, 256)), 64 | transforms.ToTensor(), 65 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]), 66 | 'transform_inference': transforms.Compose([ 67 | transforms.Resize((192, 256)), 68 | transforms.ToTensor(), 69 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]) 70 | } 71 | return transforms_dict 72 | -------------------------------------------------------------------------------- /criteria/.ipynb_checkpoints/id_loss-checkpoint.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | from configs.paths_config import model_paths 4 | from models.encoders.model_irse import Backbone 5 | 6 | 7 | class IDLoss(nn.Module): 8 | def __init__(self): 9 | super(IDLoss, self).__init__() 10 | print('Loading ResNet ArcFace') 11 | self.facenet = Backbone(input_size=112, num_layers=50, drop_ratio=0.6, mode='ir_se') 12 | self.facenet.load_state_dict(torch.load(model_paths['ir_se50'])) 13 | self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112)) 14 | self.facenet.eval() 15 | 16 | def extract_feats(self, x): 17 | x = x[:, :, 35:223, 32:220] # Crop interesting region 18 | x = self.face_pool(x) 19 | x_feats = self.facenet(x) 20 | return x_feats 21 | 22 | def forward(self, y_hat, y, x, target_id_feat=None): 23 | n_samples = x.shape[0] 24 | x_feats = self.extract_feats(x) 25 | 26 | if target_id_feat is not None: 27 | y_feats = target_id_feat#self.extract_feats(id_x) 28 | else: 29 | y_feats = self.extract_feats(y) # Otherwise use the feature from there 30 | 31 | y_hat_feats = self.extract_feats(y_hat) 32 | y_feats = y_feats.detach() 33 | loss = 0 34 | sim_improvement = 0 35 | id_logs = [] 36 | count = 0 37 | for i in range(n_samples): 38 | diff_target = y_hat_feats[i].dot(y_feats[i]) 39 | diff_input = y_hat_feats[i].dot(x_feats[i]) 40 | diff_views = y_feats[i].dot(x_feats[i]) 41 | id_logs.append({'diff_target': float(diff_target), 42 | 'diff_input': float(diff_input), 43 | 'diff_views': float(diff_views)}) 44 | loss += 1 - diff_target 45 | id_diff = float(diff_target) - float(diff_views) 46 | sim_improvement += id_diff 47 | count += 1 48 | 49 | return loss / count, sim_improvement / count, id_logs 50 | -------------------------------------------------------------------------------- /criteria/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/criteria/__init__.py -------------------------------------------------------------------------------- /criteria/id_loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | from configs.paths_config import model_paths 4 | from models.encoders.model_irse import Backbone 5 | 6 | 7 | class IDLoss(nn.Module): 8 | def __init__(self): 9 | super(IDLoss, self).__init__() 10 | print('Loading ResNet ArcFace') 11 | self.facenet = Backbone(input_size=112, num_layers=50, drop_ratio=0.6, mode='ir_se') 12 | self.facenet.load_state_dict(torch.load(model_paths['ir_se50'])) 13 | self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112)) 14 | self.facenet.eval() 15 | 16 | def extract_feats(self, x): 17 | x = x[:, :, 35:223, 32:220] # Crop interesting region 18 | x = self.face_pool(x) 19 | x_feats = self.facenet(x) 20 | return x_feats 21 | 22 | def forward(self, y_hat, y, x, target_id_feat=None): 23 | n_samples = x.shape[0] 24 | x_feats = self.extract_feats(x) 25 | 26 | if target_id_feat is not None: 27 | y_feats = target_id_feat#self.extract_feats(id_x) 28 | else: 29 | y_feats = self.extract_feats(y) # Otherwise use the feature from there 30 | 31 | y_hat_feats = self.extract_feats(y_hat) 32 | y_feats = y_feats.detach() 33 | loss = 0 34 | sim_improvement = 0 35 | id_logs = [] 36 | count = 0 37 | for i in range(n_samples): 38 | diff_target = y_hat_feats[i].dot(y_feats[i]) 39 | diff_input = y_hat_feats[i].dot(x_feats[i]) 40 | diff_views = y_feats[i].dot(x_feats[i]) 41 | id_logs.append({'diff_target': float(diff_target), 42 | 'diff_input': float(diff_input), 43 | 'diff_views': float(diff_views)}) 44 | loss += 1 - diff_target 45 | id_diff = float(diff_target) - float(diff_views) 46 | sim_improvement += id_diff 47 | count += 1 48 | 49 | return loss / count, sim_improvement / count, id_logs 50 | -------------------------------------------------------------------------------- /criteria/lpips/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/criteria/lpips/__init__.py -------------------------------------------------------------------------------- /criteria/lpips/lpips.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | from criteria.lpips.networks import get_network, LinLayers 5 | from criteria.lpips.utils import get_state_dict 6 | 7 | 8 | class LPIPS(nn.Module): 9 | r"""Creates a criterion that measures 10 | Learned Perceptual Image Patch Similarity (LPIPS). 11 | Arguments: 12 | net_type (str): the network type to compare the features: 13 | 'alex' | 'squeeze' | 'vgg'. Default: 'alex'. 14 | version (str): the version of LPIPS. Default: 0.1. 15 | """ 16 | def __init__(self, net_type: str = 'alex', version: str = '0.1'): 17 | 18 | assert version in ['0.1'], 'v0.1 is only supported now' 19 | 20 | super(LPIPS, self).__init__() 21 | 22 | # pretrained network 23 | self.net = get_network(net_type).to("cuda") 24 | 25 | # linear layers 26 | self.lin = LinLayers(self.net.n_channels_list).to("cuda") 27 | self.lin.load_state_dict(get_state_dict(net_type, version)) 28 | 29 | def forward(self, x: torch.Tensor, y: torch.Tensor): 30 | feat_x, feat_y = self.net(x), self.net(y) 31 | 32 | diff = [(fx - fy) ** 2 for fx, fy in zip(feat_x, feat_y)] 33 | res = [l(d).mean((2, 3), True) for d, l in zip(diff, self.lin)] 34 | 35 | return torch.sum(torch.cat(res, 0)) / x.shape[0] 36 | -------------------------------------------------------------------------------- /criteria/lpips/networks.py: -------------------------------------------------------------------------------- 1 | from typing import Sequence 2 | 3 | from itertools import chain 4 | 5 | import torch 6 | import torch.nn as nn 7 | from torchvision import models 8 | 9 | from criteria.lpips.utils import normalize_activation 10 | 11 | 12 | def get_network(net_type: str): 13 | if net_type == 'alex': 14 | return AlexNet() 15 | elif net_type == 'squeeze': 16 | return SqueezeNet() 17 | elif net_type == 'vgg': 18 | return VGG16() 19 | else: 20 | raise NotImplementedError('choose net_type from [alex, squeeze, vgg].') 21 | 22 | 23 | class LinLayers(nn.ModuleList): 24 | def __init__(self, n_channels_list: Sequence[int]): 25 | super(LinLayers, self).__init__([ 26 | nn.Sequential( 27 | nn.Identity(), 28 | nn.Conv2d(nc, 1, 1, 1, 0, bias=False) 29 | ) for nc in n_channels_list 30 | ]) 31 | 32 | for param in self.parameters(): 33 | param.requires_grad = False 34 | 35 | 36 | class BaseNet(nn.Module): 37 | def __init__(self): 38 | super(BaseNet, self).__init__() 39 | 40 | # register buffer 41 | self.register_buffer( 42 | 'mean', torch.Tensor([-.030, -.088, -.188])[None, :, None, None]) 43 | self.register_buffer( 44 | 'std', torch.Tensor([.458, .448, .450])[None, :, None, None]) 45 | 46 | def set_requires_grad(self, state: bool): 47 | for param in chain(self.parameters(), self.buffers()): 48 | param.requires_grad = state 49 | 50 | def z_score(self, x: torch.Tensor): 51 | return (x - self.mean) / self.std 52 | 53 | def forward(self, x: torch.Tensor): 54 | x = self.z_score(x) 55 | 56 | output = [] 57 | for i, (_, layer) in enumerate(self.layers._modules.items(), 1): 58 | x = layer(x) 59 | if i in self.target_layers: 60 | output.append(normalize_activation(x)) 61 | if len(output) == len(self.target_layers): 62 | break 63 | return output 64 | 65 | 66 | class SqueezeNet(BaseNet): 67 | def __init__(self): 68 | super(SqueezeNet, self).__init__() 69 | 70 | self.layers = models.squeezenet1_1(True).features 71 | self.target_layers = [2, 5, 8, 10, 11, 12, 13] 72 | self.n_channels_list = [64, 128, 256, 384, 384, 512, 512] 73 | 74 | self.set_requires_grad(False) 75 | 76 | 77 | class AlexNet(BaseNet): 78 | def __init__(self): 79 | super(AlexNet, self).__init__() 80 | 81 | self.layers = models.alexnet(True).features 82 | self.target_layers = [2, 5, 8, 10, 12] 83 | self.n_channels_list = [64, 192, 384, 256, 256] 84 | 85 | self.set_requires_grad(False) 86 | 87 | 88 | class VGG16(BaseNet): 89 | def __init__(self): 90 | super(VGG16, self).__init__() 91 | 92 | self.layers = models.vgg16(True).features 93 | self.target_layers = [4, 9, 16, 23, 30] 94 | self.n_channels_list = [64, 128, 256, 512, 512] 95 | 96 | self.set_requires_grad(False) -------------------------------------------------------------------------------- /criteria/lpips/utils.py: -------------------------------------------------------------------------------- 1 | from collections import OrderedDict 2 | 3 | import torch 4 | 5 | 6 | def normalize_activation(x, eps=1e-10): 7 | norm_factor = torch.sqrt(torch.sum(x ** 2, dim=1, keepdim=True)) 8 | return x / (norm_factor + eps) 9 | 10 | 11 | def get_state_dict(net_type: str = 'alex', version: str = '0.1'): 12 | # build url 13 | url = 'https://raw.githubusercontent.com/richzhang/PerceptualSimilarity/' \ 14 | + f'master/lpips/weights/v{version}/{net_type}.pth' 15 | 16 | # download 17 | old_state_dict = torch.hub.load_state_dict_from_url( 18 | url, progress=True, 19 | map_location=None if torch.cuda.is_available() else torch.device('cpu') 20 | ) 21 | 22 | # rename keys 23 | new_state_dict = OrderedDict() 24 | for key, val in old_state_dict.items(): 25 | new_key = key 26 | new_key = new_key.replace('lin', '') 27 | new_key = new_key.replace('model.', '') 28 | new_state_dict[new_key] = val 29 | 30 | return new_state_dict 31 | -------------------------------------------------------------------------------- /criteria/moco_loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | import torch.nn.functional as F 4 | from configs.paths_config import model_paths 5 | 6 | 7 | class MocoLoss(nn.Module): 8 | 9 | def __init__(self): 10 | super(MocoLoss, self).__init__() 11 | print("Loading MOCO model from path: {}".format(model_paths["moco"])) 12 | self.model = self.__load_model() 13 | self.model.cuda() 14 | self.model.eval() 15 | 16 | @staticmethod 17 | def __load_model(): 18 | import torchvision.models as models 19 | model = models.__dict__["resnet50"]() 20 | # freeze all layers but the last fc 21 | for name, param in model.named_parameters(): 22 | if name not in ['fc.weight', 'fc.bias']: 23 | param.requires_grad = False 24 | checkpoint = torch.load(model_paths['moco'], map_location="cpu") 25 | state_dict = checkpoint['state_dict'] 26 | # rename moco pre-trained keys 27 | for k in list(state_dict.keys()): 28 | # retain only encoder_q up to before the embedding layer 29 | if k.startswith('module.encoder_q') and not k.startswith('module.encoder_q.fc'): 30 | # remove prefix 31 | state_dict[k[len("module.encoder_q."):]] = state_dict[k] 32 | # delete renamed or unused k 33 | del state_dict[k] 34 | msg = model.load_state_dict(state_dict, strict=False) 35 | assert set(msg.missing_keys) == {"fc.weight", "fc.bias"} 36 | # remove output layer 37 | model = nn.Sequential(*list(model.children())[:-1]).cuda() 38 | return model 39 | 40 | def extract_feats(self, x): 41 | x = F.interpolate(x, size=224) 42 | x_feats = self.model(x) 43 | x_feats = nn.functional.normalize(x_feats, dim=1) 44 | x_feats = x_feats.squeeze() 45 | return x_feats 46 | 47 | def forward(self, y_hat, y, x): 48 | n_samples = x.shape[0] 49 | x_feats = self.extract_feats(x) 50 | y_feats = self.extract_feats(y) 51 | y_hat_feats = self.extract_feats(y_hat) 52 | y_feats = y_feats.detach() 53 | loss = 0 54 | sim_improvement = 0 55 | sim_logs = [] 56 | count = 0 57 | for i in range(n_samples): 58 | diff_target = y_hat_feats[i].dot(y_feats[i]) 59 | diff_input = y_hat_feats[i].dot(x_feats[i]) 60 | diff_views = y_feats[i].dot(x_feats[i]) 61 | sim_logs.append({'diff_target': float(diff_target), 62 | 'diff_input': float(diff_input), 63 | 'diff_views': float(diff_views)}) 64 | loss += 1 - diff_target 65 | sim_diff = float(diff_target) - float(diff_views) 66 | sim_improvement += sim_diff 67 | count += 1 68 | 69 | return loss / count, sim_improvement / count, sim_logs 70 | -------------------------------------------------------------------------------- /criteria/w_norm.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | 4 | 5 | class WNormLoss(nn.Module): 6 | 7 | def __init__(self, start_from_latent_avg=True): 8 | super(WNormLoss, self).__init__() 9 | self.start_from_latent_avg = start_from_latent_avg 10 | 11 | def forward(self, latent, latent_avg=None): 12 | if self.start_from_latent_avg: 13 | latent = latent - latent_avg 14 | return torch.sum(latent.norm(2, dim=(1, 2))) / latent.shape[0] 15 | -------------------------------------------------------------------------------- /docs/custom-color-edits.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/docs/custom-color-edits.png -------------------------------------------------------------------------------- /docs/prog-synthesis.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/docs/prog-synthesis.png -------------------------------------------------------------------------------- /docs/watercolor-synthesis.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/docs/watercolor-synthesis.png -------------------------------------------------------------------------------- /environment/paint2pix_env.yml: -------------------------------------------------------------------------------- 1 | name: intelli-paint 2 | channels: 3 | - conda-forge 4 | - defaults 5 | dependencies: 6 | - _libgcc_mutex=0.1=main 7 | - _openmp_mutex=4.5=1_gnu 8 | - anyio=2.2.0=py36h5fab9bb_0 9 | - async_generator=1.10=py_0 10 | - babel=2.9.1=pyh44b312d_0 11 | - backcall=0.2.0=pyhd3eb1b0_0 12 | - blas=1.0=mkl 13 | - brotlipy=0.7.0=py36he6145b8_1001 14 | - bzip2=1.0.8=h7b6447c_0 15 | - ca-certificates=2021.10.8=ha878542_0 16 | - certifi=2021.5.30=py36h5fab9bb_0 17 | - charset-normalizer=2.0.0=pyhd8ed1ab_0 18 | - contextvars=2.4=py_0 19 | - cryptography=3.1=py36h45558ae_0 20 | - cudatoolkit=11.0.221=h6bb024c_0 21 | - cudatoolkit-dev=11.0.3=py36h8f6f2f9_2 22 | - dataclasses=0.8=pyh4f3eec9_6 23 | - entrypoints=0.3=pyhd8ed1ab_1003 24 | - filelock=3.0.12=pyh9f0ad1d_0 25 | - freetype=2.10.2=h5ab3b9f_0 26 | - gdown=3.13.0=pyhd8ed1ab_0 27 | - gmp=6.1.2=h6c8ec71_1 28 | - gnutls=3.6.5=h71b1129_1002 29 | - htop=2.2.0=hf8c457e_1000 30 | - immutables=0.16=py36h7f8727e_0 31 | - intel-openmp=2021.2.0=h06a4308_610 32 | - ipython_genutils=0.2.0=pyhd3eb1b0_1 33 | - jedi=0.17.2=py36h5fab9bb_1 34 | - jpeg=9b=h024ee3a_2 35 | - json5=0.9.5=pyh9f0ad1d_0 36 | - jsonschema=3.2.0=pyhd8ed1ab_3 37 | - jupyter_client=6.1.12=pyhd3eb1b0_0 38 | - jupyter_core=4.7.1=py36h06a4308_0 39 | - jupyter_server=1.4.1=py36h06a4308_0 40 | - jupyterlab=3.2.5=pyhd8ed1ab_0 41 | - jupyterlab_server=2.8.2=pyhd8ed1ab_0 42 | - jupyterlab_widgets=1.0.2=pyhd8ed1ab_0 43 | - lame=3.100=h7b6447c_0 44 | - lcms2=2.12=h3be6417_0 45 | - ld_impl_linux-64=2.33.1=h53a641e_7 46 | - libblas=3.8.0=21_mkl 47 | - libcblas=3.8.0=21_mkl 48 | - libedit=3.1.20191231=h14c3975_1 49 | - libffi=3.3=he6710b0_2 50 | - libgcc-ng=9.1.0=hdf63c60_0 51 | - libgomp=9.3.0=h5101ec6_17 52 | - libiconv=1.16=h516909a_0 53 | - liblapack=3.8.0=21_mkl 54 | - libopus=1.3.1=h7b6447c_0 55 | - libpng=1.6.37=hbc83047_0 56 | - libsodium=1.0.18=h7b6447c_0 57 | - libstdcxx-ng=9.1.0=hdf63c60_0 58 | - libtiff=4.2.0=h85742a9_0 59 | - libuv=1.40.0=h7b6447c_0 60 | - libvpx=1.7.0=h439df22_0 61 | - libwebp-base=1.2.0=h27cfd23_0 62 | - lz4-c=1.9.3=h2531618_0 63 | - mistune=0.8.4=py36h1d69622_1002 64 | - mkl=2020.2=256 65 | - mkl-service=2.3.0=py36he8ac12f_0 66 | - mkl_fft=1.3.0=py36h54f3939_0 67 | - mkl_random=1.1.1=py36h0573a6f_0 68 | - nano=2.9.8=hddfc1eb_1001 69 | - nbclassic=0.2.6=pyhd3eb1b0_0 70 | - nbconvert=5.6.1=pyhd8ed1ab_2 71 | - ncurses=6.2=he6710b0_1 72 | - nest-asyncio=1.5.1=pyhd3eb1b0_0 73 | - nettle=3.4.1=hbb512f6_0 74 | - ninja=1.10.2=hff7bd54_1 75 | - numpy-base=1.19.2=py36hfa32c7d_0 76 | - olefile=0.46=py36_0 77 | - openh264=2.1.0=hd408876_0 78 | - openssl=1.1.1h=h516909a_0 79 | - parso=0.7.1=pyh9f0ad1d_0 80 | - pexpect=4.8.0=pyhd3eb1b0_3 81 | - pickleshare=0.7.5=pyhd3eb1b0_1003 82 | - pillow=7.2.0=py36hb39fc2d_0 83 | - pip=20.2.2=py36_0 84 | - prometheus_client=0.12.0=pyhd8ed1ab_0 85 | - prompt_toolkit=1.0.15=py_1 86 | - pycparser=2.20=pyh9f0ad1d_2 87 | - pyopenssl=19.1.0=py_1 88 | - pysocks=1.7.1=py36h5fab9bb_3 89 | - python=3.6.10=h7579374_2 90 | - python-dateutil=2.8.1=pyhd3eb1b0_0 91 | - python_abi=3.6=1_cp36m 92 | - readline=8.0=h7b6447c_0 93 | - screen=4.8.0=he28a2e2_0 94 | - setuptools=49.6.0=py36_0 95 | - sniffio=1.2.0=py36h5fab9bb_1 96 | - sqlite=3.32.3=h62c20be_0 97 | - tk=8.6.10=hbc83047_0 98 | - tqdm=4.61.1=pyhd8ed1ab_0 99 | - traitlets=4.3.3=py36h9f0ad1d_1 100 | - typing_extensions=3.10.0.2=pyh06a4308_0 101 | - unzip=6.0=h611a1e1_0 102 | - wcwidth=0.2.5=py_0 103 | - wheel=0.37.0=pyhd8ed1ab_1 104 | - widgetsnbextension=3.5.1=py36h5fab9bb_4 105 | - x264=1!157.20191217=h7b6447c_0 106 | - xz=5.2.5=h7b6447c_0 107 | - zeromq=4.3.4=h2531618_0 108 | - zlib=1.2.11=h7b6447c_3 109 | - zstd=1.4.9=haebb681_0 110 | - pip: 111 | - absl-py==0.10.0 112 | - altair==4.1.0 113 | - argon2-cffi==20.1.0 114 | - astor==0.8.1 115 | - astunparse==1.6.3 116 | - attrs==20.1.0 117 | - backports-entry-points-selectable==1.1.0 118 | - backports-zoneinfo==0.2.1 119 | - base58==2.1.0 120 | - bce-python-sdk==0.8.62 121 | - bleach==3.1.5 122 | - blinker==1.4 123 | - boto3==1.18.27 124 | - botocore==1.21.27 125 | - cached-property==1.5.2 126 | - cachetools==4.1.1 127 | - cffi==1.14.2 128 | - cfgv==3.3.0 129 | - chardet==3.0.4 130 | - clang==5.0 131 | - click==7.1.2 132 | - clip==1.0 133 | - clipboard==0.0.4 134 | - cloudpickle==1.2.2 135 | - colorama==0.4.4 136 | - colorlog==5.0.1 137 | - cycler==0.10.0 138 | - cython==0.29.21 139 | - decorator==4.4.2 140 | - defusedxml==0.6.0 141 | - deprecation==2.1.0 142 | - dill==0.3.4 143 | - distlib==0.3.2 144 | - dlib==19.22.1 145 | - dm-control==0.0.364896371 146 | - dm-env==1.5 147 | - dm-tree==0.1.7 148 | - dnnlib==0.0.1 149 | - dominate==2.6.0 150 | - easydict==1.9 151 | - efficientnet-pytorch==0.7.0 152 | - et-xmlfile==1.1.0 153 | - ffmpeg==1.4 154 | - ffmpeg-python==0.2.0 155 | - flake8==3.9.2 156 | - flask==2.0.1 157 | - flask-babel==2.0.0 158 | - flatbuffers==1.12 159 | - ftfy==6.0.3 160 | - future==0.18.2 161 | - fuzzywuzzy==0.18.0 162 | - gast==0.2.2 163 | - gevent==21.8.0 164 | - gitdb==4.0.7 165 | - gitpython==3.1.18 166 | - giturlparse==0.10.0 167 | - glcontext==2.3.5 168 | - glfw==1.12.0 169 | - google-api-core==2.3.1 170 | - google-api-python-client==2.33.0 171 | - google-auth==2.3.3 172 | - google-auth-httplib2==0.1.0 173 | - google-auth-oauthlib==0.4.1 174 | - google-pasta==0.2.0 175 | - googleapis-common-protos==1.54.0 176 | - greenlet==1.1.2 177 | - grpcio==1.39.0 178 | - gunicorn==20.1.0 179 | - gym==0.15.7 180 | - gym3==0.3.3 181 | - h5py==3.1.0 182 | - httplib2==0.20.2 183 | - hvac==0.11.2 184 | - identify==2.2.13 185 | - idna==2.8 186 | - imageio==2.9.0 187 | - imageio-ffmpeg==0.3.0 188 | - imagenetv2-pytorch==0.1 189 | - importlib-metadata==4.8.3 190 | - importlib-resources==5.2.2 191 | - ipykernel==4.6.1 192 | - ipython==5.5.0 193 | - ipywebrtc==0.6.0 194 | - ipywidgets==7.5.1 195 | - itsdangerous==2.0.1 196 | - jieba==0.42.1 197 | - jinja2==3.0.1 198 | - jmespath==0.10.0 199 | - joblib==0.16.0 200 | - jsonpatch==1.32 201 | - jsonpointer==2.3 202 | - jupyter==1.0.0 203 | - jupyter-client==6.1.6 204 | - jupyter-console==6.1.0 205 | - jupyter-core==4.6.3 206 | - keras==2.6.0 207 | - keras-applications==1.0.8 208 | - keras-preprocessing==1.1.2 209 | - kiwisolver==1.2.0 210 | - kornia==0.5.8 211 | - labmaze==1.0.5 212 | - lpips==0.1.4 213 | - lxml==4.8.0 214 | - markdown==3.2.2 215 | - markupsafe==2.0.1 216 | - matplotlib==3.3.1 217 | - mccabe==0.6.1 218 | - mock==4.0.3 219 | - moderngl==5.6.4 220 | - multiprocess==0.70.12.2 221 | - munch==2.5.0 222 | - nbformat==5.0.7 223 | - networkx==2.5.1 224 | - neuralgym==0.0.1 225 | - nodeenv==1.6.0 226 | - notebook==5.2.2 227 | - numpy==1.19.5 228 | - nvidia-htop==1.0.5 229 | - oauth2client==4.1.3 230 | - oauthlib==3.1.0 231 | - opencv-python==4.4.0.42 232 | - openpyxl==3.0.9 233 | - opt-einsum==3.3.0 234 | - packaging==20.4 235 | - paddle2onnx==0.7 236 | - paddlehub==2.1.0 237 | - paddlenlp==2.0.7 238 | - paddlepaddle==2.1.2 239 | - paddlepaddle-gpu==2.2.1 240 | - pandas==0.24.2 241 | - pandocfilters==1.4.2 242 | - platformdirs==2.2.0 243 | - portpicker==1.2.0 244 | - pre-commit==2.14.0 245 | - procgen==0.10.4 246 | - prometheus-client==0.8.0 247 | - prompt-toolkit==1.0.18 248 | - protobuf==3.19.4 249 | - ptyprocess==0.6.0 250 | - pyarrow==5.0.0 251 | - pyasn1==0.4.8 252 | - pyasn1-modules==0.2.8 253 | - pycodestyle==2.7.0 254 | - pycryptodome==3.10.1 255 | - pydeck==0.6.2 256 | - pydrive==1.3.1 257 | - pydrive2==1.10.0 258 | - pyflakes==2.3.1 259 | - pyglet==1.5.0 260 | - pygments==2.6.1 261 | - pyopengl==3.1.6 262 | - pyparsing==2.4.7 263 | - pyperclip==1.8.2 264 | - pyrsistent==0.16.0 265 | - python-levenshtein==0.12.2 266 | - pytorch-fid==0.2.1 267 | - pytz==2020.1 268 | - pytz-deprecation-shim==0.1.0.post0 269 | - pywavelets==1.1.1 270 | - pyyaml==5.4.1 271 | - pyzmq==20.0.0 272 | - qtconsole==4.7.6 273 | - qtpy==1.9.0 274 | - rarfile==4.0 275 | - regex==2021.7.6 276 | - requests==2.21.0 277 | - requests-oauthlib==1.3.0 278 | - rsa==4.6 279 | - ruamel-yaml==0.17.4 280 | - ruamel-yaml-clib==0.2.6 281 | - s3transfer==0.5.0 282 | - scikit-image==0.17.2 283 | - scikit-learn==0.23.2 284 | - scipy==1.2.0 285 | - seaborn==0.10.1 286 | - send2trash==1.5.0 287 | - seqeval==1.2.2 288 | - shellcheck-py==0.7.2.1 289 | - simplegeneric==0.8.1 290 | - six==1.12.0 291 | - smmap==4.0.0 292 | - streamlit==1.0.0 293 | - streamlit-drawable-canvas==0.8.0 294 | - tabulate==0.8.9 295 | - tensorboard==1.15.0 296 | - tensorboard-data-server==0.6.1 297 | - tensorboard-plugin-wit==1.7.0 298 | - tensorboardx==2.1 299 | - tensorflow-estimator==1.15.1 300 | - tensorflow-gpu==1.15.0 301 | - termcolor==1.1.0 302 | - terminado==0.8.3 303 | - testpath==0.4.4 304 | - threadpoolctl==2.1.0 305 | - tifffile==2020.9.3 306 | - timm==0.4.12 307 | - toml==0.10.2 308 | - toolz==0.11.1 309 | - torch==1.7.1+cu110 310 | - torchaudio==0.7.2 311 | - torchfile==0.1.0 312 | - torchgeometry==0.1.2 313 | - torchvision==0.8.2+cu110 314 | - tornado==4.5.3 315 | - typing-extensions==3.7.4.3 316 | - tzdata==2021.4 317 | - tzlocal==4.0.1 318 | - uritemplate==4.1.1 319 | - urllib3==1.25.10 320 | - validators==0.18.2 321 | - virtualenv==20.7.2 322 | - visdom==0.1.8.9 323 | - visualdl==2.2.0 324 | - watchdog==2.1.6 325 | - webencodings==0.5.1 326 | - websocket-client==1.3.1 327 | - werkzeug==2.0.1 328 | - wrapt==1.12.1 329 | - zipp==3.1.0 330 | - zope-event==4.5.0 331 | - zope-interface==5.4.0 332 | prefix: /home/jsingh/anaconda3/envs/intelli-paint 333 | -------------------------------------------------------------------------------- /input-images/143.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/input-images/143.jpg -------------------------------------------------------------------------------- /input-images/20.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/input-images/20.jpg -------------------------------------------------------------------------------- /input-images/214.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/input-images/214.jpg -------------------------------------------------------------------------------- /input-images/262.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/input-images/262.jpg -------------------------------------------------------------------------------- /input-images/823.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/input-images/823.jpg -------------------------------------------------------------------------------- /licenses/LICENSE_S-aiueo32: -------------------------------------------------------------------------------- 1 | BSD 2-Clause License 2 | 3 | Copyright (c) 2020, Sou Uchida 4 | All rights reserved. 5 | 6 | Redistribution and use in source and binary forms, with or without 7 | modification, are permitted provided that the following conditions are met: 8 | 9 | 1. Redistributions of source code must retain the above copyright notice, this 10 | list of conditions and the following disclaimer. 11 | 12 | 2. Redistributions in binary form must reproduce the above copyright notice, 13 | this list of conditions and the following disclaimer in the documentation 14 | and/or other materials provided with the distribution. 15 | 16 | THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 17 | AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 | IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 19 | DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE 20 | FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 | DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 22 | SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 23 | CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 24 | OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 25 | OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. -------------------------------------------------------------------------------- /licenses/LICENSE_TreB1eN: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2018 TreB1eN 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. -------------------------------------------------------------------------------- /licenses/LICENSE_alaluf: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2022 Yuval Alaluf 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /licenses/LICENSE_eladrich: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2020 Elad Richardson, Yuval Alaluf 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. -------------------------------------------------------------------------------- /licenses/LICENSE_lessw2020: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. -------------------------------------------------------------------------------- /licenses/LICENSE_omertov: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2021 omertov 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. -------------------------------------------------------------------------------- /licenses/LICENSE_rosinality: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2019 Kim Seonghyeon 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. -------------------------------------------------------------------------------- /models/.ipynb_checkpoints/e4e-checkpoint.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file defines the core research contribution 3 | """ 4 | import math 5 | import torch 6 | from torch import nn 7 | 8 | from models.stylegan2.model import Generator 9 | from configs.paths_config import model_paths 10 | from models.encoders import restyle_e4e_encoders 11 | from utils.model_utils import RESNET_MAPPING 12 | 13 | from models.stylegan2.model import EqualLinear 14 | 15 | class e4e(nn.Module): 16 | 17 | def __init__(self, opts): 18 | super(e4e, self).__init__() 19 | self.set_opts(opts) 20 | self.n_styles = int(math.log(self.opts.output_size, 2)) * 2 - 2 21 | # Define architecture 22 | self.encoder = self.set_encoder() 23 | self.decoder = Generator(self.opts.output_size, 512, 8, channel_multiplier=2) 24 | self.face_pool = torch.nn.AdaptiveAvgPool2d((256, 256)) 25 | # Load weights if needed 26 | self.load_weights() 27 | 28 | def set_encoder(self): 29 | if self.opts.encoder_type == 'ProgressiveBackboneEncoder': 30 | encoder = restyle_e4e_encoders.ProgressiveBackboneEncoder(50, 'ir_se', self.n_styles, self.opts) 31 | elif self.opts.encoder_type == 'ResNetProgressiveBackboneEncoder': 32 | encoder = restyle_e4e_encoders.ResNetProgressiveBackboneEncoder(self.n_styles, self.opts) 33 | else: 34 | raise Exception(f'{self.opts.encoder_type} is not a valid encoders') 35 | return encoder 36 | 37 | def load_weights(self): 38 | if self.opts.checkpoint_path is not None: 39 | print(f'Loading ReStyle e4e from checkpoint: {self.opts.checkpoint_path}') 40 | ckpt = torch.load(self.opts.checkpoint_path, map_location='cpu') 41 | self.encoder.load_state_dict(self.__get_keys(ckpt, 'encoder'), strict=False) 42 | self.decoder.load_state_dict(self.__get_keys(ckpt, 'decoder'), strict=True) 43 | self.__load_latent_avg(ckpt) 44 | else: 45 | encoder_ckpt = self.__get_encoder_checkpoint() 46 | self.encoder.load_state_dict(encoder_ckpt, strict=False) 47 | print(f'Loading decoder weights from pretrained path: {self.opts.stylegan_weights}') 48 | ckpt = torch.load(self.opts.stylegan_weights) 49 | self.decoder.load_state_dict(ckpt['g_ema'], strict=True) 50 | self.__load_latent_avg(ckpt, repeat=self.n_styles) 51 | 52 | def forward(self, x, target_id_feat=None, latent=None, resize=True, latent_mask=None, input_code=False, randomize_noise=True, 53 | inject_latent=None, return_latents=False, alpha=None, average_code=False, input_is_full=False): 54 | if input_code: 55 | codes = x 56 | else: 57 | codes = self.encoder(x,target_id_feat) 58 | # residual step 59 | if x.shape[1] == self.opts.input_nc and latent is not None: 60 | # learn error with respect to previous iteration 61 | codes = codes + latent 62 | else: 63 | # first iteration is with respect to the avg latent code 64 | codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1) 65 | 66 | if latent_mask is not None: 67 | for i in latent_mask: 68 | if inject_latent is not None: 69 | if alpha is not None: 70 | codes[:, i] = alpha * inject_latent[:, i] + (1 - alpha) * codes[:, i] 71 | else: 72 | codes[:, i] = inject_latent[:, i] 73 | else: 74 | codes[:, i] = 0 75 | 76 | if average_code: 77 | input_is_latent = True 78 | else: 79 | input_is_latent = (not input_code) or (input_is_full) 80 | 81 | images, result_latent = self.decoder([codes], 82 | input_is_latent=input_is_latent, 83 | randomize_noise=randomize_noise, 84 | return_latents=return_latents) 85 | 86 | if resize: 87 | images = self.face_pool(images) 88 | 89 | if return_latents: 90 | return images, result_latent 91 | else: 92 | return images 93 | 94 | def set_opts(self, opts): 95 | self.opts = opts 96 | 97 | def __load_latent_avg(self, ckpt, repeat=None): 98 | if 'latent_avg' in ckpt: 99 | self.latent_avg = ckpt['latent_avg'].to(self.opts.device) 100 | if repeat is not None: 101 | self.latent_avg = self.latent_avg.repeat(repeat, 1) 102 | else: 103 | self.latent_avg = None 104 | 105 | def __get_encoder_checkpoint(self): 106 | if "ffhq" in self.opts.dataset_type: 107 | print('Loading encoders weights from irse50!') 108 | encoder_ckpt = torch.load(model_paths['ir_se50']) 109 | # Transfer the RGB input of the irse50 network to the first 3 input channels of pSp's encoder 110 | if self.opts.input_nc != 3: 111 | shape = encoder_ckpt['input_layer.0.weight'].shape 112 | altered_input_layer = torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 113 | altered_input_layer[:, :3, :, :] = encoder_ckpt['input_layer.0.weight'] 114 | encoder_ckpt['input_layer.0.weight'] = altered_input_layer 115 | return encoder_ckpt 116 | else: 117 | print('Loading encoders weights from resnet34!') 118 | encoder_ckpt = torch.load(model_paths['resnet34']) 119 | # Transfer the RGB input of the resnet34 network to the first 3 input channels of pSp's encoder 120 | if self.opts.input_nc != 3: 121 | shape = encoder_ckpt['conv1.weight'].shape 122 | altered_input_layer = torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 123 | altered_input_layer[:, :3, :, :] = encoder_ckpt['conv1.weight'] 124 | encoder_ckpt['conv1.weight'] = altered_input_layer 125 | mapped_encoder_ckpt = dict(encoder_ckpt) 126 | for p, v in encoder_ckpt.items(): 127 | for original_name, psp_name in RESNET_MAPPING.items(): 128 | if original_name in p: 129 | mapped_encoder_ckpt[p.replace(original_name, psp_name)] = v 130 | mapped_encoder_ckpt.pop(p) 131 | return encoder_ckpt 132 | 133 | @staticmethod 134 | def __get_keys(d, name): 135 | if 'state_dict' in d: 136 | d = d['state_dict'] 137 | d_filt = {k[len(name) + 1:]: v for k, v in d.items() if k[:len(name)] == name} 138 | return d_filt 139 | -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/__init__.py -------------------------------------------------------------------------------- /models/e4e.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file defines the core research contribution 3 | """ 4 | import math 5 | import torch 6 | from torch import nn 7 | 8 | from models.stylegan2.model import Generator 9 | from configs.paths_config import model_paths 10 | from models.encoders import restyle_e4e_encoders 11 | from utils.model_utils import RESNET_MAPPING 12 | 13 | from models.stylegan2.model import EqualLinear 14 | import streamlit as st 15 | 16 | class e4e(nn.Module): 17 | 18 | def __init__(self, opts): 19 | super(e4e, self).__init__() 20 | self.set_opts(opts) 21 | self.n_styles = int(math.log(self.opts.output_size, 2)) * 2 - 2 22 | # Define architecture 23 | self.encoder = self.set_encoder() 24 | self.decoder = Generator(self.opts.output_size, 512, 8, channel_multiplier=2) 25 | self.face_pool = torch.nn.AdaptiveAvgPool2d((256, 256)) 26 | # Load weights if needed 27 | self.load_weights() 28 | 29 | def set_encoder(self): 30 | if self.opts.encoder_type == 'ProgressiveBackboneEncoder': 31 | encoder = restyle_e4e_encoders.ProgressiveBackboneEncoder(50, 'ir_se', self.n_styles, self.opts) 32 | elif self.opts.encoder_type == 'ResNetProgressiveBackboneEncoder': 33 | encoder = restyle_e4e_encoders.ResNetProgressiveBackboneEncoder(self.n_styles, self.opts) 34 | else: 35 | raise Exception(f'{self.opts.encoder_type} is not a valid encoders') 36 | return encoder 37 | 38 | def load_weights(self): 39 | if self.opts.checkpoint_path is not None: 40 | print(f'Loading ReStyle e4e from checkpoint: {self.opts.checkpoint_path}') 41 | ckpt = torch.load(self.opts.checkpoint_path, map_location='cpu') 42 | encoder_ckpt = self.__get_keys(ckpt, 'encoder') 43 | 44 | # if self.opts.input_nc != 6: 45 | # print ("here ...") 46 | # shape = encoder_ckpt['input_layer.0.weight'].shape 47 | # altered_input_layer1 = 0.*torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 48 | # altered_input_layer1[:, :6, :, :] = encoder_ckpt['input_layer.0.weight'] 49 | 50 | # altered_input_layer2 = 0.*torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 51 | # altered_input_layer2[:, 6:, :, :] = encoder_ckpt['input_layer.0.weight'] 52 | # encoder_ckpt['input_layer_updated_canvas.0.weight'] = altered_input_layer2 53 | # encoder_ckpt['input_layer.0.weight'] = altered_input_layer1 54 | 55 | self.encoder.load_state_dict(encoder_ckpt, strict=False) 56 | # self.encoder.load_state_dict(self.__get_keys(ckpt, 'encoder'), strict=False) 57 | self.decoder.load_state_dict(self.__get_keys(ckpt, 'decoder'), strict=True) 58 | self.__load_latent_avg(ckpt) 59 | else: 60 | encoder_ckpt = self.__get_encoder_checkpoint() 61 | self.encoder.load_state_dict(encoder_ckpt, strict=False) 62 | print(f'Loading decoder weights from pretrained path: {self.opts.stylegan_weights}') 63 | ckpt = torch.load(self.opts.stylegan_weights) 64 | self.decoder.load_state_dict(ckpt['g_ema'], strict=True) 65 | self.__load_latent_avg(ckpt, repeat=self.n_styles) 66 | 67 | def forward(self, x, target_id_feat=None, latent=None, resize=True, latent_mask=None, input_code=False, randomize_noise=True, 68 | inject_latent=None, return_latents=False, alpha=None, average_code=False, input_is_full=False, get_multiple_codes=False): 69 | 70 | if input_code: 71 | codes = x 72 | else: 73 | if get_multiple_codes: 74 | codes_canvas0, codes_canvas1 = self.encoder(x,target_id_feat,get_multiple_codes) 75 | codes = torch.cat([codes_canvas0, codes_canvas1],axis=0) 76 | else: 77 | codes = self.encoder(x,target_id_feat) 78 | # print (codes.shape,self.latent_avg.shape,self.latent_avg.repeat(codes.shape[0], 1, 1).shape) 79 | # residual step 80 | if x.shape[1] == self.opts.input_nc and latent is not None: 81 | # learn error with respect to previous iteration 82 | codes = codes + latent 83 | else: 84 | # first iteration is with respect to the avg latent code 85 | codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1) 86 | 87 | if latent_mask is not None: 88 | for i in latent_mask: 89 | if inject_latent is not None: 90 | if alpha is not None: 91 | codes[:, i] = alpha * inject_latent[:, i] + (1 - alpha) * codes[:, i] 92 | else: 93 | codes[:, i] = inject_latent[:, i] 94 | else: 95 | codes[:, i] = 0 96 | 97 | if average_code: 98 | input_is_latent = True 99 | else: 100 | input_is_latent = (not input_code) or (input_is_full) 101 | 102 | images, result_latent = self.decoder([codes], 103 | input_is_latent=input_is_latent, 104 | randomize_noise=randomize_noise, 105 | return_latents=return_latents) 106 | 107 | if resize: 108 | images = self.face_pool(images) 109 | 110 | if return_latents: 111 | return images, result_latent 112 | else: 113 | return images 114 | 115 | def set_opts(self, opts): 116 | self.opts = opts 117 | 118 | def __load_latent_avg(self, ckpt, repeat=None): 119 | if 'latent_avg' in ckpt: 120 | self.latent_avg = ckpt['latent_avg'].to(self.opts.device) 121 | if repeat is not None: 122 | self.latent_avg = self.latent_avg.repeat(repeat, 1) 123 | else: 124 | self.latent_avg = None 125 | 126 | def __get_encoder_checkpoint(self): 127 | if "ffhq" in self.opts.dataset_type: 128 | print('Loading encoders weights from irse50!') 129 | encoder_ckpt = torch.load(model_paths['ir_se50']) 130 | # Transfer the RGB input of the irse50 network to the first 3 input channels of pSp's encoder 131 | if self.opts.input_nc != 3: 132 | shape = encoder_ckpt['input_layer.0.weight'].shape 133 | altered_input_layer = torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 134 | altered_input_layer[:, :3, :, :] = encoder_ckpt['input_layer.0.weight'] 135 | encoder_ckpt['input_layer.0.weight'] = altered_input_layer 136 | return encoder_ckpt 137 | else: 138 | print('Loading encoders weights from resnet34!') 139 | encoder_ckpt = torch.load(model_paths['resnet34']) 140 | # Transfer the RGB input of the resnet34 network to the first 3 input channels of pSp's encoder 141 | if self.opts.input_nc != 3: 142 | shape = encoder_ckpt['conv1.weight'].shape 143 | altered_input_layer = torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 144 | altered_input_layer[:, :3, :, :] = encoder_ckpt['conv1.weight'] 145 | encoder_ckpt['conv1.weight'] = altered_input_layer 146 | mapped_encoder_ckpt = dict(encoder_ckpt) 147 | for p, v in encoder_ckpt.items(): 148 | for original_name, psp_name in RESNET_MAPPING.items(): 149 | if original_name in p: 150 | mapped_encoder_ckpt[p.replace(original_name, psp_name)] = v 151 | mapped_encoder_ckpt.pop(p) 152 | return encoder_ckpt 153 | 154 | @staticmethod 155 | def __get_keys(d, name): 156 | if 'state_dict' in d: 157 | d = d['state_dict'] 158 | d_filt = {k[len(name) + 1:]: v for k, v in d.items() if k[:len(name)] == name} 159 | return d_filt 160 | -------------------------------------------------------------------------------- /models/e4e_modules/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/e4e_modules/__init__.py -------------------------------------------------------------------------------- /models/e4e_modules/discriminator.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | 4 | class LatentCodesDiscriminator(nn.Module): 5 | def __init__(self, style_dim, n_mlp): 6 | super().__init__() 7 | 8 | self.style_dim = style_dim 9 | 10 | layers = [] 11 | for i in range(n_mlp-1): 12 | layers.append( 13 | nn.Linear(style_dim, style_dim) 14 | ) 15 | layers.append(nn.LeakyReLU(0.2)) 16 | layers.append(nn.Linear(512, 1)) 17 | self.mlp = nn.Sequential(*layers) 18 | 19 | def forward(self, w): 20 | return self.mlp(w) 21 | -------------------------------------------------------------------------------- /models/e4e_modules/latent_codes_pool.py: -------------------------------------------------------------------------------- 1 | import random 2 | import torch 3 | 4 | 5 | class LatentCodesPool: 6 | """This class implements latent codes buffer that stores previously generated w latent codes. 7 | This buffer enables us to update discriminators using a history of generated w's 8 | rather than the ones produced by the latest encoder. 9 | """ 10 | 11 | def __init__(self, pool_size): 12 | """Initialize the ImagePool class 13 | Parameters: 14 | pool_size (int) -- the size of image buffer, if pool_size=0, no buffer will be created 15 | """ 16 | self.pool_size = pool_size 17 | if self.pool_size > 0: # create an empty pool 18 | self.num_ws = 0 19 | self.ws = [] 20 | 21 | def query(self, ws): 22 | """Return w's from the pool. 23 | Parameters: 24 | ws: the latest generated w's from the generator 25 | Returns w's from the buffer. 26 | By 50/100, the buffer will return input w's. 27 | By 50/100, the buffer will return w's previously stored in the buffer, 28 | and insert the current w's to the buffer. 29 | """ 30 | if self.pool_size == 0: # if the buffer size is 0, do nothing 31 | return ws 32 | return_ws = [] 33 | for w in ws: # ws.shape: (batch, 512) or (batch, n_latent, 512) 34 | # w = torch.unsqueeze(image.data, 0) 35 | if w.ndim == 2: 36 | i = random.randint(0, len(w) - 1) # apply a random latent index as a candidate 37 | w = w[i] 38 | self.handle_w(w, return_ws) 39 | return_ws = torch.stack(return_ws, 0) # collect all the images and return 40 | return return_ws 41 | 42 | def handle_w(self, w, return_ws): 43 | if self.num_ws < self.pool_size: # if the buffer is not full; keep inserting current codes to the buffer 44 | self.num_ws = self.num_ws + 1 45 | self.ws.append(w) 46 | return_ws.append(w) 47 | else: 48 | p = random.uniform(0, 1) 49 | if p > 0.5: # by 50% chance, the buffer will return a previously stored latent code, and insert the current code into the buffer 50 | random_id = random.randint(0, self.pool_size - 1) # randint is inclusive 51 | tmp = self.ws[random_id].clone() 52 | self.ws[random_id] = w 53 | return_ws.append(tmp) 54 | else: # by another 50% chance, the buffer will return the current image 55 | return_ws.append(w) 56 | -------------------------------------------------------------------------------- /models/encoders/.ipynb_checkpoints/map2style-checkpoint.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from torch import nn 3 | from torch.nn import Conv2d, Module 4 | 5 | from models.stylegan2.model import EqualLinear 6 | 7 | 8 | class GradualStyleBlock(Module): 9 | def __init__(self, in_c, out_c, spatial): 10 | super(GradualStyleBlock, self).__init__() 11 | self.out_c = out_c 12 | self.spatial = spatial 13 | num_pools = int(np.log2(spatial)) 14 | modules = [] 15 | modules += [Conv2d(in_c, out_c, kernel_size=3, stride=2, padding=1), 16 | nn.LeakyReLU()] 17 | for i in range(num_pools - 1): 18 | modules += [ 19 | Conv2d(out_c, out_c, kernel_size=3, stride=2, padding=1), 20 | nn.LeakyReLU() 21 | ] 22 | self.convs = nn.Sequential(*modules) 23 | self.linear = EqualLinear(out_c, out_c, lr_mul=1) 24 | 25 | def forward(self, x): 26 | x = self.convs(x) 27 | x = x.view(-1, self.out_c) 28 | x = self.linear(x) 29 | return x 30 | -------------------------------------------------------------------------------- /models/encoders/.ipynb_checkpoints/restyle_e4e_encoders-checkpoint.py: -------------------------------------------------------------------------------- 1 | from enum import Enum 2 | from torch import nn 3 | from torch.nn import Conv2d, BatchNorm2d, PReLU, Sequential, Module 4 | from torchvision.models import resnet34 5 | 6 | from models.encoders.helpers import get_blocks, bottleneck_IR, bottleneck_IR_SE 7 | from models.encoders.map2style import GradualStyleBlock 8 | 9 | from models.stylegan2.model import EqualLinear 10 | 11 | class ProgressiveStage(Enum): 12 | WTraining = 0 13 | Delta1Training = 1 14 | Delta2Training = 2 15 | Delta3Training = 3 16 | Delta4Training = 4 17 | Delta5Training = 5 18 | Delta6Training = 6 19 | Delta7Training = 7 20 | Delta8Training = 8 21 | Delta9Training = 9 22 | Delta10Training = 10 23 | Delta11Training = 11 24 | Delta12Training = 12 25 | Delta13Training = 13 26 | Delta14Training = 14 27 | Delta15Training = 15 28 | Delta16Training = 16 29 | Delta17Training = 17 30 | Inference = 18 31 | 32 | 33 | class ProgressiveBackboneEncoder(Module): 34 | """ 35 | The simpler backbone architecture used by ReStyle where all style vectors are extracted from the final 16x16 feature 36 | map of the encoder. This classes uses the simplified architecture applied over an ResNet IRSE50 backbone with the 37 | progressive training scheme from e4e_modules. 38 | Note this class is designed to be used for the human facial domain. 39 | """ 40 | def __init__(self, num_layers, mode='ir', n_styles=18, opts=None): 41 | super(ProgressiveBackboneEncoder, self).__init__() 42 | assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152' 43 | assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se' 44 | blocks = get_blocks(num_layers) 45 | if mode == 'ir': 46 | unit_module = bottleneck_IR 47 | elif mode == 'ir_se': 48 | unit_module = bottleneck_IR_SE 49 | 50 | self.input_layer = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False), 51 | BatchNorm2d(64), 52 | PReLU(64)) 53 | modules = [] 54 | for block in blocks: 55 | for bottleneck in block: 56 | modules.append(unit_module(bottleneck.in_channel, 57 | bottleneck.depth, 58 | bottleneck.stride)) 59 | self.body = Sequential(*modules) 60 | 61 | self.styles = nn.ModuleList() 62 | self.style_count = n_styles 63 | for i in range(self.style_count): 64 | style = GradualStyleBlock(512, 512, 16) 65 | self.styles.append(style) 66 | self.progressive_stage = ProgressiveStage.Inference 67 | 68 | self.id_layers = nn.ModuleList() 69 | for i in range(self.style_count): 70 | id_layer = EqualLinear(512, 512, lr_mul=1) 71 | self.id_layers.append(id_layer) 72 | 73 | def get_deltas_starting_dimensions(self): 74 | ''' Get a list of the initial dimension of every delta from which it is applied ''' 75 | return list(range(self.style_count)) # Each dimension has a delta applied to 76 | 77 | def set_progressive_stage(self, new_stage: ProgressiveStage): 78 | # In this encoder we train all the pyramid (At least as a first stage experiment 79 | self.progressive_stage = new_stage 80 | print('Changed progressive stage to: ', new_stage) 81 | 82 | def forward(self, x, target_id_feat=None): 83 | x = self.input_layer(x) 84 | x = self.body(x) 85 | 86 | # get initial w0 from first map2style layer 87 | w0 = self.styles[0](x) 88 | if target_id_feat is not None: 89 | w0 += self.id_layers[0](target_id_feat) 90 | w = w0.repeat(self.style_count, 1, 1).permute(1, 0, 2) 91 | 92 | # learn the deltas up to the current stage 93 | stage = self.progressive_stage.value 94 | for i in range(1, min(stage + 1, self.style_count)): 95 | delta_i = self.styles[i](x) 96 | if target_id_feat is not None: 97 | delta_i += self.id_layers[i](target_id_feat) 98 | w[:, i] += delta_i 99 | return w 100 | 101 | 102 | class ResNetProgressiveBackboneEncoder(Module): 103 | """ 104 | The simpler backbone architecture used by ReStyle where all style vectors are extracted from the final 16x16 feature 105 | map of the encoder. This classes uses the simplified architecture applied over an ResNet34 backbone with the 106 | progressive training scheme from e4e_modules. 107 | """ 108 | def __init__(self, n_styles=18, opts=None): 109 | super(ResNetProgressiveBackboneEncoder, self).__init__() 110 | 111 | self.conv1 = nn.Conv2d(opts.input_nc, 64, kernel_size=7, stride=2, padding=3, bias=False) 112 | self.bn1 = BatchNorm2d(64) 113 | self.relu = PReLU(64) 114 | 115 | resnet_basenet = resnet34(pretrained=True) 116 | blocks = [ 117 | resnet_basenet.layer1, 118 | resnet_basenet.layer2, 119 | resnet_basenet.layer3, 120 | resnet_basenet.layer4 121 | ] 122 | modules = [] 123 | for block in blocks: 124 | for bottleneck in block: 125 | modules.append(bottleneck) 126 | self.body = Sequential(*modules) 127 | 128 | self.styles = nn.ModuleList() 129 | self.style_count = n_styles 130 | for i in range(self.style_count): 131 | style = GradualStyleBlock(512, 512, 16) 132 | self.styles.append(style) 133 | self.progressive_stage = ProgressiveStage.Inference 134 | 135 | def get_deltas_starting_dimensions(self): 136 | ''' Get a list of the initial dimension of every delta from which it is applied ''' 137 | return list(range(self.style_count)) # Each dimension has a delta applied to 138 | 139 | def set_progressive_stage(self, new_stage: ProgressiveStage): 140 | # In this encoder we train all the pyramid (At least as a first stage experiment 141 | self.progressive_stage = new_stage 142 | print('Changed progressive stage to: ', new_stage) 143 | 144 | def forward(self, x): 145 | x = self.conv1(x) 146 | x = self.bn1(x) 147 | x = self.relu(x) 148 | x = self.body(x) 149 | 150 | # get initial w0 from first map2style layer 151 | w0 = self.styles[0](x) 152 | w = w0.repeat(self.style_count, 1, 1).permute(1, 0, 2) 153 | 154 | # learn the deltas up to the current stage 155 | stage = self.progressive_stage.value 156 | for i in range(1, min(stage + 1, self.style_count)): 157 | delta_i = self.styles[i](x) 158 | w[:, i] += delta_i 159 | return w 160 | -------------------------------------------------------------------------------- /models/encoders/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/encoders/__init__.py -------------------------------------------------------------------------------- /models/encoders/fpn_encoders.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.functional as F 3 | from torch import nn 4 | from torch.nn import Conv2d, BatchNorm2d, PReLU, Sequential, Module 5 | from torchvision.models.resnet import resnet34 6 | 7 | from models.encoders.helpers import get_blocks, bottleneck_IR, bottleneck_IR_SE 8 | from models.encoders.map2style import GradualStyleBlock 9 | 10 | 11 | class GradualStyleEncoder(Module): 12 | """ 13 | Original encoder architecture from pixel2style2pixel. This classes uses an FPN-based architecture applied over 14 | an ResNet IRSE-50 backbone. 15 | Note this class is designed to be used for the human facial domain. 16 | """ 17 | def __init__(self, num_layers, mode='ir', n_styles=18, opts=None): 18 | super(GradualStyleEncoder, self).__init__() 19 | assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152' 20 | assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se' 21 | blocks = get_blocks(num_layers) 22 | if mode == 'ir': 23 | unit_module = bottleneck_IR 24 | elif mode == 'ir_se': 25 | unit_module = bottleneck_IR_SE 26 | self.input_layer = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False), 27 | BatchNorm2d(64), 28 | PReLU(64)) 29 | modules = [] 30 | for block in blocks: 31 | for bottleneck in block: 32 | modules.append(unit_module(bottleneck.in_channel, 33 | bottleneck.depth, 34 | bottleneck.stride)) 35 | self.body = Sequential(*modules) 36 | 37 | self.styles = nn.ModuleList() 38 | self.style_count = n_styles 39 | self.coarse_ind = 3 40 | self.middle_ind = 7 41 | for i in range(self.style_count): 42 | if i < self.coarse_ind: 43 | style = GradualStyleBlock(512, 512, 16) 44 | elif i < self.middle_ind: 45 | style = GradualStyleBlock(512, 512, 32) 46 | else: 47 | style = GradualStyleBlock(512, 512, 64) 48 | self.styles.append(style) 49 | self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0) 50 | self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0) 51 | 52 | def _upsample_add(self, x, y): 53 | _, _, H, W = y.size() 54 | return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y 55 | 56 | def forward(self, x): 57 | x = self.input_layer(x) 58 | 59 | latents = [] 60 | modulelist = list(self.body._modules.values()) 61 | for i, l in enumerate(modulelist): 62 | x = l(x) 63 | if i == 6: 64 | c1 = x 65 | elif i == 20: 66 | c2 = x 67 | elif i == 23: 68 | c3 = x 69 | 70 | for j in range(self.coarse_ind): 71 | latents.append(self.styles[j](c3)) 72 | 73 | p2 = self._upsample_add(c3, self.latlayer1(c2)) 74 | for j in range(self.coarse_ind, self.middle_ind): 75 | latents.append(self.styles[j](p2)) 76 | 77 | p1 = self._upsample_add(p2, self.latlayer2(c1)) 78 | for j in range(self.middle_ind, self.style_count): 79 | latents.append(self.styles[j](p1)) 80 | 81 | out = torch.stack(latents, dim=1) 82 | return out 83 | 84 | 85 | class ResNetGradualStyleEncoder(Module): 86 | """ 87 | Original encoder architecture from pixel2style2pixel. This classes uses an FPN-based architecture applied over 88 | an ResNet34 backbone. 89 | """ 90 | def __init__(self, n_styles=18, opts=None): 91 | super(ResNetGradualStyleEncoder, self).__init__() 92 | 93 | self.conv1 = nn.Conv2d(opts.input_nc, 64, kernel_size=7, stride=2, padding=3, bias=False) 94 | self.bn1 = BatchNorm2d(64) 95 | self.relu = PReLU(64) 96 | 97 | resnet_basenet = resnet34(pretrained=True) 98 | blocks = [ 99 | resnet_basenet.layer1, 100 | resnet_basenet.layer2, 101 | resnet_basenet.layer3, 102 | resnet_basenet.layer4 103 | ] 104 | 105 | modules = [] 106 | for block in blocks: 107 | for bottleneck in block: 108 | modules.append(bottleneck) 109 | 110 | self.body = Sequential(*modules) 111 | 112 | self.styles = nn.ModuleList() 113 | self.style_count = n_styles 114 | self.coarse_ind = 3 115 | self.middle_ind = 7 116 | for i in range(self.style_count): 117 | if i < self.coarse_ind: 118 | style = GradualStyleBlock(512, 512, 16) 119 | elif i < self.middle_ind: 120 | style = GradualStyleBlock(512, 512, 32) 121 | else: 122 | style = GradualStyleBlock(512, 512, 64) 123 | self.styles.append(style) 124 | self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0) 125 | self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0) 126 | 127 | def _upsample_add(self, x, y): 128 | _, _, H, W = y.size() 129 | return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y 130 | 131 | def forward(self, x): 132 | x = self.conv1(x) 133 | x = self.bn1(x) 134 | x = self.relu(x) 135 | 136 | latents = [] 137 | modulelist = list(self.body._modules.values()) 138 | for i, l in enumerate(modulelist): 139 | x = l(x) 140 | if i == 6: 141 | c1 = x 142 | elif i == 12: 143 | c2 = x 144 | elif i == 15: 145 | c3 = x 146 | 147 | for j in range(self.coarse_ind): 148 | latents.append(self.styles[j](c3)) 149 | 150 | p2 = self._upsample_add(c3, self.latlayer1(c2)) 151 | for j in range(self.coarse_ind, self.middle_ind): 152 | latents.append(self.styles[j](p2)) 153 | 154 | p1 = self._upsample_add(p2, self.latlayer2(c1)) 155 | for j in range(self.middle_ind, self.style_count): 156 | latents.append(self.styles[j](p1)) 157 | 158 | out = torch.stack(latents, dim=1) 159 | return out 160 | -------------------------------------------------------------------------------- /models/encoders/helpers.py: -------------------------------------------------------------------------------- 1 | from collections import namedtuple 2 | import torch 3 | from torch.nn import Conv2d, BatchNorm2d, PReLU, ReLU, Sigmoid, MaxPool2d, AdaptiveAvgPool2d, Sequential, Module 4 | 5 | """ 6 | ArcFace implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch) 7 | """ 8 | 9 | 10 | class Flatten(Module): 11 | def forward(self, input): 12 | return input.view(input.size(0), -1) 13 | 14 | 15 | def l2_norm(input, axis=1): 16 | norm = torch.norm(input, 2, axis, True) 17 | output = torch.div(input, norm) 18 | return output 19 | 20 | 21 | class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])): 22 | """ A named tuple describing a ResNet block. """ 23 | 24 | 25 | def get_block(in_channel, depth, num_units, stride=2): 26 | return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)] 27 | 28 | 29 | def get_blocks(num_layers): 30 | if num_layers == 50: 31 | blocks = [ 32 | get_block(in_channel=64, depth=64, num_units=3), 33 | get_block(in_channel=64, depth=128, num_units=4), 34 | get_block(in_channel=128, depth=256, num_units=14), 35 | get_block(in_channel=256, depth=512, num_units=3) 36 | ] 37 | elif num_layers == 100: 38 | blocks = [ 39 | get_block(in_channel=64, depth=64, num_units=3), 40 | get_block(in_channel=64, depth=128, num_units=13), 41 | get_block(in_channel=128, depth=256, num_units=30), 42 | get_block(in_channel=256, depth=512, num_units=3) 43 | ] 44 | elif num_layers == 152: 45 | blocks = [ 46 | get_block(in_channel=64, depth=64, num_units=3), 47 | get_block(in_channel=64, depth=128, num_units=8), 48 | get_block(in_channel=128, depth=256, num_units=36), 49 | get_block(in_channel=256, depth=512, num_units=3) 50 | ] 51 | else: 52 | raise ValueError("Invalid number of layers: {}. Must be one of [50, 100, 152]".format(num_layers)) 53 | return blocks 54 | 55 | 56 | class SEModule(Module): 57 | def __init__(self, channels, reduction): 58 | super(SEModule, self).__init__() 59 | self.avg_pool = AdaptiveAvgPool2d(1) 60 | self.fc1 = Conv2d(channels, channels // reduction, kernel_size=1, padding=0, bias=False) 61 | self.relu = ReLU(inplace=True) 62 | self.fc2 = Conv2d(channels // reduction, channels, kernel_size=1, padding=0, bias=False) 63 | self.sigmoid = Sigmoid() 64 | 65 | def forward(self, x): 66 | module_input = x 67 | x = self.avg_pool(x) 68 | x = self.fc1(x) 69 | x = self.relu(x) 70 | x = self.fc2(x) 71 | x = self.sigmoid(x) 72 | return module_input * x 73 | 74 | 75 | class bottleneck_IR(Module): 76 | def __init__(self, in_channel, depth, stride): 77 | super(bottleneck_IR, self).__init__() 78 | if in_channel == depth: 79 | self.shortcut_layer = MaxPool2d(1, stride) 80 | else: 81 | self.shortcut_layer = Sequential( 82 | Conv2d(in_channel, depth, (1, 1), stride, bias=False), 83 | BatchNorm2d(depth) 84 | ) 85 | self.res_layer = Sequential( 86 | BatchNorm2d(in_channel), 87 | Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth), 88 | Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth) 89 | ) 90 | 91 | def forward(self, x): 92 | shortcut = self.shortcut_layer(x) 93 | res = self.res_layer(x) 94 | return res + shortcut 95 | 96 | 97 | class bottleneck_IR_SE(Module): 98 | def __init__(self, in_channel, depth, stride): 99 | super(bottleneck_IR_SE, self).__init__() 100 | if in_channel == depth: 101 | self.shortcut_layer = MaxPool2d(1, stride) 102 | else: 103 | self.shortcut_layer = Sequential( 104 | Conv2d(in_channel, depth, (1, 1), stride, bias=False), 105 | BatchNorm2d(depth) 106 | ) 107 | self.res_layer = Sequential( 108 | BatchNorm2d(in_channel), 109 | Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), 110 | PReLU(depth), 111 | Conv2d(depth, depth, (3, 3), stride, 1, bias=False), 112 | BatchNorm2d(depth), 113 | SEModule(depth, 16) 114 | ) 115 | 116 | def forward(self, x): 117 | shortcut = self.shortcut_layer(x) 118 | res = self.res_layer(x) 119 | return res + shortcut 120 | -------------------------------------------------------------------------------- /models/encoders/map2style.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from torch import nn 3 | from torch.nn import Conv2d, Module 4 | 5 | from models.stylegan2.model import EqualLinear 6 | 7 | 8 | class GradualStyleBlock(Module): 9 | def __init__(self, in_c, out_c, spatial): 10 | super(GradualStyleBlock, self).__init__() 11 | self.out_c = out_c 12 | self.spatial = spatial 13 | num_pools = int(np.log2(spatial)) 14 | modules = [] 15 | modules += [Conv2d(in_c, out_c, kernel_size=3, stride=2, padding=1), 16 | nn.LeakyReLU()] 17 | for i in range(num_pools - 1): 18 | modules += [ 19 | Conv2d(out_c, out_c, kernel_size=3, stride=2, padding=1), 20 | nn.LeakyReLU() 21 | ] 22 | self.convs = nn.Sequential(*modules) 23 | self.linear = EqualLinear(out_c, out_c, lr_mul=1) 24 | 25 | def forward(self, x): 26 | x = self.convs(x) 27 | x = x.view(-1, self.out_c) 28 | x = self.linear(x) 29 | return x 30 | -------------------------------------------------------------------------------- /models/encoders/model_irse.py: -------------------------------------------------------------------------------- 1 | from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Dropout, Sequential, Module 2 | from models.encoders.helpers import get_blocks, Flatten, bottleneck_IR, bottleneck_IR_SE, l2_norm 3 | 4 | """ 5 | Modified Backbone implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch) 6 | """ 7 | 8 | 9 | class Backbone(Module): 10 | def __init__(self, input_size, num_layers, mode='ir', drop_ratio=0.4, affine=True): 11 | super(Backbone, self).__init__() 12 | assert input_size in [112, 224], "input_size should be 112 or 224" 13 | assert num_layers in [50, 100, 152], "num_layers should be 50, 100 or 152" 14 | assert mode in ['ir', 'ir_se'], "mode should be ir or ir_se" 15 | blocks = get_blocks(num_layers) 16 | if mode == 'ir': 17 | unit_module = bottleneck_IR 18 | elif mode == 'ir_se': 19 | unit_module = bottleneck_IR_SE 20 | self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False), 21 | BatchNorm2d(64), 22 | PReLU(64)) 23 | if input_size == 112: 24 | self.output_layer = Sequential(BatchNorm2d(512), 25 | Dropout(drop_ratio), 26 | Flatten(), 27 | Linear(512 * 7 * 7, 512), 28 | BatchNorm1d(512, affine=affine)) 29 | else: 30 | self.output_layer = Sequential(BatchNorm2d(512), 31 | Dropout(drop_ratio), 32 | Flatten(), 33 | Linear(512 * 14 * 14, 512), 34 | BatchNorm1d(512, affine=affine)) 35 | 36 | modules = [] 37 | for block in blocks: 38 | for bottleneck in block: 39 | modules.append(unit_module(bottleneck.in_channel, 40 | bottleneck.depth, 41 | bottleneck.stride)) 42 | self.body = Sequential(*modules) 43 | 44 | def forward(self, x): 45 | x = self.input_layer(x) 46 | x = self.body(x) 47 | x = self.output_layer(x) 48 | return l2_norm(x) 49 | 50 | 51 | def IR_50(input_size): 52 | """Constructs a ir-50 model.""" 53 | model = Backbone(input_size, num_layers=50, mode='ir', drop_ratio=0.4, affine=False) 54 | return model 55 | 56 | 57 | def IR_101(input_size): 58 | """Constructs a ir-101 model.""" 59 | model = Backbone(input_size, num_layers=100, mode='ir', drop_ratio=0.4, affine=False) 60 | return model 61 | 62 | 63 | def IR_152(input_size): 64 | """Constructs a ir-152 model.""" 65 | model = Backbone(input_size, num_layers=152, mode='ir', drop_ratio=0.4, affine=False) 66 | return model 67 | 68 | 69 | def IR_SE_50(input_size): 70 | """Constructs a ir_se-50 model.""" 71 | model = Backbone(input_size, num_layers=50, mode='ir_se', drop_ratio=0.4, affine=False) 72 | return model 73 | 74 | 75 | def IR_SE_101(input_size): 76 | """Constructs a ir_se-101 model.""" 77 | model = Backbone(input_size, num_layers=100, mode='ir_se', drop_ratio=0.4, affine=False) 78 | return model 79 | 80 | 81 | def IR_SE_152(input_size): 82 | """Constructs a ir_se-152 model.""" 83 | model = Backbone(input_size, num_layers=152, mode='ir_se', drop_ratio=0.4, affine=False) 84 | return model 85 | -------------------------------------------------------------------------------- /models/encoders/restyle_e4e_encoders.py: -------------------------------------------------------------------------------- 1 | from enum import Enum 2 | from torch import nn 3 | from torch.nn import Conv2d, BatchNorm2d, PReLU, Sequential, Module 4 | from torchvision.models import resnet34 5 | 6 | from models.encoders.helpers import get_blocks, bottleneck_IR, bottleneck_IR_SE 7 | from models.encoders.map2style import GradualStyleBlock 8 | 9 | from models.stylegan2.model import EqualLinear 10 | 11 | class ProgressiveStage(Enum): 12 | WTraining = 0 13 | Delta1Training = 1 14 | Delta2Training = 2 15 | Delta3Training = 3 16 | Delta4Training = 4 17 | Delta5Training = 5 18 | Delta6Training = 6 19 | Delta7Training = 7 20 | Delta8Training = 8 21 | Delta9Training = 9 22 | Delta10Training = 10 23 | Delta11Training = 11 24 | Delta12Training = 12 25 | Delta13Training = 13 26 | Delta14Training = 14 27 | Delta15Training = 15 28 | Delta16Training = 16 29 | Delta17Training = 17 30 | Inference = 18 31 | 32 | 33 | class ProgressiveBackboneEncoder(Module): 34 | """ 35 | Paint2pix uses a Restyle-like architecture for building the canvas and identity encoder. 36 | This is a combined class which can be used as either the canvas or identity encoders 37 | depending on the input arguements 'is_canvas_encoder=False' 38 | """ 39 | def __init__(self, num_layers, mode='ir', n_styles=18, opts=None, is_canvas_encoder=False): 40 | super(ProgressiveBackboneEncoder, self).__init__() 41 | assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152' 42 | assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se' 43 | blocks = get_blocks(num_layers) 44 | if mode == 'ir': 45 | unit_module = bottleneck_IR 46 | elif mode == 'ir_se': 47 | unit_module = bottleneck_IR_SE 48 | 49 | self.input_layer = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False), 50 | BatchNorm2d(64), 51 | PReLU(64)) 52 | 53 | self.is_canvas_encoder = opts.is_canvas_encoder 54 | if self.is_canvas_encoder: 55 | self.input_layer_updated_canvas = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False), 56 | BatchNorm2d(64), 57 | PReLU(64)) 58 | 59 | modules = [] 60 | for block in blocks: 61 | for bottleneck in block: 62 | modules.append(unit_module(bottleneck.in_channel, 63 | bottleneck.depth, 64 | bottleneck.stride)) 65 | self.body = Sequential(*modules) 66 | 67 | self.styles = nn.ModuleList() 68 | self.style_count = n_styles 69 | for i in range(self.style_count): 70 | style = GradualStyleBlock(512, 512, 16) 71 | self.styles.append(style) 72 | self.progressive_stage = ProgressiveStage.Inference 73 | 74 | self.id_layers = nn.ModuleList() 75 | for i in range(self.style_count): 76 | id_layer = EqualLinear(512, 512, lr_mul=1) 77 | self.id_layers.append(id_layer) 78 | 79 | def get_deltas_starting_dimensions(self): 80 | ''' Get a list of the initial dimension of every delta from which it is applied ''' 81 | return list(range(self.style_count)) # Each dimension has a delta applied to 82 | 83 | def set_progressive_stage(self, new_stage: ProgressiveStage): 84 | # In this encoder we train all the pyramid (At least as a first stage experiment 85 | self.progressive_stage = new_stage 86 | print('Changed progressive stage to: ', new_stage) 87 | 88 | def forward(self, x_, target_id_feat=None, get_multiple_codes=False): 89 | x = self.input_layer(x_) 90 | x = self.body(x) 91 | 92 | # get initial w0 from first map2style layer 93 | w0 = self.styles[0](x) 94 | if target_id_feat is not None: 95 | w0 += self.id_layers[0](target_id_feat) 96 | w = w0.repeat(self.style_count, 1, 1).permute(1, 0, 2) 97 | 98 | # learn the deltas up to the current stage 99 | stage = self.progressive_stage.value 100 | for i in range(1, min(stage + 1, self.style_count)): 101 | delta_i = self.styles[i](x) 102 | if target_id_feat is not None: 103 | delta_i += self.id_layers[i](target_id_feat) 104 | w[:, i] += delta_i 105 | 106 | if self.is_canvas_encoder and get_multiple_codes: 107 | w_canvas0 = w 108 | x = self.input_layer_updated_canvas(x_) 109 | x = self.body(x) 110 | 111 | # get initial w0 from first map2style layer 112 | w0 = self.styles[0](x) 113 | if target_id_feat is not None: 114 | w0 += self.id_layers[0](target_id_feat) 115 | w = w0.repeat(self.style_count, 1, 1).permute(1, 0, 2) 116 | 117 | # learn the deltas up to the current stage 118 | stage = self.progressive_stage.value 119 | for i in range(1, min(stage + 1, self.style_count)): 120 | delta_i = self.styles[i](x) 121 | if target_id_feat is not None: 122 | delta_i += self.id_layers[i](target_id_feat) 123 | w[:, i] += delta_i 124 | w_canvas1 = w 125 | return w_canvas0, w_canvas1 126 | else: 127 | return w -------------------------------------------------------------------------------- /models/encoders/restyle_psp_encoders.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | from torch.nn import Conv2d, BatchNorm2d, PReLU, Sequential, Module 4 | from torchvision.models.resnet import resnet34 5 | 6 | from models.encoders.helpers import get_blocks, bottleneck_IR, bottleneck_IR_SE 7 | from models.encoders.map2style import GradualStyleBlock 8 | 9 | 10 | class BackboneEncoder(Module): 11 | """ 12 | The simpler backbone architecture used by ReStyle where all style vectors are extracted from the final 16x16 feature 13 | map of the encoder. This classes uses the simplified architecture applied over an ResNet IRSE-50 backbone. 14 | Note this class is designed to be used for the human facial domain. 15 | """ 16 | def __init__(self, num_layers, mode='ir', n_styles=18, opts=None): 17 | super(BackboneEncoder, self).__init__() 18 | assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152' 19 | assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se' 20 | blocks = get_blocks(num_layers) 21 | if mode == 'ir': 22 | unit_module = bottleneck_IR 23 | elif mode == 'ir_se': 24 | unit_module = bottleneck_IR_SE 25 | 26 | self.input_layer = Sequential(Conv2d(opts.input_nc, 64, (3, 3), 1, 1, bias=False), 27 | BatchNorm2d(64), 28 | PReLU(64)) 29 | modules = [] 30 | for block in blocks: 31 | for bottleneck in block: 32 | modules.append(unit_module(bottleneck.in_channel, 33 | bottleneck.depth, 34 | bottleneck.stride)) 35 | self.body = Sequential(*modules) 36 | 37 | self.styles = nn.ModuleList() 38 | self.style_count = n_styles 39 | for i in range(self.style_count): 40 | style = GradualStyleBlock(512, 512, 16) 41 | self.styles.append(style) 42 | 43 | def forward(self, x): 44 | x = self.input_layer(x) 45 | x = self.body(x) 46 | latents = [] 47 | for j in range(self.style_count): 48 | latents.append(self.styles[j](x)) 49 | out = torch.stack(latents, dim=1) 50 | return out 51 | 52 | 53 | class ResNetBackboneEncoder(Module): 54 | """ 55 | The simpler backbone architecture used by ReStyle where all style vectors are extracted from the final 16x16 feature 56 | map of the encoder. This classes uses the simplified architecture applied over an ResNet34 backbone. 57 | """ 58 | def __init__(self, n_styles=18, opts=None): 59 | super(ResNetBackboneEncoder, self).__init__() 60 | 61 | self.conv1 = nn.Conv2d(opts.input_nc, 64, kernel_size=7, stride=2, padding=3, bias=False) 62 | self.bn1 = BatchNorm2d(64) 63 | self.relu = PReLU(64) 64 | 65 | resnet_basenet = resnet34(pretrained=True) 66 | blocks = [ 67 | resnet_basenet.layer1, 68 | resnet_basenet.layer2, 69 | resnet_basenet.layer3, 70 | resnet_basenet.layer4 71 | ] 72 | modules = [] 73 | for block in blocks: 74 | for bottleneck in block: 75 | modules.append(bottleneck) 76 | self.body = Sequential(*modules) 77 | 78 | self.styles = nn.ModuleList() 79 | self.style_count = n_styles 80 | for i in range(self.style_count): 81 | style = GradualStyleBlock(512, 512, 16) 82 | self.styles.append(style) 83 | 84 | def forward(self, x): 85 | x = self.conv1(x) 86 | x = self.bn1(x) 87 | x = self.relu(x) 88 | x = self.body(x) 89 | latents = [] 90 | for j in range(self.style_count): 91 | latents.append(self.styles[j](x)) 92 | out = torch.stack(latents, dim=1) 93 | return out 94 | -------------------------------------------------------------------------------- /models/mtcnn/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/mtcnn/__init__.py -------------------------------------------------------------------------------- /models/mtcnn/mtcnn.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | from PIL import Image 4 | from models.mtcnn.mtcnn_pytorch.src.get_nets import PNet, RNet, ONet 5 | from models.mtcnn.mtcnn_pytorch.src.box_utils import nms, calibrate_box, get_image_boxes, convert_to_square 6 | from models.mtcnn.mtcnn_pytorch.src.first_stage import run_first_stage 7 | from models.mtcnn.mtcnn_pytorch.src.align_trans import get_reference_facial_points, warp_and_crop_face 8 | 9 | device = 'cuda:0' 10 | 11 | 12 | class MTCNN(): 13 | def __init__(self): 14 | print(device) 15 | self.pnet = PNet().to(device) 16 | self.rnet = RNet().to(device) 17 | self.onet = ONet().to(device) 18 | self.pnet.eval() 19 | self.rnet.eval() 20 | self.onet.eval() 21 | self.refrence = get_reference_facial_points(default_square=True) 22 | 23 | def align(self, img): 24 | _, landmarks = self.detect_faces(img) 25 | if len(landmarks) == 0: 26 | return None, None 27 | facial5points = [[landmarks[0][j], landmarks[0][j + 5]] for j in range(5)] 28 | warped_face, tfm = warp_and_crop_face(np.array(img), facial5points, self.refrence, crop_size=(112, 112)) 29 | return Image.fromarray(warped_face), tfm 30 | 31 | def align_multi(self, img, limit=None, min_face_size=30.0): 32 | boxes, landmarks = self.detect_faces(img, min_face_size) 33 | if limit: 34 | boxes = boxes[:limit] 35 | landmarks = landmarks[:limit] 36 | faces = [] 37 | tfms = [] 38 | for landmark in landmarks: 39 | facial5points = [[landmark[j], landmark[j + 5]] for j in range(5)] 40 | warped_face, tfm = warp_and_crop_face(np.array(img), facial5points, self.refrence, crop_size=(112, 112)) 41 | faces.append(Image.fromarray(warped_face)) 42 | tfms.append(tfm) 43 | return boxes, faces, tfms 44 | 45 | def detect_faces(self, image, min_face_size=20.0, 46 | thresholds=[0.15, 0.25, 0.35], 47 | nms_thresholds=[0.7, 0.7, 0.7]): 48 | """ 49 | Arguments: 50 | image: an instance of PIL.Image. 51 | min_face_size: a float number. 52 | thresholds: a list of length 3. 53 | nms_thresholds: a list of length 3. 54 | 55 | Returns: 56 | two float numpy arrays of shapes [n_boxes, 4] and [n_boxes, 10], 57 | bounding boxes and facial landmarks. 58 | """ 59 | 60 | # BUILD AN IMAGE PYRAMID 61 | width, height = image.size 62 | min_length = min(height, width) 63 | 64 | min_detection_size = 12 65 | factor = 0.707 # sqrt(0.5) 66 | 67 | # scales for scaling the image 68 | scales = [] 69 | 70 | # scales the image so that 71 | # minimum size that we can detect equals to 72 | # minimum face size that we want to detect 73 | m = min_detection_size / min_face_size 74 | min_length *= m 75 | 76 | factor_count = 0 77 | while min_length > min_detection_size: 78 | scales.append(m * factor ** factor_count) 79 | min_length *= factor 80 | factor_count += 1 81 | 82 | # STAGE 1 83 | 84 | # it will be returned 85 | bounding_boxes = [] 86 | 87 | with torch.no_grad(): 88 | # run P-Net on different scales 89 | for s in scales: 90 | boxes = run_first_stage(image, self.pnet, scale=s, threshold=thresholds[0]) 91 | bounding_boxes.append(boxes) 92 | 93 | # collect boxes (and offsets, and scores) from different scales 94 | bounding_boxes = [i for i in bounding_boxes if i is not None] 95 | bounding_boxes = np.vstack(bounding_boxes) 96 | 97 | keep = nms(bounding_boxes[:, 0:5], nms_thresholds[0]) 98 | bounding_boxes = bounding_boxes[keep] 99 | 100 | # use offsets predicted by pnet to transform bounding boxes 101 | bounding_boxes = calibrate_box(bounding_boxes[:, 0:5], bounding_boxes[:, 5:]) 102 | # shape [n_boxes, 5] 103 | 104 | bounding_boxes = convert_to_square(bounding_boxes) 105 | bounding_boxes[:, 0:4] = np.round(bounding_boxes[:, 0:4]) 106 | 107 | # STAGE 2 108 | 109 | img_boxes = get_image_boxes(bounding_boxes, image, size=24) 110 | img_boxes = torch.FloatTensor(img_boxes).to(device) 111 | 112 | output = self.rnet(img_boxes) 113 | offsets = output[0].cpu().data.numpy() # shape [n_boxes, 4] 114 | probs = output[1].cpu().data.numpy() # shape [n_boxes, 2] 115 | 116 | keep = np.where(probs[:, 1] > thresholds[1])[0] 117 | bounding_boxes = bounding_boxes[keep] 118 | bounding_boxes[:, 4] = probs[keep, 1].reshape((-1,)) 119 | offsets = offsets[keep] 120 | 121 | keep = nms(bounding_boxes, nms_thresholds[1]) 122 | bounding_boxes = bounding_boxes[keep] 123 | bounding_boxes = calibrate_box(bounding_boxes, offsets[keep]) 124 | bounding_boxes = convert_to_square(bounding_boxes) 125 | bounding_boxes[:, 0:4] = np.round(bounding_boxes[:, 0:4]) 126 | 127 | # STAGE 3 128 | 129 | img_boxes = get_image_boxes(bounding_boxes, image, size=48) 130 | if len(img_boxes) == 0: 131 | return [], [] 132 | img_boxes = torch.FloatTensor(img_boxes).to(device) 133 | output = self.onet(img_boxes) 134 | landmarks = output[0].cpu().data.numpy() # shape [n_boxes, 10] 135 | offsets = output[1].cpu().data.numpy() # shape [n_boxes, 4] 136 | probs = output[2].cpu().data.numpy() # shape [n_boxes, 2] 137 | 138 | keep = np.where(probs[:, 1] > thresholds[2])[0] 139 | bounding_boxes = bounding_boxes[keep] 140 | bounding_boxes[:, 4] = probs[keep, 1].reshape((-1,)) 141 | offsets = offsets[keep] 142 | landmarks = landmarks[keep] 143 | 144 | # compute landmark points 145 | width = bounding_boxes[:, 2] - bounding_boxes[:, 0] + 1.0 146 | height = bounding_boxes[:, 3] - bounding_boxes[:, 1] + 1.0 147 | xmin, ymin = bounding_boxes[:, 0], bounding_boxes[:, 1] 148 | landmarks[:, 0:5] = np.expand_dims(xmin, 1) + np.expand_dims(width, 1) * landmarks[:, 0:5] 149 | landmarks[:, 5:10] = np.expand_dims(ymin, 1) + np.expand_dims(height, 1) * landmarks[:, 5:10] 150 | 151 | bounding_boxes = calibrate_box(bounding_boxes, offsets) 152 | keep = nms(bounding_boxes, nms_thresholds[2], mode='min') 153 | bounding_boxes = bounding_boxes[keep] 154 | landmarks = landmarks[keep] 155 | 156 | return bounding_boxes, landmarks 157 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/mtcnn/mtcnn_pytorch/__init__.py -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/__init__.py: -------------------------------------------------------------------------------- 1 | from .visualization_utils import show_bboxes 2 | from .detector import detect_faces 3 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/align_trans.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | Created on Mon Apr 24 15:43:29 2017 4 | @author: zhaoy 5 | """ 6 | import numpy as np 7 | import cv2 8 | 9 | # from scipy.linalg import lstsq 10 | # from scipy.ndimage import geometric_transform # , map_coordinates 11 | 12 | from models.mtcnn.mtcnn_pytorch.src.matlab_cp2tform import get_similarity_transform_for_cv2 13 | 14 | # reference facial points, a list of coordinates (x,y) 15 | REFERENCE_FACIAL_POINTS = [ 16 | [30.29459953, 51.69630051], 17 | [65.53179932, 51.50139999], 18 | [48.02519989, 71.73660278], 19 | [33.54930115, 92.3655014], 20 | [62.72990036, 92.20410156] 21 | ] 22 | 23 | DEFAULT_CROP_SIZE = (96, 112) 24 | 25 | 26 | class FaceWarpException(Exception): 27 | def __str__(self): 28 | return 'In File {}:{}'.format( 29 | __file__, super.__str__(self)) 30 | 31 | 32 | def get_reference_facial_points(output_size=None, 33 | inner_padding_factor=0.0, 34 | outer_padding=(0, 0), 35 | default_square=False): 36 | """ 37 | Function: 38 | ---------- 39 | get reference 5 key points according to crop settings: 40 | 0. Set default crop_size: 41 | if default_square: 42 | crop_size = (112, 112) 43 | else: 44 | crop_size = (96, 112) 45 | 1. Pad the crop_size by inner_padding_factor in each side; 46 | 2. Resize crop_size into (output_size - outer_padding*2), 47 | pad into output_size with outer_padding; 48 | 3. Output reference_5point; 49 | Parameters: 50 | ---------- 51 | @output_size: (w, h) or None 52 | size of aligned face image 53 | @inner_padding_factor: (w_factor, h_factor) 54 | padding factor for inner (w, h) 55 | @outer_padding: (w_pad, h_pad) 56 | each row is a pair of coordinates (x, y) 57 | @default_square: True or False 58 | if True: 59 | default crop_size = (112, 112) 60 | else: 61 | default crop_size = (96, 112); 62 | !!! make sure, if output_size is not None: 63 | (output_size - outer_padding) 64 | = some_scale * (default crop_size * (1.0 + inner_padding_factor)) 65 | Returns: 66 | ---------- 67 | @reference_5point: 5x2 np.array 68 | each row is a pair of transformed coordinates (x, y) 69 | """ 70 | # print('\n===> get_reference_facial_points():') 71 | 72 | # print('---> Params:') 73 | # print(' output_size: ', output_size) 74 | # print(' inner_padding_factor: ', inner_padding_factor) 75 | # print(' outer_padding:', outer_padding) 76 | # print(' default_square: ', default_square) 77 | 78 | tmp_5pts = np.array(REFERENCE_FACIAL_POINTS) 79 | tmp_crop_size = np.array(DEFAULT_CROP_SIZE) 80 | 81 | # 0) make the inner region a square 82 | if default_square: 83 | size_diff = max(tmp_crop_size) - tmp_crop_size 84 | tmp_5pts += size_diff / 2 85 | tmp_crop_size += size_diff 86 | 87 | # print('---> default:') 88 | # print(' crop_size = ', tmp_crop_size) 89 | # print(' reference_5pts = ', tmp_5pts) 90 | 91 | if (output_size and 92 | output_size[0] == tmp_crop_size[0] and 93 | output_size[1] == tmp_crop_size[1]): 94 | # print('output_size == DEFAULT_CROP_SIZE {}: return default reference points'.format(tmp_crop_size)) 95 | return tmp_5pts 96 | 97 | if (inner_padding_factor == 0 and 98 | outer_padding == (0, 0)): 99 | if output_size is None: 100 | # print('No paddings to do: return default reference points') 101 | return tmp_5pts 102 | else: 103 | raise FaceWarpException( 104 | 'No paddings to do, output_size must be None or {}'.format(tmp_crop_size)) 105 | 106 | # check output size 107 | if not (0 <= inner_padding_factor <= 1.0): 108 | raise FaceWarpException('Not (0 <= inner_padding_factor <= 1.0)') 109 | 110 | if ((inner_padding_factor > 0 or outer_padding[0] > 0 or outer_padding[1] > 0) 111 | and output_size is None): 112 | output_size = tmp_crop_size * \ 113 | (1 + inner_padding_factor * 2).astype(np.int32) 114 | output_size += np.array(outer_padding) 115 | # print(' deduced from paddings, output_size = ', output_size) 116 | 117 | if not (outer_padding[0] < output_size[0] 118 | and outer_padding[1] < output_size[1]): 119 | raise FaceWarpException('Not (outer_padding[0] < output_size[0]' 120 | 'and outer_padding[1] < output_size[1])') 121 | 122 | # 1) pad the inner region according inner_padding_factor 123 | # print('---> STEP1: pad the inner region according inner_padding_factor') 124 | if inner_padding_factor > 0: 125 | size_diff = tmp_crop_size * inner_padding_factor * 2 126 | tmp_5pts += size_diff / 2 127 | tmp_crop_size += np.round(size_diff).astype(np.int32) 128 | 129 | # print(' crop_size = ', tmp_crop_size) 130 | # print(' reference_5pts = ', tmp_5pts) 131 | 132 | # 2) resize the padded inner region 133 | # print('---> STEP2: resize the padded inner region') 134 | size_bf_outer_pad = np.array(output_size) - np.array(outer_padding) * 2 135 | # print(' crop_size = ', tmp_crop_size) 136 | # print(' size_bf_outer_pad = ', size_bf_outer_pad) 137 | 138 | if size_bf_outer_pad[0] * tmp_crop_size[1] != size_bf_outer_pad[1] * tmp_crop_size[0]: 139 | raise FaceWarpException('Must have (output_size - outer_padding)' 140 | '= some_scale * (crop_size * (1.0 + inner_padding_factor)') 141 | 142 | scale_factor = size_bf_outer_pad[0].astype(np.float32) / tmp_crop_size[0] 143 | # print(' resize scale_factor = ', scale_factor) 144 | tmp_5pts = tmp_5pts * scale_factor 145 | # size_diff = tmp_crop_size * (scale_factor - min(scale_factor)) 146 | # tmp_5pts = tmp_5pts + size_diff / 2 147 | tmp_crop_size = size_bf_outer_pad 148 | # print(' crop_size = ', tmp_crop_size) 149 | # print(' reference_5pts = ', tmp_5pts) 150 | 151 | # 3) add outer_padding to make output_size 152 | reference_5point = tmp_5pts + np.array(outer_padding) 153 | tmp_crop_size = output_size 154 | # print('---> STEP3: add outer_padding to make output_size') 155 | # print(' crop_size = ', tmp_crop_size) 156 | # print(' reference_5pts = ', tmp_5pts) 157 | 158 | # print('===> end get_reference_facial_points\n') 159 | 160 | return reference_5point 161 | 162 | 163 | def get_affine_transform_matrix(src_pts, dst_pts): 164 | """ 165 | Function: 166 | ---------- 167 | get affine transform matrix 'tfm' from src_pts to dst_pts 168 | Parameters: 169 | ---------- 170 | @src_pts: Kx2 np.array 171 | source points matrix, each row is a pair of coordinates (x, y) 172 | @dst_pts: Kx2 np.array 173 | destination points matrix, each row is a pair of coordinates (x, y) 174 | Returns: 175 | ---------- 176 | @tfm: 2x3 np.array 177 | transform matrix from src_pts to dst_pts 178 | """ 179 | 180 | tfm = np.float32([[1, 0, 0], [0, 1, 0]]) 181 | n_pts = src_pts.shape[0] 182 | ones = np.ones((n_pts, 1), src_pts.dtype) 183 | src_pts_ = np.hstack([src_pts, ones]) 184 | dst_pts_ = np.hstack([dst_pts, ones]) 185 | 186 | # #print(('src_pts_:\n' + str(src_pts_)) 187 | # #print(('dst_pts_:\n' + str(dst_pts_)) 188 | 189 | A, res, rank, s = np.linalg.lstsq(src_pts_, dst_pts_) 190 | 191 | # #print(('np.linalg.lstsq return A: \n' + str(A)) 192 | # #print(('np.linalg.lstsq return res: \n' + str(res)) 193 | # #print(('np.linalg.lstsq return rank: \n' + str(rank)) 194 | # #print(('np.linalg.lstsq return s: \n' + str(s)) 195 | 196 | if rank == 3: 197 | tfm = np.float32([ 198 | [A[0, 0], A[1, 0], A[2, 0]], 199 | [A[0, 1], A[1, 1], A[2, 1]] 200 | ]) 201 | elif rank == 2: 202 | tfm = np.float32([ 203 | [A[0, 0], A[1, 0], 0], 204 | [A[0, 1], A[1, 1], 0] 205 | ]) 206 | 207 | return tfm 208 | 209 | 210 | def warp_and_crop_face(src_img, 211 | facial_pts, 212 | reference_pts=None, 213 | crop_size=(96, 112), 214 | align_type='smilarity'): 215 | """ 216 | Function: 217 | ---------- 218 | apply affine transform 'trans' to uv 219 | Parameters: 220 | ---------- 221 | @src_img: 3x3 np.array 222 | input image 223 | @facial_pts: could be 224 | 1)a list of K coordinates (x,y) 225 | or 226 | 2) Kx2 or 2xK np.array 227 | each row or col is a pair of coordinates (x, y) 228 | @reference_pts: could be 229 | 1) a list of K coordinates (x,y) 230 | or 231 | 2) Kx2 or 2xK np.array 232 | each row or col is a pair of coordinates (x, y) 233 | or 234 | 3) None 235 | if None, use default reference facial points 236 | @crop_size: (w, h) 237 | output face image size 238 | @align_type: transform type, could be one of 239 | 1) 'similarity': use similarity transform 240 | 2) 'cv2_affine': use the first 3 points to do affine transform, 241 | by calling cv2.getAffineTransform() 242 | 3) 'affine': use all points to do affine transform 243 | Returns: 244 | ---------- 245 | @face_img: output face image with size (w, h) = @crop_size 246 | """ 247 | 248 | if reference_pts is None: 249 | if crop_size[0] == 96 and crop_size[1] == 112: 250 | reference_pts = REFERENCE_FACIAL_POINTS 251 | else: 252 | default_square = False 253 | inner_padding_factor = 0 254 | outer_padding = (0, 0) 255 | output_size = crop_size 256 | 257 | reference_pts = get_reference_facial_points(output_size, 258 | inner_padding_factor, 259 | outer_padding, 260 | default_square) 261 | 262 | ref_pts = np.float32(reference_pts) 263 | ref_pts_shp = ref_pts.shape 264 | if max(ref_pts_shp) < 3 or min(ref_pts_shp) != 2: 265 | raise FaceWarpException( 266 | 'reference_pts.shape must be (K,2) or (2,K) and K>2') 267 | 268 | if ref_pts_shp[0] == 2: 269 | ref_pts = ref_pts.T 270 | 271 | src_pts = np.float32(facial_pts) 272 | src_pts_shp = src_pts.shape 273 | if max(src_pts_shp) < 3 or min(src_pts_shp) != 2: 274 | raise FaceWarpException( 275 | 'facial_pts.shape must be (K,2) or (2,K) and K>2') 276 | 277 | if src_pts_shp[0] == 2: 278 | src_pts = src_pts.T 279 | 280 | # #print('--->src_pts:\n', src_pts 281 | # #print('--->ref_pts\n', ref_pts 282 | 283 | if src_pts.shape != ref_pts.shape: 284 | raise FaceWarpException( 285 | 'facial_pts and reference_pts must have the same shape') 286 | 287 | if align_type is 'cv2_affine': 288 | tfm = cv2.getAffineTransform(src_pts[0:3], ref_pts[0:3]) 289 | # #print(('cv2.getAffineTransform() returns tfm=\n' + str(tfm)) 290 | elif align_type is 'affine': 291 | tfm = get_affine_transform_matrix(src_pts, ref_pts) 292 | # #print(('get_affine_transform_matrix() returns tfm=\n' + str(tfm)) 293 | else: 294 | tfm = get_similarity_transform_for_cv2(src_pts, ref_pts) 295 | # #print(('get_similarity_transform_for_cv2() returns tfm=\n' + str(tfm)) 296 | 297 | # #print('--->Transform matrix: ' 298 | # #print(('type(tfm):' + str(type(tfm))) 299 | # #print(('tfm.dtype:' + str(tfm.dtype)) 300 | # #print( tfm 301 | 302 | face_img = cv2.warpAffine(src_img, tfm, (crop_size[0], crop_size[1])) 303 | 304 | return face_img, tfm 305 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/box_utils.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from PIL import Image 3 | 4 | 5 | def nms(boxes, overlap_threshold=0.5, mode='union'): 6 | """Non-maximum suppression. 7 | 8 | Arguments: 9 | boxes: a float numpy array of shape [n, 5], 10 | where each row is (xmin, ymin, xmax, ymax, score). 11 | overlap_threshold: a float number. 12 | mode: 'union' or 'min'. 13 | 14 | Returns: 15 | list with indices of the selected boxes 16 | """ 17 | 18 | # if there are no boxes, return the empty list 19 | if len(boxes) == 0: 20 | return [] 21 | 22 | # list of picked indices 23 | pick = [] 24 | 25 | # grab the coordinates of the bounding boxes 26 | x1, y1, x2, y2, score = [boxes[:, i] for i in range(5)] 27 | 28 | area = (x2 - x1 + 1.0) * (y2 - y1 + 1.0) 29 | ids = np.argsort(score) # in increasing order 30 | 31 | while len(ids) > 0: 32 | 33 | # grab index of the largest value 34 | last = len(ids) - 1 35 | i = ids[last] 36 | pick.append(i) 37 | 38 | # compute intersections 39 | # of the box with the largest score 40 | # with the rest of boxes 41 | 42 | # left top corner of intersection boxes 43 | ix1 = np.maximum(x1[i], x1[ids[:last]]) 44 | iy1 = np.maximum(y1[i], y1[ids[:last]]) 45 | 46 | # right bottom corner of intersection boxes 47 | ix2 = np.minimum(x2[i], x2[ids[:last]]) 48 | iy2 = np.minimum(y2[i], y2[ids[:last]]) 49 | 50 | # width and height of intersection boxes 51 | w = np.maximum(0.0, ix2 - ix1 + 1.0) 52 | h = np.maximum(0.0, iy2 - iy1 + 1.0) 53 | 54 | # intersections' areas 55 | inter = w * h 56 | if mode == 'min': 57 | overlap = inter / np.minimum(area[i], area[ids[:last]]) 58 | elif mode == 'union': 59 | # intersection over union (IoU) 60 | overlap = inter / (area[i] + area[ids[:last]] - inter) 61 | 62 | # delete all boxes where overlap is too big 63 | ids = np.delete( 64 | ids, 65 | np.concatenate([[last], np.where(overlap > overlap_threshold)[0]]) 66 | ) 67 | 68 | return pick 69 | 70 | 71 | def convert_to_square(bboxes): 72 | """Convert bounding boxes to a square form. 73 | 74 | Arguments: 75 | bboxes: a float numpy array of shape [n, 5]. 76 | 77 | Returns: 78 | a float numpy array of shape [n, 5], 79 | squared bounding boxes. 80 | """ 81 | 82 | square_bboxes = np.zeros_like(bboxes) 83 | x1, y1, x2, y2 = [bboxes[:, i] for i in range(4)] 84 | h = y2 - y1 + 1.0 85 | w = x2 - x1 + 1.0 86 | max_side = np.maximum(h, w) 87 | square_bboxes[:, 0] = x1 + w * 0.5 - max_side * 0.5 88 | square_bboxes[:, 1] = y1 + h * 0.5 - max_side * 0.5 89 | square_bboxes[:, 2] = square_bboxes[:, 0] + max_side - 1.0 90 | square_bboxes[:, 3] = square_bboxes[:, 1] + max_side - 1.0 91 | return square_bboxes 92 | 93 | 94 | def calibrate_box(bboxes, offsets): 95 | """Transform bounding boxes to be more like true bounding boxes. 96 | 'offsets' is one of the outputs of the nets. 97 | 98 | Arguments: 99 | bboxes: a float numpy array of shape [n, 5]. 100 | offsets: a float numpy array of shape [n, 4]. 101 | 102 | Returns: 103 | a float numpy array of shape [n, 5]. 104 | """ 105 | x1, y1, x2, y2 = [bboxes[:, i] for i in range(4)] 106 | w = x2 - x1 + 1.0 107 | h = y2 - y1 + 1.0 108 | w = np.expand_dims(w, 1) 109 | h = np.expand_dims(h, 1) 110 | 111 | # this is what happening here: 112 | # tx1, ty1, tx2, ty2 = [offsets[:, i] for i in range(4)] 113 | # x1_true = x1 + tx1*w 114 | # y1_true = y1 + ty1*h 115 | # x2_true = x2 + tx2*w 116 | # y2_true = y2 + ty2*h 117 | # below is just more compact form of this 118 | 119 | # are offsets always such that 120 | # x1 < x2 and y1 < y2 ? 121 | 122 | translation = np.hstack([w, h, w, h]) * offsets 123 | bboxes[:, 0:4] = bboxes[:, 0:4] + translation 124 | return bboxes 125 | 126 | 127 | def get_image_boxes(bounding_boxes, img, size=24): 128 | """Cut out boxes from the image. 129 | 130 | Arguments: 131 | bounding_boxes: a float numpy array of shape [n, 5]. 132 | img: an instance of PIL.Image. 133 | size: an integer, size of cutouts. 134 | 135 | Returns: 136 | a float numpy array of shape [n, 3, size, size]. 137 | """ 138 | 139 | num_boxes = len(bounding_boxes) 140 | width, height = img.size 141 | 142 | [dy, edy, dx, edx, y, ey, x, ex, w, h] = correct_bboxes(bounding_boxes, width, height) 143 | img_boxes = np.zeros((num_boxes, 3, size, size), 'float32') 144 | 145 | for i in range(num_boxes): 146 | img_box = np.zeros((h[i], w[i], 3), 'uint8') 147 | 148 | img_array = np.asarray(img, 'uint8') 149 | img_box[dy[i]:(edy[i] + 1), dx[i]:(edx[i] + 1), :] = \ 150 | img_array[y[i]:(ey[i] + 1), x[i]:(ex[i] + 1), :] 151 | 152 | # resize 153 | img_box = Image.fromarray(img_box) 154 | img_box = img_box.resize((size, size), Image.BILINEAR) 155 | img_box = np.asarray(img_box, 'float32') 156 | 157 | img_boxes[i, :, :, :] = _preprocess(img_box) 158 | 159 | return img_boxes 160 | 161 | 162 | def correct_bboxes(bboxes, width, height): 163 | """Crop boxes that are too big and get coordinates 164 | with respect to cutouts. 165 | 166 | Arguments: 167 | bboxes: a float numpy array of shape [n, 5], 168 | where each row is (xmin, ymin, xmax, ymax, score). 169 | width: a float number. 170 | height: a float number. 171 | 172 | Returns: 173 | dy, dx, edy, edx: a int numpy arrays of shape [n], 174 | coordinates of the boxes with respect to the cutouts. 175 | y, x, ey, ex: a int numpy arrays of shape [n], 176 | corrected ymin, xmin, ymax, xmax. 177 | h, w: a int numpy arrays of shape [n], 178 | just heights and widths of boxes. 179 | 180 | in the following order: 181 | [dy, edy, dx, edx, y, ey, x, ex, w, h]. 182 | """ 183 | 184 | x1, y1, x2, y2 = [bboxes[:, i] for i in range(4)] 185 | w, h = x2 - x1 + 1.0, y2 - y1 + 1.0 186 | num_boxes = bboxes.shape[0] 187 | 188 | # 'e' stands for end 189 | # (x, y) -> (ex, ey) 190 | x, y, ex, ey = x1, y1, x2, y2 191 | 192 | # we need to cut out a box from the image. 193 | # (x, y, ex, ey) are corrected coordinates of the box 194 | # in the image. 195 | # (dx, dy, edx, edy) are coordinates of the box in the cutout 196 | # from the image. 197 | dx, dy = np.zeros((num_boxes,)), np.zeros((num_boxes,)) 198 | edx, edy = w.copy() - 1.0, h.copy() - 1.0 199 | 200 | # if box's bottom right corner is too far right 201 | ind = np.where(ex > width - 1.0)[0] 202 | edx[ind] = w[ind] + width - 2.0 - ex[ind] 203 | ex[ind] = width - 1.0 204 | 205 | # if box's bottom right corner is too low 206 | ind = np.where(ey > height - 1.0)[0] 207 | edy[ind] = h[ind] + height - 2.0 - ey[ind] 208 | ey[ind] = height - 1.0 209 | 210 | # if box's top left corner is too far left 211 | ind = np.where(x < 0.0)[0] 212 | dx[ind] = 0.0 - x[ind] 213 | x[ind] = 0.0 214 | 215 | # if box's top left corner is too high 216 | ind = np.where(y < 0.0)[0] 217 | dy[ind] = 0.0 - y[ind] 218 | y[ind] = 0.0 219 | 220 | return_list = [dy, edy, dx, edx, y, ey, x, ex, w, h] 221 | return_list = [i.astype('int32') for i in return_list] 222 | 223 | return return_list 224 | 225 | 226 | def _preprocess(img): 227 | """Preprocessing step before feeding the network. 228 | 229 | Arguments: 230 | img: a float numpy array of shape [h, w, c]. 231 | 232 | Returns: 233 | a float numpy array of shape [1, c, h, w]. 234 | """ 235 | img = img.transpose((2, 0, 1)) 236 | img = np.expand_dims(img, 0) 237 | img = (img - 127.5) * 0.0078125 238 | return img 239 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/detector.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | from .get_nets import PNet, RNet, ONet 4 | from .box_utils import nms, calibrate_box, get_image_boxes, convert_to_square 5 | from .first_stage import run_first_stage 6 | 7 | 8 | def detect_faces(image, min_face_size=20.0, 9 | thresholds=[0.6, 0.7, 0.8], 10 | nms_thresholds=[0.7, 0.7, 0.7]): 11 | """ 12 | Arguments: 13 | image: an instance of PIL.Image. 14 | min_face_size: a float number. 15 | thresholds: a list of length 3. 16 | nms_thresholds: a list of length 3. 17 | 18 | Returns: 19 | two float numpy arrays of shapes [n_boxes, 4] and [n_boxes, 10], 20 | bounding boxes and facial landmarks. 21 | """ 22 | 23 | # LOAD MODELS 24 | pnet = PNet() 25 | rnet = RNet() 26 | onet = ONet() 27 | onet.eval() 28 | 29 | # BUILD AN IMAGE PYRAMID 30 | width, height = image.size 31 | min_length = min(height, width) 32 | 33 | min_detection_size = 12 34 | factor = 0.707 # sqrt(0.5) 35 | 36 | # scales for scaling the image 37 | scales = [] 38 | 39 | # scales the image so that 40 | # minimum size that we can detect equals to 41 | # minimum face size that we want to detect 42 | m = min_detection_size / min_face_size 43 | min_length *= m 44 | 45 | factor_count = 0 46 | while min_length > min_detection_size: 47 | scales.append(m * factor ** factor_count) 48 | min_length *= factor 49 | factor_count += 1 50 | 51 | # STAGE 1 52 | 53 | # it will be returned 54 | bounding_boxes = [] 55 | 56 | with torch.no_grad(): 57 | # run P-Net on different scales 58 | for s in scales: 59 | boxes = run_first_stage(image, pnet, scale=s, threshold=thresholds[0]) 60 | bounding_boxes.append(boxes) 61 | 62 | # collect boxes (and offsets, and scores) from different scales 63 | bounding_boxes = [i for i in bounding_boxes if i is not None] 64 | bounding_boxes = np.vstack(bounding_boxes) 65 | 66 | keep = nms(bounding_boxes[:, 0:5], nms_thresholds[0]) 67 | bounding_boxes = bounding_boxes[keep] 68 | 69 | # use offsets predicted by pnet to transform bounding boxes 70 | bounding_boxes = calibrate_box(bounding_boxes[:, 0:5], bounding_boxes[:, 5:]) 71 | # shape [n_boxes, 5] 72 | 73 | bounding_boxes = convert_to_square(bounding_boxes) 74 | bounding_boxes[:, 0:4] = np.round(bounding_boxes[:, 0:4]) 75 | 76 | # STAGE 2 77 | 78 | img_boxes = get_image_boxes(bounding_boxes, image, size=24) 79 | img_boxes = torch.FloatTensor(img_boxes) 80 | 81 | output = rnet(img_boxes) 82 | offsets = output[0].data.numpy() # shape [n_boxes, 4] 83 | probs = output[1].data.numpy() # shape [n_boxes, 2] 84 | 85 | keep = np.where(probs[:, 1] > thresholds[1])[0] 86 | bounding_boxes = bounding_boxes[keep] 87 | bounding_boxes[:, 4] = probs[keep, 1].reshape((-1,)) 88 | offsets = offsets[keep] 89 | 90 | keep = nms(bounding_boxes, nms_thresholds[1]) 91 | bounding_boxes = bounding_boxes[keep] 92 | bounding_boxes = calibrate_box(bounding_boxes, offsets[keep]) 93 | bounding_boxes = convert_to_square(bounding_boxes) 94 | bounding_boxes[:, 0:4] = np.round(bounding_boxes[:, 0:4]) 95 | 96 | # STAGE 3 97 | 98 | img_boxes = get_image_boxes(bounding_boxes, image, size=48) 99 | if len(img_boxes) == 0: 100 | return [], [] 101 | img_boxes = torch.FloatTensor(img_boxes) 102 | output = onet(img_boxes) 103 | landmarks = output[0].data.numpy() # shape [n_boxes, 10] 104 | offsets = output[1].data.numpy() # shape [n_boxes, 4] 105 | probs = output[2].data.numpy() # shape [n_boxes, 2] 106 | 107 | keep = np.where(probs[:, 1] > thresholds[2])[0] 108 | bounding_boxes = bounding_boxes[keep] 109 | bounding_boxes[:, 4] = probs[keep, 1].reshape((-1,)) 110 | offsets = offsets[keep] 111 | landmarks = landmarks[keep] 112 | 113 | # compute landmark points 114 | width = bounding_boxes[:, 2] - bounding_boxes[:, 0] + 1.0 115 | height = bounding_boxes[:, 3] - bounding_boxes[:, 1] + 1.0 116 | xmin, ymin = bounding_boxes[:, 0], bounding_boxes[:, 1] 117 | landmarks[:, 0:5] = np.expand_dims(xmin, 1) + np.expand_dims(width, 1) * landmarks[:, 0:5] 118 | landmarks[:, 5:10] = np.expand_dims(ymin, 1) + np.expand_dims(height, 1) * landmarks[:, 5:10] 119 | 120 | bounding_boxes = calibrate_box(bounding_boxes, offsets) 121 | keep = nms(bounding_boxes, nms_thresholds[2], mode='min') 122 | bounding_boxes = bounding_boxes[keep] 123 | landmarks = landmarks[keep] 124 | 125 | return bounding_boxes, landmarks 126 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/first_stage.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import math 3 | from PIL import Image 4 | import numpy as np 5 | from .box_utils import nms, _preprocess 6 | 7 | # device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 8 | device = 'cuda:0' 9 | 10 | 11 | def run_first_stage(image, net, scale, threshold): 12 | """Run P-Net, generate bounding boxes, and do NMS. 13 | 14 | Arguments: 15 | image: an instance of PIL.Image. 16 | net: an instance of pytorch's nn.Module, P-Net. 17 | scale: a float number, 18 | scale width and height of the image by this number. 19 | threshold: a float number, 20 | threshold on the probability of a face when generating 21 | bounding boxes from predictions of the net. 22 | 23 | Returns: 24 | a float numpy array of shape [n_boxes, 9], 25 | bounding boxes with scores and offsets (4 + 1 + 4). 26 | """ 27 | 28 | # scale the image and convert it to a float array 29 | width, height = image.size 30 | sw, sh = math.ceil(width * scale), math.ceil(height * scale) 31 | img = image.resize((sw, sh), Image.BILINEAR) 32 | img = np.asarray(img, 'float32') 33 | 34 | img = torch.FloatTensor(_preprocess(img)).to(device) 35 | with torch.no_grad(): 36 | output = net(img) 37 | probs = output[1].cpu().data.numpy()[0, 1, :, :] 38 | offsets = output[0].cpu().data.numpy() 39 | # probs: probability of a face at each sliding window 40 | # offsets: transformations to true bounding boxes 41 | 42 | boxes = _generate_bboxes(probs, offsets, scale, threshold) 43 | if len(boxes) == 0: 44 | return None 45 | 46 | keep = nms(boxes[:, 0:5], overlap_threshold=0.5) 47 | return boxes[keep] 48 | 49 | 50 | def _generate_bboxes(probs, offsets, scale, threshold): 51 | """Generate bounding boxes at places 52 | where there is probably a face. 53 | 54 | Arguments: 55 | probs: a float numpy array of shape [n, m]. 56 | offsets: a float numpy array of shape [1, 4, n, m]. 57 | scale: a float number, 58 | width and height of the image were scaled by this number. 59 | threshold: a float number. 60 | 61 | Returns: 62 | a float numpy array of shape [n_boxes, 9] 63 | """ 64 | 65 | # applying P-Net is equivalent, in some sense, to 66 | # moving 12x12 window with stride 2 67 | stride = 2 68 | cell_size = 12 69 | 70 | # indices of boxes where there is probably a face 71 | inds = np.where(probs > threshold) 72 | 73 | if inds[0].size == 0: 74 | return np.array([]) 75 | 76 | # transformations of bounding boxes 77 | tx1, ty1, tx2, ty2 = [offsets[0, i, inds[0], inds[1]] for i in range(4)] 78 | # they are defined as: 79 | # w = x2 - x1 + 1 80 | # h = y2 - y1 + 1 81 | # x1_true = x1 + tx1*w 82 | # x2_true = x2 + tx2*w 83 | # y1_true = y1 + ty1*h 84 | # y2_true = y2 + ty2*h 85 | 86 | offsets = np.array([tx1, ty1, tx2, ty2]) 87 | score = probs[inds[0], inds[1]] 88 | 89 | # P-Net is applied to scaled images 90 | # so we need to rescale bounding boxes back 91 | bounding_boxes = np.vstack([ 92 | np.round((stride * inds[1] + 1.0) / scale), 93 | np.round((stride * inds[0] + 1.0) / scale), 94 | np.round((stride * inds[1] + 1.0 + cell_size) / scale), 95 | np.round((stride * inds[0] + 1.0 + cell_size) / scale), 96 | score, offsets 97 | ]) 98 | # why one is added? 99 | 100 | return bounding_boxes.T 101 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/get_nets.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from collections import OrderedDict 5 | import numpy as np 6 | 7 | from configs.paths_config import model_paths 8 | PNET_PATH = model_paths["mtcnn_pnet"] 9 | ONET_PATH = model_paths["mtcnn_onet"] 10 | RNET_PATH = model_paths["mtcnn_rnet"] 11 | 12 | 13 | class Flatten(nn.Module): 14 | 15 | def __init__(self): 16 | super(Flatten, self).__init__() 17 | 18 | def forward(self, x): 19 | """ 20 | Arguments: 21 | x: a float tensor with shape [batch_size, c, h, w]. 22 | Returns: 23 | a float tensor with shape [batch_size, c*h*w]. 24 | """ 25 | 26 | # without this pretrained model isn't working 27 | x = x.transpose(3, 2).contiguous() 28 | 29 | return x.view(x.size(0), -1) 30 | 31 | 32 | class PNet(nn.Module): 33 | 34 | def __init__(self): 35 | super().__init__() 36 | 37 | # suppose we have input with size HxW, then 38 | # after first layer: H - 2, 39 | # after pool: ceil((H - 2)/2), 40 | # after second conv: ceil((H - 2)/2) - 2, 41 | # after last conv: ceil((H - 2)/2) - 4, 42 | # and the same for W 43 | 44 | self.features = nn.Sequential(OrderedDict([ 45 | ('conv1', nn.Conv2d(3, 10, 3, 1)), 46 | ('prelu1', nn.PReLU(10)), 47 | ('pool1', nn.MaxPool2d(2, 2, ceil_mode=True)), 48 | 49 | ('conv2', nn.Conv2d(10, 16, 3, 1)), 50 | ('prelu2', nn.PReLU(16)), 51 | 52 | ('conv3', nn.Conv2d(16, 32, 3, 1)), 53 | ('prelu3', nn.PReLU(32)) 54 | ])) 55 | 56 | self.conv4_1 = nn.Conv2d(32, 2, 1, 1) 57 | self.conv4_2 = nn.Conv2d(32, 4, 1, 1) 58 | 59 | weights = np.load(PNET_PATH, allow_pickle=True)[()] 60 | for n, p in self.named_parameters(): 61 | p.data = torch.FloatTensor(weights[n]) 62 | 63 | def forward(self, x): 64 | """ 65 | Arguments: 66 | x: a float tensor with shape [batch_size, 3, h, w]. 67 | Returns: 68 | b: a float tensor with shape [batch_size, 4, h', w']. 69 | a: a float tensor with shape [batch_size, 2, h', w']. 70 | """ 71 | x = self.features(x) 72 | a = self.conv4_1(x) 73 | b = self.conv4_2(x) 74 | a = F.softmax(a, dim=-1) 75 | return b, a 76 | 77 | 78 | class RNet(nn.Module): 79 | 80 | def __init__(self): 81 | super().__init__() 82 | 83 | self.features = nn.Sequential(OrderedDict([ 84 | ('conv1', nn.Conv2d(3, 28, 3, 1)), 85 | ('prelu1', nn.PReLU(28)), 86 | ('pool1', nn.MaxPool2d(3, 2, ceil_mode=True)), 87 | 88 | ('conv2', nn.Conv2d(28, 48, 3, 1)), 89 | ('prelu2', nn.PReLU(48)), 90 | ('pool2', nn.MaxPool2d(3, 2, ceil_mode=True)), 91 | 92 | ('conv3', nn.Conv2d(48, 64, 2, 1)), 93 | ('prelu3', nn.PReLU(64)), 94 | 95 | ('flatten', Flatten()), 96 | ('conv4', nn.Linear(576, 128)), 97 | ('prelu4', nn.PReLU(128)) 98 | ])) 99 | 100 | self.conv5_1 = nn.Linear(128, 2) 101 | self.conv5_2 = nn.Linear(128, 4) 102 | 103 | weights = np.load(RNET_PATH, allow_pickle=True)[()] 104 | for n, p in self.named_parameters(): 105 | p.data = torch.FloatTensor(weights[n]) 106 | 107 | def forward(self, x): 108 | """ 109 | Arguments: 110 | x: a float tensor with shape [batch_size, 3, h, w]. 111 | Returns: 112 | b: a float tensor with shape [batch_size, 4]. 113 | a: a float tensor with shape [batch_size, 2]. 114 | """ 115 | x = self.features(x) 116 | a = self.conv5_1(x) 117 | b = self.conv5_2(x) 118 | a = F.softmax(a, dim=-1) 119 | return b, a 120 | 121 | 122 | class ONet(nn.Module): 123 | 124 | def __init__(self): 125 | super().__init__() 126 | 127 | self.features = nn.Sequential(OrderedDict([ 128 | ('conv1', nn.Conv2d(3, 32, 3, 1)), 129 | ('prelu1', nn.PReLU(32)), 130 | ('pool1', nn.MaxPool2d(3, 2, ceil_mode=True)), 131 | 132 | ('conv2', nn.Conv2d(32, 64, 3, 1)), 133 | ('prelu2', nn.PReLU(64)), 134 | ('pool2', nn.MaxPool2d(3, 2, ceil_mode=True)), 135 | 136 | ('conv3', nn.Conv2d(64, 64, 3, 1)), 137 | ('prelu3', nn.PReLU(64)), 138 | ('pool3', nn.MaxPool2d(2, 2, ceil_mode=True)), 139 | 140 | ('conv4', nn.Conv2d(64, 128, 2, 1)), 141 | ('prelu4', nn.PReLU(128)), 142 | 143 | ('flatten', Flatten()), 144 | ('conv5', nn.Linear(1152, 256)), 145 | ('drop5', nn.Dropout(0.25)), 146 | ('prelu5', nn.PReLU(256)), 147 | ])) 148 | 149 | self.conv6_1 = nn.Linear(256, 2) 150 | self.conv6_2 = nn.Linear(256, 4) 151 | self.conv6_3 = nn.Linear(256, 10) 152 | 153 | weights = np.load(ONET_PATH, allow_pickle=True)[()] 154 | for n, p in self.named_parameters(): 155 | p.data = torch.FloatTensor(weights[n]) 156 | 157 | def forward(self, x): 158 | """ 159 | Arguments: 160 | x: a float tensor with shape [batch_size, 3, h, w]. 161 | Returns: 162 | c: a float tensor with shape [batch_size, 10]. 163 | b: a float tensor with shape [batch_size, 4]. 164 | a: a float tensor with shape [batch_size, 2]. 165 | """ 166 | x = self.features(x) 167 | a = self.conv6_1(x) 168 | b = self.conv6_2(x) 169 | c = self.conv6_3(x) 170 | a = F.softmax(a, dim=-1) 171 | return c, b, a 172 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/matlab_cp2tform.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | Created on Tue Jul 11 06:54:28 2017 4 | 5 | @author: zhaoyafei 6 | """ 7 | 8 | import numpy as np 9 | from numpy.linalg import inv, norm, lstsq 10 | from numpy.linalg import matrix_rank as rank 11 | 12 | 13 | class MatlabCp2tormException(Exception): 14 | def __str__(self): 15 | return 'In File {}:{}'.format( 16 | __file__, super.__str__(self)) 17 | 18 | 19 | def tformfwd(trans, uv): 20 | """ 21 | Function: 22 | ---------- 23 | apply affine transform 'trans' to uv 24 | 25 | Parameters: 26 | ---------- 27 | @trans: 3x3 np.array 28 | transform matrix 29 | @uv: Kx2 np.array 30 | each row is a pair of coordinates (x, y) 31 | 32 | Returns: 33 | ---------- 34 | @xy: Kx2 np.array 35 | each row is a pair of transformed coordinates (x, y) 36 | """ 37 | uv = np.hstack(( 38 | uv, np.ones((uv.shape[0], 1)) 39 | )) 40 | xy = np.dot(uv, trans) 41 | xy = xy[:, 0:-1] 42 | return xy 43 | 44 | 45 | def tforminv(trans, uv): 46 | """ 47 | Function: 48 | ---------- 49 | apply the inverse of affine transform 'trans' to uv 50 | 51 | Parameters: 52 | ---------- 53 | @trans: 3x3 np.array 54 | transform matrix 55 | @uv: Kx2 np.array 56 | each row is a pair of coordinates (x, y) 57 | 58 | Returns: 59 | ---------- 60 | @xy: Kx2 np.array 61 | each row is a pair of inverse-transformed coordinates (x, y) 62 | """ 63 | Tinv = inv(trans) 64 | xy = tformfwd(Tinv, uv) 65 | return xy 66 | 67 | 68 | def findNonreflectiveSimilarity(uv, xy, options=None): 69 | options = {'K': 2} 70 | 71 | K = options['K'] 72 | M = xy.shape[0] 73 | x = xy[:, 0].reshape((-1, 1)) # use reshape to keep a column vector 74 | y = xy[:, 1].reshape((-1, 1)) # use reshape to keep a column vector 75 | # print('--->x, y:\n', x, y 76 | 77 | tmp1 = np.hstack((x, y, np.ones((M, 1)), np.zeros((M, 1)))) 78 | tmp2 = np.hstack((y, -x, np.zeros((M, 1)), np.ones((M, 1)))) 79 | X = np.vstack((tmp1, tmp2)) 80 | # print('--->X.shape: ', X.shape 81 | # print('X:\n', X 82 | 83 | u = uv[:, 0].reshape((-1, 1)) # use reshape to keep a column vector 84 | v = uv[:, 1].reshape((-1, 1)) # use reshape to keep a column vector 85 | U = np.vstack((u, v)) 86 | # print('--->U.shape: ', U.shape 87 | # print('U:\n', U 88 | 89 | # We know that X * r = U 90 | if rank(X) >= 2 * K: 91 | r, _, _, _ = lstsq(X, U, rcond=None) # Make sure this is what I want 92 | r = np.squeeze(r) 93 | else: 94 | raise Exception('cp2tform:twoUniquePointsReq') 95 | 96 | # print('--->r:\n', r 97 | 98 | sc = r[0] 99 | ss = r[1] 100 | tx = r[2] 101 | ty = r[3] 102 | 103 | Tinv = np.array([ 104 | [sc, -ss, 0], 105 | [ss, sc, 0], 106 | [tx, ty, 1] 107 | ]) 108 | 109 | # print('--->Tinv:\n', Tinv 110 | 111 | T = inv(Tinv) 112 | # print('--->T:\n', T 113 | 114 | T[:, 2] = np.array([0, 0, 1]) 115 | 116 | return T, Tinv 117 | 118 | 119 | def findSimilarity(uv, xy, options=None): 120 | options = {'K': 2} 121 | 122 | # uv = np.array(uv) 123 | # xy = np.array(xy) 124 | 125 | # Solve for trans1 126 | trans1, trans1_inv = findNonreflectiveSimilarity(uv, xy, options) 127 | 128 | # Solve for trans2 129 | 130 | # manually reflect the xy data across the Y-axis 131 | xyR = xy 132 | xyR[:, 0] = -1 * xyR[:, 0] 133 | 134 | trans2r, trans2r_inv = findNonreflectiveSimilarity(uv, xyR, options) 135 | 136 | # manually reflect the tform to undo the reflection done on xyR 137 | TreflectY = np.array([ 138 | [-1, 0, 0], 139 | [0, 1, 0], 140 | [0, 0, 1] 141 | ]) 142 | 143 | trans2 = np.dot(trans2r, TreflectY) 144 | 145 | # Figure out if trans1 or trans2 is better 146 | xy1 = tformfwd(trans1, uv) 147 | norm1 = norm(xy1 - xy) 148 | 149 | xy2 = tformfwd(trans2, uv) 150 | norm2 = norm(xy2 - xy) 151 | 152 | if norm1 <= norm2: 153 | return trans1, trans1_inv 154 | else: 155 | trans2_inv = inv(trans2) 156 | return trans2, trans2_inv 157 | 158 | 159 | def get_similarity_transform(src_pts, dst_pts, reflective=True): 160 | """ 161 | Function: 162 | ---------- 163 | Find Similarity Transform Matrix 'trans': 164 | u = src_pts[:, 0] 165 | v = src_pts[:, 1] 166 | x = dst_pts[:, 0] 167 | y = dst_pts[:, 1] 168 | [x, y, 1] = [u, v, 1] * trans 169 | 170 | Parameters: 171 | ---------- 172 | @src_pts: Kx2 np.array 173 | source points, each row is a pair of coordinates (x, y) 174 | @dst_pts: Kx2 np.array 175 | destination points, each row is a pair of transformed 176 | coordinates (x, y) 177 | @reflective: True or False 178 | if True: 179 | use reflective similarity transform 180 | else: 181 | use non-reflective similarity transform 182 | 183 | Returns: 184 | ---------- 185 | @trans: 3x3 np.array 186 | transform matrix from uv to xy 187 | trans_inv: 3x3 np.array 188 | inverse of trans, transform matrix from xy to uv 189 | """ 190 | 191 | if reflective: 192 | trans, trans_inv = findSimilarity(src_pts, dst_pts) 193 | else: 194 | trans, trans_inv = findNonreflectiveSimilarity(src_pts, dst_pts) 195 | 196 | return trans, trans_inv 197 | 198 | 199 | def cvt_tform_mat_for_cv2(trans): 200 | """ 201 | Function: 202 | ---------- 203 | Convert Transform Matrix 'trans' into 'cv2_trans' which could be 204 | directly used by cv2.warpAffine(): 205 | u = src_pts[:, 0] 206 | v = src_pts[:, 1] 207 | x = dst_pts[:, 0] 208 | y = dst_pts[:, 1] 209 | [x, y].T = cv_trans * [u, v, 1].T 210 | 211 | Parameters: 212 | ---------- 213 | @trans: 3x3 np.array 214 | transform matrix from uv to xy 215 | 216 | Returns: 217 | ---------- 218 | @cv2_trans: 2x3 np.array 219 | transform matrix from src_pts to dst_pts, could be directly used 220 | for cv2.warpAffine() 221 | """ 222 | cv2_trans = trans[:, 0:2].T 223 | 224 | return cv2_trans 225 | 226 | 227 | def get_similarity_transform_for_cv2(src_pts, dst_pts, reflective=True): 228 | """ 229 | Function: 230 | ---------- 231 | Find Similarity Transform Matrix 'cv2_trans' which could be 232 | directly used by cv2.warpAffine(): 233 | u = src_pts[:, 0] 234 | v = src_pts[:, 1] 235 | x = dst_pts[:, 0] 236 | y = dst_pts[:, 1] 237 | [x, y].T = cv_trans * [u, v, 1].T 238 | 239 | Parameters: 240 | ---------- 241 | @src_pts: Kx2 np.array 242 | source points, each row is a pair of coordinates (x, y) 243 | @dst_pts: Kx2 np.array 244 | destination points, each row is a pair of transformed 245 | coordinates (x, y) 246 | reflective: True or False 247 | if True: 248 | use reflective similarity transform 249 | else: 250 | use non-reflective similarity transform 251 | 252 | Returns: 253 | ---------- 254 | @cv2_trans: 2x3 np.array 255 | transform matrix from src_pts to dst_pts, could be directly used 256 | for cv2.warpAffine() 257 | """ 258 | trans, trans_inv = get_similarity_transform(src_pts, dst_pts, reflective) 259 | cv2_trans = cvt_tform_mat_for_cv2(trans) 260 | 261 | return cv2_trans 262 | 263 | 264 | if __name__ == '__main__': 265 | """ 266 | u = [0, 6, -2] 267 | v = [0, 3, 5] 268 | x = [-1, 0, 4] 269 | y = [-1, -10, 4] 270 | 271 | # In Matlab, run: 272 | # 273 | # uv = [u'; v']; 274 | # xy = [x'; y']; 275 | # tform_sim=cp2tform(uv,xy,'similarity'); 276 | # 277 | # trans = tform_sim.tdata.T 278 | # ans = 279 | # -0.0764 -1.6190 0 280 | # 1.6190 -0.0764 0 281 | # -3.2156 0.0290 1.0000 282 | # trans_inv = tform_sim.tdata.Tinv 283 | # ans = 284 | # 285 | # -0.0291 0.6163 0 286 | # -0.6163 -0.0291 0 287 | # -0.0756 1.9826 1.0000 288 | # xy_m=tformfwd(tform_sim, u,v) 289 | # 290 | # xy_m = 291 | # 292 | # -3.2156 0.0290 293 | # 1.1833 -9.9143 294 | # 5.0323 2.8853 295 | # uv_m=tforminv(tform_sim, x,y) 296 | # 297 | # uv_m = 298 | # 299 | # 0.5698 1.3953 300 | # 6.0872 2.2733 301 | # -2.6570 4.3314 302 | """ 303 | u = [0, 6, -2] 304 | v = [0, 3, 5] 305 | x = [-1, 0, 4] 306 | y = [-1, -10, 4] 307 | 308 | uv = np.array((u, v)).T 309 | xy = np.array((x, y)).T 310 | 311 | print('\n--->uv:') 312 | print(uv) 313 | print('\n--->xy:') 314 | print(xy) 315 | 316 | trans, trans_inv = get_similarity_transform(uv, xy) 317 | 318 | print('\n--->trans matrix:') 319 | print(trans) 320 | 321 | print('\n--->trans_inv matrix:') 322 | print(trans_inv) 323 | 324 | print('\n---> apply transform to uv') 325 | print('\nxy_m = uv_augmented * trans') 326 | uv_aug = np.hstack(( 327 | uv, np.ones((uv.shape[0], 1)) 328 | )) 329 | xy_m = np.dot(uv_aug, trans) 330 | print(xy_m) 331 | 332 | print('\nxy_m = tformfwd(trans, uv)') 333 | xy_m = tformfwd(trans, uv) 334 | print(xy_m) 335 | 336 | print('\n---> apply inverse transform to xy') 337 | print('\nuv_m = xy_augmented * trans_inv') 338 | xy_aug = np.hstack(( 339 | xy, np.ones((xy.shape[0], 1)) 340 | )) 341 | uv_m = np.dot(xy_aug, trans_inv) 342 | print(uv_m) 343 | 344 | print('\nuv_m = tformfwd(trans_inv, xy)') 345 | uv_m = tformfwd(trans_inv, xy) 346 | print(uv_m) 347 | 348 | uv_m = tforminv(trans, xy) 349 | print('\nuv_m = tforminv(trans, xy)') 350 | print(uv_m) 351 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/visualization_utils.py: -------------------------------------------------------------------------------- 1 | from PIL import ImageDraw 2 | 3 | 4 | def show_bboxes(img, bounding_boxes, facial_landmarks=[]): 5 | """Draw bounding boxes and facial landmarks. 6 | 7 | Arguments: 8 | img: an instance of PIL.Image. 9 | bounding_boxes: a float numpy array of shape [n, 5]. 10 | facial_landmarks: a float numpy array of shape [n, 10]. 11 | 12 | Returns: 13 | an instance of PIL.Image. 14 | """ 15 | 16 | img_copy = img.copy() 17 | draw = ImageDraw.Draw(img_copy) 18 | 19 | for b in bounding_boxes: 20 | draw.rectangle([ 21 | (b[0], b[1]), (b[2], b[3]) 22 | ], outline='white') 23 | 24 | for p in facial_landmarks: 25 | for i in range(5): 26 | draw.ellipse([ 27 | (p[i] - 1.0, p[i + 5] - 1.0), 28 | (p[i] + 1.0, p[i + 5] + 1.0) 29 | ], outline='blue') 30 | 31 | return img_copy 32 | -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/weights/onet.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/mtcnn/mtcnn_pytorch/src/weights/onet.npy -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/weights/pnet.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/mtcnn/mtcnn_pytorch/src/weights/pnet.npy -------------------------------------------------------------------------------- /models/mtcnn/mtcnn_pytorch/src/weights/rnet.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/mtcnn/mtcnn_pytorch/src/weights/rnet.npy -------------------------------------------------------------------------------- /models/psp.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file defines the core research contribution 3 | """ 4 | import math 5 | import torch 6 | from torch import nn 7 | 8 | from models.stylegan2.model import Generator 9 | from configs.paths_config import model_paths 10 | from models.encoders import fpn_encoders, restyle_psp_encoders 11 | from utils.model_utils import RESNET_MAPPING 12 | 13 | 14 | class pSp(nn.Module): 15 | 16 | def __init__(self, opts): 17 | super(pSp, self).__init__() 18 | self.set_opts(opts) 19 | self.n_styles = int(math.log(self.opts.output_size, 2)) * 2 - 2 20 | # Define architecture 21 | self.encoder = self.set_encoder() 22 | self.decoder = Generator(self.opts.output_size, 512, 8, channel_multiplier=2) 23 | self.face_pool = torch.nn.AdaptiveAvgPool2d((256, 256)) 24 | # Load weights if needed 25 | self.load_weights() 26 | 27 | def set_encoder(self): 28 | if self.opts.encoder_type == 'GradualStyleEncoder': 29 | encoder = fpn_encoders.GradualStyleEncoder(50, 'ir_se', self.n_styles, self.opts) 30 | elif self.opts.encoder_type == 'ResNetGradualStyleEncoder': 31 | encoder = fpn_encoders.ResNetGradualStyleEncoder(self.n_styles, self.opts) 32 | elif self.opts.encoder_type == 'BackboneEncoder': 33 | encoder = restyle_psp_encoders.BackboneEncoder(50, 'ir_se', self.n_styles, self.opts) 34 | elif self.opts.encoder_type == 'ResNetBackboneEncoder': 35 | encoder = restyle_psp_encoders.ResNetBackboneEncoder(self.n_styles, self.opts) 36 | else: 37 | raise Exception(f'{self.opts.encoder_type} is not a valid encoders') 38 | return encoder 39 | 40 | def load_weights(self): 41 | if self.opts.checkpoint_path is not None: 42 | print(f'Loading ReStyle pSp from checkpoint: {self.opts.checkpoint_path}') 43 | ckpt = torch.load(self.opts.checkpoint_path, map_location='cpu') 44 | self.encoder.load_state_dict(self.__get_keys(ckpt, 'encoder'), strict=False) 45 | self.decoder.load_state_dict(self.__get_keys(ckpt, 'decoder'), strict=True) 46 | self.__load_latent_avg(ckpt) 47 | else: 48 | encoder_ckpt = self.__get_encoder_checkpoint() 49 | self.encoder.load_state_dict(encoder_ckpt, strict=False) 50 | print(f'Loading decoder weights from pretrained path: {self.opts.stylegan_weights}') 51 | ckpt = torch.load(self.opts.stylegan_weights) 52 | self.decoder.load_state_dict(ckpt['g_ema'], strict=True) 53 | self.__load_latent_avg(ckpt, repeat=self.n_styles) 54 | 55 | def forward(self, x, latent=None, resize=True, latent_mask=None, input_code=False, randomize_noise=True, 56 | inject_latent=None, return_latents=False, alpha=None, average_code=False, input_is_full=False): 57 | if input_code: 58 | codes = x 59 | else: 60 | codes = self.encoder(x) 61 | # residual step 62 | if x.shape[1] == 6 and latent is not None: 63 | # learn error with respect to previous iteration 64 | codes = codes + latent 65 | else: 66 | # first iteration is with respect to the avg latent code 67 | codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1) 68 | 69 | if latent_mask is not None: 70 | for i in latent_mask: 71 | if inject_latent is not None: 72 | if alpha is not None: 73 | codes[:, i] = alpha * inject_latent[:, i] + (1 - alpha) * codes[:, i] 74 | else: 75 | codes[:, i] = inject_latent[:, i] 76 | else: 77 | codes[:, i] = 0 78 | 79 | if average_code: 80 | input_is_latent = True 81 | else: 82 | input_is_latent = (not input_code) or (input_is_full) 83 | 84 | images, result_latent = self.decoder([codes], 85 | input_is_latent=input_is_latent, 86 | randomize_noise=randomize_noise, 87 | return_latents=return_latents) 88 | 89 | if resize: 90 | images = self.face_pool(images) 91 | 92 | if return_latents: 93 | return images, result_latent 94 | else: 95 | return images 96 | 97 | def set_opts(self, opts): 98 | self.opts = opts 99 | 100 | def __load_latent_avg(self, ckpt, repeat=None): 101 | if 'latent_avg' in ckpt: 102 | self.latent_avg = ckpt['latent_avg'].to(self.opts.device) 103 | if repeat is not None: 104 | self.latent_avg = self.latent_avg.repeat(repeat, 1) 105 | else: 106 | self.latent_avg = None 107 | 108 | def __get_encoder_checkpoint(self): 109 | if "ffhq" in self.opts.dataset_type: 110 | print('Loading encoders weights from irse50!') 111 | encoder_ckpt = torch.load(model_paths['ir_se50']) 112 | # Transfer the RGB input of the irse50 network to the first 3 input channels of pSp's encoder 113 | if self.opts.input_nc != 3: 114 | shape = encoder_ckpt['input_layer.0.weight'].shape 115 | altered_input_layer = torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 116 | altered_input_layer[:, :3, :, :] = encoder_ckpt['input_layer.0.weight'] 117 | encoder_ckpt['input_layer.0.weight'] = altered_input_layer 118 | return encoder_ckpt 119 | else: 120 | print('Loading encoders weights from resnet34!') 121 | encoder_ckpt = torch.load(model_paths['resnet34']) 122 | # Transfer the RGB input of the resnet34 network to the first 3 input channels of pSp's encoder 123 | if self.opts.input_nc != 3: 124 | shape = encoder_ckpt['conv1.weight'].shape 125 | altered_input_layer = torch.randn(shape[0], self.opts.input_nc, shape[2], shape[3], dtype=torch.float32) 126 | altered_input_layer[:, :3, :, :] = encoder_ckpt['conv1.weight'] 127 | encoder_ckpt['conv1.weight'] = altered_input_layer 128 | mapped_encoder_ckpt = dict(encoder_ckpt) 129 | for p, v in encoder_ckpt.items(): 130 | for original_name, psp_name in RESNET_MAPPING.items(): 131 | if original_name in p: 132 | mapped_encoder_ckpt[p.replace(original_name, psp_name)] = v 133 | mapped_encoder_ckpt.pop(p) 134 | return encoder_ckpt 135 | 136 | @staticmethod 137 | def __get_keys(d, name): 138 | if 'state_dict' in d: 139 | d = d['state_dict'] 140 | d_filt = {k[len(name) + 1:]: v for k, v in d.items() if k[:len(name)] == name} 141 | return d_filt 142 | -------------------------------------------------------------------------------- /models/stylegan2/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/models/stylegan2/__init__.py -------------------------------------------------------------------------------- /models/stylegan2/op/__init__.py: -------------------------------------------------------------------------------- 1 | from .fused_act import FusedLeakyReLU, fused_leaky_relu 2 | from .upfirdn2d import upfirdn2d 3 | -------------------------------------------------------------------------------- /models/stylegan2/op/fused_act.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | import torch 4 | from torch import nn 5 | from torch.autograd import Function 6 | from torch.utils.cpp_extension import load 7 | 8 | module_path = os.path.dirname(__file__) 9 | fused = load( 10 | 'fused', 11 | sources=[ 12 | os.path.join(module_path, 'fused_bias_act.cpp'), 13 | os.path.join(module_path, 'fused_bias_act_kernel.cu'), 14 | ], 15 | ) 16 | 17 | 18 | class FusedLeakyReLUFunctionBackward(Function): 19 | @staticmethod 20 | def forward(ctx, grad_output, out, negative_slope, scale): 21 | ctx.save_for_backward(out) 22 | ctx.negative_slope = negative_slope 23 | ctx.scale = scale 24 | 25 | empty = grad_output.new_empty(0) 26 | 27 | grad_input = fused.fused_bias_act( 28 | grad_output, empty, out, 3, 1, negative_slope, scale 29 | ) 30 | 31 | dim = [0] 32 | 33 | if grad_input.ndim > 2: 34 | dim += list(range(2, grad_input.ndim)) 35 | 36 | grad_bias = grad_input.sum(dim).detach() 37 | 38 | return grad_input, grad_bias 39 | 40 | @staticmethod 41 | def backward(ctx, gradgrad_input, gradgrad_bias): 42 | out, = ctx.saved_tensors 43 | gradgrad_out = fused.fused_bias_act( 44 | gradgrad_input, gradgrad_bias, out, 3, 1, ctx.negative_slope, ctx.scale 45 | ) 46 | 47 | return gradgrad_out, None, None, None 48 | 49 | 50 | class FusedLeakyReLUFunction(Function): 51 | @staticmethod 52 | def forward(ctx, input, bias, negative_slope, scale): 53 | empty = input.new_empty(0) 54 | out = fused.fused_bias_act(input, bias, empty, 3, 0, negative_slope, scale) 55 | ctx.save_for_backward(out) 56 | ctx.negative_slope = negative_slope 57 | ctx.scale = scale 58 | 59 | return out 60 | 61 | @staticmethod 62 | def backward(ctx, grad_output): 63 | out, = ctx.saved_tensors 64 | 65 | grad_input, grad_bias = FusedLeakyReLUFunctionBackward.apply( 66 | grad_output, out, ctx.negative_slope, ctx.scale 67 | ) 68 | 69 | return grad_input, grad_bias, None, None 70 | 71 | 72 | class FusedLeakyReLU(nn.Module): 73 | def __init__(self, channel, negative_slope=0.2, scale=2 ** 0.5): 74 | super().__init__() 75 | 76 | self.bias = nn.Parameter(torch.zeros(channel)) 77 | self.negative_slope = negative_slope 78 | self.scale = scale 79 | 80 | def forward(self, input): 81 | return fused_leaky_relu(input, self.bias, self.negative_slope, self.scale) 82 | 83 | 84 | def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5): 85 | return FusedLeakyReLUFunction.apply(input, bias, negative_slope, scale) 86 | -------------------------------------------------------------------------------- /models/stylegan2/op/fused_bias_act.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | 3 | 4 | torch::Tensor fused_bias_act_op(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer, 5 | int act, int grad, float alpha, float scale); 6 | 7 | #define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") 8 | #define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") 9 | #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) 10 | 11 | torch::Tensor fused_bias_act(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer, 12 | int act, int grad, float alpha, float scale) { 13 | CHECK_CUDA(input); 14 | CHECK_CUDA(bias); 15 | 16 | return fused_bias_act_op(input, bias, refer, act, grad, alpha, scale); 17 | } 18 | 19 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 20 | m.def("fused_bias_act", &fused_bias_act, "fused bias act (CUDA)"); 21 | } -------------------------------------------------------------------------------- /models/stylegan2/op/fused_bias_act_kernel.cu: -------------------------------------------------------------------------------- 1 | // Copyright (c) 2019, NVIDIA Corporation. All rights reserved. 2 | // 3 | // This work is made available under the Nvidia Source Code License-NC. 4 | // To view a copy of this license, visit 5 | // https://nvlabs.github.io/stylegan2/license.html 6 | 7 | #include 8 | 9 | #include 10 | #include 11 | #include 12 | #include 13 | 14 | #include 15 | #include 16 | 17 | 18 | template 19 | static __global__ void fused_bias_act_kernel(scalar_t* out, const scalar_t* p_x, const scalar_t* p_b, const scalar_t* p_ref, 20 | int act, int grad, scalar_t alpha, scalar_t scale, int loop_x, int size_x, int step_b, int size_b, int use_bias, int use_ref) { 21 | int xi = blockIdx.x * loop_x * blockDim.x + threadIdx.x; 22 | 23 | scalar_t zero = 0.0; 24 | 25 | for (int loop_idx = 0; loop_idx < loop_x && xi < size_x; loop_idx++, xi += blockDim.x) { 26 | scalar_t x = p_x[xi]; 27 | 28 | if (use_bias) { 29 | x += p_b[(xi / step_b) % size_b]; 30 | } 31 | 32 | scalar_t ref = use_ref ? p_ref[xi] : zero; 33 | 34 | scalar_t y; 35 | 36 | switch (act * 10 + grad) { 37 | default: 38 | case 10: y = x; break; 39 | case 11: y = x; break; 40 | case 12: y = 0.0; break; 41 | 42 | case 30: y = (x > 0.0) ? x : x * alpha; break; 43 | case 31: y = (ref > 0.0) ? x : x * alpha; break; 44 | case 32: y = 0.0; break; 45 | } 46 | 47 | out[xi] = y * scale; 48 | } 49 | } 50 | 51 | 52 | torch::Tensor fused_bias_act_op(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer, 53 | int act, int grad, float alpha, float scale) { 54 | int curDevice = -1; 55 | cudaGetDevice(&curDevice); 56 | cudaStream_t stream = at::cuda::getCurrentCUDAStream(curDevice); 57 | 58 | auto x = input.contiguous(); 59 | auto b = bias.contiguous(); 60 | auto ref = refer.contiguous(); 61 | 62 | int use_bias = b.numel() ? 1 : 0; 63 | int use_ref = ref.numel() ? 1 : 0; 64 | 65 | int size_x = x.numel(); 66 | int size_b = b.numel(); 67 | int step_b = 1; 68 | 69 | for (int i = 1 + 1; i < x.dim(); i++) { 70 | step_b *= x.size(i); 71 | } 72 | 73 | int loop_x = 4; 74 | int block_size = 4 * 32; 75 | int grid_size = (size_x - 1) / (loop_x * block_size) + 1; 76 | 77 | auto y = torch::empty_like(x); 78 | 79 | AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "fused_bias_act_kernel", [&] { 80 | fused_bias_act_kernel<<>>( 81 | y.data_ptr(), 82 | x.data_ptr(), 83 | b.data_ptr(), 84 | ref.data_ptr(), 85 | act, 86 | grad, 87 | alpha, 88 | scale, 89 | loop_x, 90 | size_x, 91 | step_b, 92 | size_b, 93 | use_bias, 94 | use_ref 95 | ); 96 | }); 97 | 98 | return y; 99 | } -------------------------------------------------------------------------------- /models/stylegan2/op/upfirdn2d.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | 3 | 4 | torch::Tensor upfirdn2d_op(const torch::Tensor& input, const torch::Tensor& kernel, 5 | int up_x, int up_y, int down_x, int down_y, 6 | int pad_x0, int pad_x1, int pad_y0, int pad_y1); 7 | 8 | #define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") 9 | #define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") 10 | #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) 11 | 12 | torch::Tensor upfirdn2d(const torch::Tensor& input, const torch::Tensor& kernel, 13 | int up_x, int up_y, int down_x, int down_y, 14 | int pad_x0, int pad_x1, int pad_y0, int pad_y1) { 15 | CHECK_CUDA(input); 16 | CHECK_CUDA(kernel); 17 | 18 | return upfirdn2d_op(input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1); 19 | } 20 | 21 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 22 | m.def("upfirdn2d", &upfirdn2d, "upfirdn2d (CUDA)"); 23 | } -------------------------------------------------------------------------------- /models/stylegan2/op/upfirdn2d.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | import torch 4 | from torch.autograd import Function 5 | from torch.utils.cpp_extension import load 6 | 7 | module_path = os.path.dirname(__file__) 8 | upfirdn2d_op = load( 9 | 'upfirdn2d', 10 | sources=[ 11 | os.path.join(module_path, 'upfirdn2d.cpp'), 12 | os.path.join(module_path, 'upfirdn2d_kernel.cu'), 13 | ], 14 | ) 15 | 16 | 17 | class UpFirDn2dBackward(Function): 18 | @staticmethod 19 | def forward( 20 | ctx, grad_output, kernel, grad_kernel, up, down, pad, g_pad, in_size, out_size 21 | ): 22 | up_x, up_y = up 23 | down_x, down_y = down 24 | g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1 = g_pad 25 | 26 | grad_output = grad_output.reshape(-1, out_size[0], out_size[1], 1) 27 | 28 | grad_input = upfirdn2d_op.upfirdn2d( 29 | grad_output, 30 | grad_kernel, 31 | down_x, 32 | down_y, 33 | up_x, 34 | up_y, 35 | g_pad_x0, 36 | g_pad_x1, 37 | g_pad_y0, 38 | g_pad_y1, 39 | ) 40 | grad_input = grad_input.view(in_size[0], in_size[1], in_size[2], in_size[3]) 41 | 42 | ctx.save_for_backward(kernel) 43 | 44 | pad_x0, pad_x1, pad_y0, pad_y1 = pad 45 | 46 | ctx.up_x = up_x 47 | ctx.up_y = up_y 48 | ctx.down_x = down_x 49 | ctx.down_y = down_y 50 | ctx.pad_x0 = pad_x0 51 | ctx.pad_x1 = pad_x1 52 | ctx.pad_y0 = pad_y0 53 | ctx.pad_y1 = pad_y1 54 | ctx.in_size = in_size 55 | ctx.out_size = out_size 56 | 57 | return grad_input 58 | 59 | @staticmethod 60 | def backward(ctx, gradgrad_input): 61 | kernel, = ctx.saved_tensors 62 | 63 | gradgrad_input = gradgrad_input.reshape(-1, ctx.in_size[2], ctx.in_size[3], 1) 64 | 65 | gradgrad_out = upfirdn2d_op.upfirdn2d( 66 | gradgrad_input, 67 | kernel, 68 | ctx.up_x, 69 | ctx.up_y, 70 | ctx.down_x, 71 | ctx.down_y, 72 | ctx.pad_x0, 73 | ctx.pad_x1, 74 | ctx.pad_y0, 75 | ctx.pad_y1, 76 | ) 77 | # gradgrad_out = gradgrad_out.view(ctx.in_size[0], ctx.out_size[0], ctx.out_size[1], ctx.in_size[3]) 78 | gradgrad_out = gradgrad_out.view( 79 | ctx.in_size[0], ctx.in_size[1], ctx.out_size[0], ctx.out_size[1] 80 | ) 81 | 82 | return gradgrad_out, None, None, None, None, None, None, None, None 83 | 84 | 85 | class UpFirDn2d(Function): 86 | @staticmethod 87 | def forward(ctx, input, kernel, up, down, pad): 88 | up_x, up_y = up 89 | down_x, down_y = down 90 | pad_x0, pad_x1, pad_y0, pad_y1 = pad 91 | 92 | kernel_h, kernel_w = kernel.shape 93 | batch, channel, in_h, in_w = input.shape 94 | ctx.in_size = input.shape 95 | 96 | input = input.reshape(-1, in_h, in_w, 1) 97 | 98 | ctx.save_for_backward(kernel, torch.flip(kernel, [0, 1])) 99 | 100 | out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1 101 | out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1 102 | ctx.out_size = (out_h, out_w) 103 | 104 | ctx.up = (up_x, up_y) 105 | ctx.down = (down_x, down_y) 106 | ctx.pad = (pad_x0, pad_x1, pad_y0, pad_y1) 107 | 108 | g_pad_x0 = kernel_w - pad_x0 - 1 109 | g_pad_y0 = kernel_h - pad_y0 - 1 110 | g_pad_x1 = in_w * up_x - out_w * down_x + pad_x0 - up_x + 1 111 | g_pad_y1 = in_h * up_y - out_h * down_y + pad_y0 - up_y + 1 112 | 113 | ctx.g_pad = (g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1) 114 | 115 | out = upfirdn2d_op.upfirdn2d( 116 | input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1 117 | ) 118 | # out = out.view(major, out_h, out_w, minor) 119 | out = out.view(-1, channel, out_h, out_w) 120 | 121 | return out 122 | 123 | @staticmethod 124 | def backward(ctx, grad_output): 125 | kernel, grad_kernel = ctx.saved_tensors 126 | 127 | grad_input = UpFirDn2dBackward.apply( 128 | grad_output, 129 | kernel, 130 | grad_kernel, 131 | ctx.up, 132 | ctx.down, 133 | ctx.pad, 134 | ctx.g_pad, 135 | ctx.in_size, 136 | ctx.out_size, 137 | ) 138 | 139 | return grad_input, None, None, None, None 140 | 141 | 142 | def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)): 143 | out = UpFirDn2d.apply( 144 | input, kernel, (up, up), (down, down), (pad[0], pad[1], pad[0], pad[1]) 145 | ) 146 | 147 | return out 148 | 149 | 150 | def upfirdn2d_native( 151 | input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1 152 | ): 153 | _, in_h, in_w, minor = input.shape 154 | kernel_h, kernel_w = kernel.shape 155 | 156 | out = input.view(-1, in_h, 1, in_w, 1, minor) 157 | out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1]) 158 | out = out.view(-1, in_h * up_y, in_w * up_x, minor) 159 | 160 | out = F.pad( 161 | out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)] 162 | ) 163 | out = out[ 164 | :, 165 | max(-pad_y0, 0): out.shape[1] - max(-pad_y1, 0), 166 | max(-pad_x0, 0): out.shape[2] - max(-pad_x1, 0), 167 | :, 168 | ] 169 | 170 | out = out.permute(0, 3, 1, 2) 171 | out = out.reshape( 172 | [-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1] 173 | ) 174 | w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w) 175 | out = F.conv2d(out, w) 176 | out = out.reshape( 177 | -1, 178 | minor, 179 | in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1, 180 | in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1, 181 | ) 182 | out = out.permute(0, 2, 3, 1) 183 | 184 | return out[:, ::down_y, ::down_x, :] 185 | -------------------------------------------------------------------------------- /models/stylegan2/op/upfirdn2d_kernel.cu: -------------------------------------------------------------------------------- 1 | // Copyright (c) 2019, NVIDIA Corporation. All rights reserved. 2 | // 3 | // This work is made available under the Nvidia Source Code License-NC. 4 | // To view a copy of this license, visit 5 | // https://nvlabs.github.io/stylegan2/license.html 6 | 7 | #include 8 | 9 | #include 10 | #include 11 | #include 12 | #include 13 | 14 | #include 15 | #include 16 | 17 | 18 | static __host__ __device__ __forceinline__ int floor_div(int a, int b) { 19 | int c = a / b; 20 | 21 | if (c * b > a) { 22 | c--; 23 | } 24 | 25 | return c; 26 | } 27 | 28 | 29 | struct UpFirDn2DKernelParams { 30 | int up_x; 31 | int up_y; 32 | int down_x; 33 | int down_y; 34 | int pad_x0; 35 | int pad_x1; 36 | int pad_y0; 37 | int pad_y1; 38 | 39 | int major_dim; 40 | int in_h; 41 | int in_w; 42 | int minor_dim; 43 | int kernel_h; 44 | int kernel_w; 45 | int out_h; 46 | int out_w; 47 | int loop_major; 48 | int loop_x; 49 | }; 50 | 51 | 52 | template 53 | __global__ void upfirdn2d_kernel(scalar_t* out, const scalar_t* input, const scalar_t* kernel, const UpFirDn2DKernelParams p) { 54 | const int tile_in_h = ((tile_out_h - 1) * down_y + kernel_h - 1) / up_y + 1; 55 | const int tile_in_w = ((tile_out_w - 1) * down_x + kernel_w - 1) / up_x + 1; 56 | 57 | __shared__ volatile float sk[kernel_h][kernel_w]; 58 | __shared__ volatile float sx[tile_in_h][tile_in_w]; 59 | 60 | int minor_idx = blockIdx.x; 61 | int tile_out_y = minor_idx / p.minor_dim; 62 | minor_idx -= tile_out_y * p.minor_dim; 63 | tile_out_y *= tile_out_h; 64 | int tile_out_x_base = blockIdx.y * p.loop_x * tile_out_w; 65 | int major_idx_base = blockIdx.z * p.loop_major; 66 | 67 | if (tile_out_x_base >= p.out_w | tile_out_y >= p.out_h | major_idx_base >= p.major_dim) { 68 | return; 69 | } 70 | 71 | for (int tap_idx = threadIdx.x; tap_idx < kernel_h * kernel_w; tap_idx += blockDim.x) { 72 | int ky = tap_idx / kernel_w; 73 | int kx = tap_idx - ky * kernel_w; 74 | scalar_t v = 0.0; 75 | 76 | if (kx < p.kernel_w & ky < p.kernel_h) { 77 | v = kernel[(p.kernel_h - 1 - ky) * p.kernel_w + (p.kernel_w - 1 - kx)]; 78 | } 79 | 80 | sk[ky][kx] = v; 81 | } 82 | 83 | for (int loop_major = 0, major_idx = major_idx_base; loop_major < p.loop_major & major_idx < p.major_dim; loop_major++, major_idx++) { 84 | for (int loop_x = 0, tile_out_x = tile_out_x_base; loop_x < p.loop_x & tile_out_x < p.out_w; loop_x++, tile_out_x += tile_out_w) { 85 | int tile_mid_x = tile_out_x * down_x + up_x - 1 - p.pad_x0; 86 | int tile_mid_y = tile_out_y * down_y + up_y - 1 - p.pad_y0; 87 | int tile_in_x = floor_div(tile_mid_x, up_x); 88 | int tile_in_y = floor_div(tile_mid_y, up_y); 89 | 90 | __syncthreads(); 91 | 92 | for (int in_idx = threadIdx.x; in_idx < tile_in_h * tile_in_w; in_idx += blockDim.x) { 93 | int rel_in_y = in_idx / tile_in_w; 94 | int rel_in_x = in_idx - rel_in_y * tile_in_w; 95 | int in_x = rel_in_x + tile_in_x; 96 | int in_y = rel_in_y + tile_in_y; 97 | 98 | scalar_t v = 0.0; 99 | 100 | if (in_x >= 0 & in_y >= 0 & in_x < p.in_w & in_y < p.in_h) { 101 | v = input[((major_idx * p.in_h + in_y) * p.in_w + in_x) * p.minor_dim + minor_idx]; 102 | } 103 | 104 | sx[rel_in_y][rel_in_x] = v; 105 | } 106 | 107 | __syncthreads(); 108 | for (int out_idx = threadIdx.x; out_idx < tile_out_h * tile_out_w; out_idx += blockDim.x) { 109 | int rel_out_y = out_idx / tile_out_w; 110 | int rel_out_x = out_idx - rel_out_y * tile_out_w; 111 | int out_x = rel_out_x + tile_out_x; 112 | int out_y = rel_out_y + tile_out_y; 113 | 114 | int mid_x = tile_mid_x + rel_out_x * down_x; 115 | int mid_y = tile_mid_y + rel_out_y * down_y; 116 | int in_x = floor_div(mid_x, up_x); 117 | int in_y = floor_div(mid_y, up_y); 118 | int rel_in_x = in_x - tile_in_x; 119 | int rel_in_y = in_y - tile_in_y; 120 | int kernel_x = (in_x + 1) * up_x - mid_x - 1; 121 | int kernel_y = (in_y + 1) * up_y - mid_y - 1; 122 | 123 | scalar_t v = 0.0; 124 | 125 | #pragma unroll 126 | for (int y = 0; y < kernel_h / up_y; y++) 127 | #pragma unroll 128 | for (int x = 0; x < kernel_w / up_x; x++) 129 | v += sx[rel_in_y + y][rel_in_x + x] * sk[kernel_y + y * up_y][kernel_x + x * up_x]; 130 | 131 | if (out_x < p.out_w & out_y < p.out_h) { 132 | out[((major_idx * p.out_h + out_y) * p.out_w + out_x) * p.minor_dim + minor_idx] = v; 133 | } 134 | } 135 | } 136 | } 137 | } 138 | 139 | 140 | torch::Tensor upfirdn2d_op(const torch::Tensor& input, const torch::Tensor& kernel, 141 | int up_x, int up_y, int down_x, int down_y, 142 | int pad_x0, int pad_x1, int pad_y0, int pad_y1) { 143 | int curDevice = -1; 144 | cudaGetDevice(&curDevice); 145 | cudaStream_t stream = at::cuda::getCurrentCUDAStream(curDevice); 146 | 147 | UpFirDn2DKernelParams p; 148 | 149 | auto x = input.contiguous(); 150 | auto k = kernel.contiguous(); 151 | 152 | p.major_dim = x.size(0); 153 | p.in_h = x.size(1); 154 | p.in_w = x.size(2); 155 | p.minor_dim = x.size(3); 156 | p.kernel_h = k.size(0); 157 | p.kernel_w = k.size(1); 158 | p.up_x = up_x; 159 | p.up_y = up_y; 160 | p.down_x = down_x; 161 | p.down_y = down_y; 162 | p.pad_x0 = pad_x0; 163 | p.pad_x1 = pad_x1; 164 | p.pad_y0 = pad_y0; 165 | p.pad_y1 = pad_y1; 166 | 167 | p.out_h = (p.in_h * p.up_y + p.pad_y0 + p.pad_y1 - p.kernel_h + p.down_y) / p.down_y; 168 | p.out_w = (p.in_w * p.up_x + p.pad_x0 + p.pad_x1 - p.kernel_w + p.down_x) / p.down_x; 169 | 170 | auto out = at::empty({p.major_dim, p.out_h, p.out_w, p.minor_dim}, x.options()); 171 | 172 | int mode = -1; 173 | 174 | int tile_out_h; 175 | int tile_out_w; 176 | 177 | if (p.up_x == 1 && p.up_y == 1 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 4 && p.kernel_w <= 4) { 178 | mode = 1; 179 | tile_out_h = 16; 180 | tile_out_w = 64; 181 | } 182 | 183 | if (p.up_x == 1 && p.up_y == 1 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 3 && p.kernel_w <= 3) { 184 | mode = 2; 185 | tile_out_h = 16; 186 | tile_out_w = 64; 187 | } 188 | 189 | if (p.up_x == 2 && p.up_y == 2 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 4 && p.kernel_w <= 4) { 190 | mode = 3; 191 | tile_out_h = 16; 192 | tile_out_w = 64; 193 | } 194 | 195 | if (p.up_x == 2 && p.up_y == 2 && p.down_x == 1 && p.down_y == 1 && p.kernel_h <= 2 && p.kernel_w <= 2) { 196 | mode = 4; 197 | tile_out_h = 16; 198 | tile_out_w = 64; 199 | } 200 | 201 | if (p.up_x == 1 && p.up_y == 1 && p.down_x == 2 && p.down_y == 2 && p.kernel_h <= 4 && p.kernel_w <= 4) { 202 | mode = 5; 203 | tile_out_h = 8; 204 | tile_out_w = 32; 205 | } 206 | 207 | if (p.up_x == 1 && p.up_y == 1 && p.down_x == 2 && p.down_y == 2 && p.kernel_h <= 2 && p.kernel_w <= 2) { 208 | mode = 6; 209 | tile_out_h = 8; 210 | tile_out_w = 32; 211 | } 212 | 213 | dim3 block_size; 214 | dim3 grid_size; 215 | 216 | if (tile_out_h > 0 && tile_out_w) { 217 | p.loop_major = (p.major_dim - 1) / 16384 + 1; 218 | p.loop_x = 1; 219 | block_size = dim3(32 * 8, 1, 1); 220 | grid_size = dim3(((p.out_h - 1) / tile_out_h + 1) * p.minor_dim, 221 | (p.out_w - 1) / (p.loop_x * tile_out_w) + 1, 222 | (p.major_dim - 1) / p.loop_major + 1); 223 | } 224 | 225 | AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "upfirdn2d_cuda", [&] { 226 | switch (mode) { 227 | case 1: 228 | upfirdn2d_kernel<<>>( 229 | out.data_ptr(), x.data_ptr(), k.data_ptr(), p 230 | ); 231 | 232 | break; 233 | 234 | case 2: 235 | upfirdn2d_kernel<<>>( 236 | out.data_ptr(), x.data_ptr(), k.data_ptr(), p 237 | ); 238 | 239 | break; 240 | 241 | case 3: 242 | upfirdn2d_kernel<<>>( 243 | out.data_ptr(), x.data_ptr(), k.data_ptr(), p 244 | ); 245 | 246 | break; 247 | 248 | case 4: 249 | upfirdn2d_kernel<<>>( 250 | out.data_ptr(), x.data_ptr(), k.data_ptr(), p 251 | ); 252 | 253 | break; 254 | 255 | case 5: 256 | upfirdn2d_kernel<<>>( 257 | out.data_ptr(), x.data_ptr(), k.data_ptr(), p 258 | ); 259 | 260 | break; 261 | 262 | case 6: 263 | upfirdn2d_kernel<<>>( 264 | out.data_ptr(), x.data_ptr(), k.data_ptr(), p 265 | ); 266 | 267 | break; 268 | } 269 | }); 270 | 271 | return out; 272 | } -------------------------------------------------------------------------------- /output/result_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/output/result_0.png -------------------------------------------------------------------------------- /output/result_mask_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/output/result_mask_0.png -------------------------------------------------------------------------------- /utils/.ipynb_checkpoints/id_utils-checkpoint.py: -------------------------------------------------------------------------------- 1 | from utils.common import tensor2im 2 | from PIL import ImageColor 3 | import torch 4 | import cv2 5 | 6 | import numpy as np 7 | from collections import deque 8 | import cv2 9 | import pandas as pd 10 | import os,sys 11 | import glob 12 | 13 | import random 14 | import torch 15 | import torch.nn as nn 16 | import torch.nn.functional as F 17 | import torch.optim as optim 18 | from torchvision import transforms, utils 19 | from PIL import Image 20 | 21 | from utils.common import tensor2im 22 | from models.psp import pSp 23 | from models.e4e import e4e 24 | # from utils.inference_utils import run_on_batch 25 | 26 | from criteria import id_loss, moco_loss 27 | 28 | # from utils.common import tensor2im 29 | # from options.train_options import TrainOptions 30 | # from models.psp import pSp 31 | 32 | import streamlit as st 33 | 34 | 35 | from argparse import Namespace 36 | 37 | import torch 38 | import clip 39 | from PIL import Image 40 | 41 | device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 42 | 43 | def load_model(experiment_type='ffhq',use_baseline=False,id_constrain=False): 44 | with torch.no_grad(): 45 | if experiment_type == 'ffhq': 46 | if use_baseline: 47 | model_path = 'pretrained_models/restyle_e4e_ffhq_encode.pt' 48 | else: 49 | #model_path = 'experiment_paint_v2/checkpoints/best_model.pt' 50 | #model_path = 'experiment_paint_v4/checkpoints/best_model.pt' 51 | #model_path = 'experiment_1024_v4/checkpoints/best_model.pt' 52 | if id_constrain: 53 | model_path = 'experiments/celeba/intelli-paint/paint_1024_id-constrain_v2/checkpoints/best_model.pt' 54 | else: 55 | model_path = 'experiments/celeba/intelli-paint/paint_1024_v1/checkpoints/best_model.pt' 56 | 57 | resize_dims = (256,256) 58 | 59 | elif experiment_type == 'cars_encode': 60 | model_path = 'pretrained_models/restyle_e4e_cars_encode.pt' 61 | model_path = 'experiments/cars196/intelli-paint/paint_512_v1/checkpoints/best_model.pt' 62 | resize_dims = (192,256) 63 | 64 | ckpt = torch.load(model_path, map_location='cpu') 65 | opts = ckpt['opts'] 66 | # pprint.pprint(opts) # Display full options used 67 | # update the training options 68 | opts['checkpoint_path'] = model_path 69 | opts['device'] = device 70 | 71 | opts = Namespace(**opts) 72 | net = e4e(opts) 73 | # if experiment_type == 'horse_encode' or experiment_type == 'ffhq_encode': 74 | # net = e4e(opts) 75 | # else: 76 | # net = pSp(opts) 77 | 78 | net.eval() 79 | net = net.to(device) 80 | print('Model successfully loaded!') 81 | 82 | transform = transforms.Compose([ 83 | transforms.Resize(resize_dims), 84 | transforms.ToTensor(), 85 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]) 86 | 87 | return net, transform, opts 88 | 89 | 90 | def get_avg_image(net, experiment_type='ffhq'): 91 | avg_image = net(net.latent_avg.unsqueeze(0), 92 | input_code=True, 93 | randomize_noise=False, 94 | return_latents=False, 95 | average_code=True)[0] 96 | avg_image = avg_image.to('cuda').float().detach() 97 | if experiment_type == "cars_encode": 98 | avg_image = avg_image[:, 32:224, :] 99 | return avg_image 100 | 101 | def run_on_batch(inputs, net, opts, avg_image, target_id_feat=None): 102 | y_hat, latent = None, None 103 | #results_batch = {idx: [] for idx in range(inputs.shape[0])} 104 | #results_latent = {idx: [] for idx in range(inputs.shape[0])} 105 | for iter in range(opts.n_iters_per_batch): 106 | if iter == 0: 107 | avg_image_for_batch = avg_image.unsqueeze(0).repeat(inputs.shape[0], 1, 1, 1) 108 | x_input = torch.cat([inputs, avg_image_for_batch], dim=1) 109 | else: 110 | x_input = torch.cat([inputs, y_hat], dim=1) 111 | 112 | y_hat, latent = net.forward(x_input, 113 | target_id_feat=target_id_feat, 114 | latent=latent, 115 | randomize_noise=False, 116 | return_latents=True, 117 | resize=opts.resize_outputs) 118 | 119 | if opts.dataset_type == "cars_encode": 120 | if opts.resize_outputs: 121 | y_hat = y_hat[:, :, 32:224, :] 122 | else: 123 | y_hat = y_hat[:, :, 64:448, :] 124 | 125 | # # store intermediate outputs 126 | # for idx in range(inputs.shape[0]): 127 | # results_batch[idx].append(y_hat[idx]) 128 | # results_latent[idx].append(latent[idx].cpu().numpy()) 129 | 130 | # resize input to 256 before feeding into next iteration 131 | if opts.dataset_type == "cars_encode": 132 | y_hat = torch.nn.AdaptiveAvgPool2d((192, 256))(y_hat) 133 | else: 134 | y_hat = net.face_pool(y_hat) 135 | 136 | return y_hat, latent #results_batch, results_latent 137 | 138 | def predict_image_completion(image, net, transform, opts, preprocess=False, experiment_type='ffhq', resize_dims=(256,256), multi_modal=False, num_multi_output=5, n_iters=5, latent_mask=None ,mix_alpha=None, id_constrain=False, target_id_feat=None): 139 | opts.n_iters_per_batch = n_iters 140 | opts.resize_outputs = False # generate outputs at full resolution 141 | 142 | if preprocess: 143 | image = transform(image).to(device).unsqueeze(0) 144 | 145 | with torch.no_grad(): 146 | avg_image = get_avg_image(net,experiment_type) 147 | images, latents = run_on_batch(image, net, opts, avg_image) 148 | #run_on_batch(transformed_image.unsqueeze(0), net, experiment_type=experiment_type) 149 | #result_images, latent = images[0], latents[0] 150 | 151 | if preprocess: 152 | result_image = tensor2im(result_images[-1]).resize(resize_dims[::-1]) 153 | return images, latents -------------------------------------------------------------------------------- /utils/.ipynb_checkpoints/inference_utils-checkpoint.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | def get_average_image(net, opts): 5 | avg_image = net(net.latent_avg.unsqueeze(0), 6 | input_code=True, 7 | randomize_noise=False, 8 | return_latents=False, 9 | average_code=True)[0] 10 | avg_image = avg_image.to('cuda').float().detach() 11 | if opts.dataset_type == "cars_encode": 12 | avg_image = avg_image[:, 32:224, :] 13 | return avg_image 14 | 15 | 16 | def run_on_batch(inputs, net, opts, avg_image, target_id_feat=None): 17 | y_hat, latent = None, None 18 | results_batch = {idx: [] for idx in range(inputs.shape[0])} 19 | results_latent = {idx: [] for idx in range(inputs.shape[0])} 20 | for iter in range(opts.n_iters_per_batch): 21 | if iter == 0: 22 | avg_image_for_batch = avg_image.unsqueeze(0).repeat(inputs.shape[0], 1, 1, 1) 23 | x_input = torch.cat([inputs, avg_image_for_batch], dim=1) 24 | else: 25 | x_input = torch.cat([inputs, y_hat], dim=1) 26 | 27 | y_hat, latent = net.forward(x_input, 28 | target_id_feat=target_id_feat, 29 | latent=latent, 30 | randomize_noise=False, 31 | return_latents=True, 32 | resize=opts.resize_outputs) 33 | 34 | if opts.dataset_type == "cars_encode": 35 | if opts.resize_outputs: 36 | y_hat = y_hat[:, :, 32:224, :] 37 | else: 38 | y_hat = y_hat[:, :, 64:448, :] 39 | 40 | # store intermediate outputs 41 | for idx in range(inputs.shape[0]): 42 | results_batch[idx].append(y_hat[idx]) 43 | results_latent[idx].append(latent[idx].cpu().numpy()) 44 | 45 | # resize input to 256 before feeding into next iteration 46 | if opts.dataset_type == "cars_encode": 47 | y_hat = torch.nn.AdaptiveAvgPool2d((192, 256))(y_hat) 48 | else: 49 | y_hat = net.face_pool(y_hat) 50 | 51 | return results_batch, results_latent 52 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/1jsingh/paint2pix/971d1ea06ef6cbcc555ad09a365bf1621ce13f08/utils/__init__.py -------------------------------------------------------------------------------- /utils/common.py: -------------------------------------------------------------------------------- 1 | from PIL import Image 2 | import matplotlib.pyplot as plt 3 | 4 | 5 | def tensor2im(var): 6 | var = var.cpu().detach().transpose(0, 2).transpose(0, 1).numpy() 7 | var = ((var + 1) / 2) 8 | var[var < 0] = 0 9 | var[var > 1] = 1 10 | var = var * 255 11 | return Image.fromarray(var.astype('uint8')) 12 | 13 | 14 | def vis_faces(log_hooks): 15 | display_count = len(log_hooks) 16 | n_outputs = len(log_hooks[0]['output_face']) if type(log_hooks[0]['output_face']) == list else 1 17 | fig = plt.figure(figsize=(6 + (n_outputs * 2), 4 * display_count)) 18 | gs = fig.add_gridspec(display_count, (2 + n_outputs)) 19 | for i in range(display_count): 20 | hooks_dict = log_hooks[i] 21 | fig.add_subplot(gs[i, 0]) 22 | vis_faces_iterative(hooks_dict, fig, gs, i) 23 | plt.tight_layout() 24 | return fig 25 | 26 | 27 | def vis_faces_iterative(hooks_dict, fig, gs, i): 28 | plt.imshow(hooks_dict['input_face']) 29 | plt.title('Input\nOut Sim={:.2f}'.format(float(hooks_dict['diff_input']))) 30 | fig.add_subplot(gs[i, 1]) 31 | plt.imshow(hooks_dict['target_face']) 32 | plt.title('Target\nIn={:.2f}, Out={:.2f}'.format(float(hooks_dict['diff_views']), float(hooks_dict['diff_target']))) 33 | for idx, output_idx in enumerate(range(len(hooks_dict['output_face']) - 1, -1, -1)): 34 | output_image, similarity = hooks_dict['output_face'][output_idx] 35 | fig.add_subplot(gs[i, 2 + idx]) 36 | plt.imshow(output_image) 37 | plt.title('Output {}\n Target Sim={:.2f}'.format(output_idx, float(similarity))) 38 | -------------------------------------------------------------------------------- /utils/data_utils.py: -------------------------------------------------------------------------------- 1 | """ 2 | Code adopted from pix2pixHD: 3 | https://github.com/NVIDIA/pix2pixHD/blob/master/data/image_folder.py 4 | """ 5 | import os 6 | 7 | IMG_EXTENSIONS = [ 8 | '.jpg', '.JPG', '.jpeg', '.JPEG', 9 | '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tiff' 10 | ] 11 | 12 | 13 | def is_image_file(filename): 14 | return any(filename.endswith(extension) for extension in IMG_EXTENSIONS) 15 | 16 | 17 | def make_dataset(dir): 18 | images = [] 19 | assert os.path.isdir(dir), '%s is not a valid directory' % dir 20 | for root, _, fnames in sorted(os.walk(dir)): 21 | for fname in fnames: 22 | if is_image_file(fname): 23 | path = os.path.join(root, fname) 24 | images.append(path) 25 | return images 26 | -------------------------------------------------------------------------------- /utils/id_utils.py: -------------------------------------------------------------------------------- 1 | from utils.common import tensor2im 2 | from PIL import ImageColor 3 | import torch 4 | import cv2 5 | 6 | import numpy as np 7 | from collections import deque 8 | import cv2 9 | import pandas as pd 10 | import os,sys 11 | import glob 12 | 13 | import random 14 | import torch 15 | import torch.nn as nn 16 | import torch.nn.functional as F 17 | import torch.optim as optim 18 | from torchvision import transforms, utils 19 | from PIL import Image 20 | 21 | from utils.common import tensor2im 22 | from models.psp import pSp 23 | from models.e4e import e4e 24 | # from utils.inference_utils import run_on_batch 25 | 26 | from criteria import id_loss, moco_loss 27 | 28 | # from utils.common import tensor2im 29 | # from options.train_options import TrainOptions 30 | # from models.psp import pSp 31 | 32 | import streamlit as st 33 | 34 | 35 | from argparse import Namespace 36 | 37 | import torch 38 | import clip 39 | from PIL import Image 40 | 41 | device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 42 | 43 | def load_model(experiment_type='ffhq',use_baseline=False,id_constrain=False): 44 | with torch.no_grad(): 45 | if experiment_type == 'ffhq': 46 | if use_baseline: 47 | model_path = 'pretrained_models/restyle_e4e_ffhq_encode.pt' 48 | else: 49 | #model_path = 'experiment_paint_v2/checkpoints/best_model.pt' 50 | #model_path = 'experiment_paint_v4/checkpoints/best_model.pt' 51 | #model_path = 'experiment_1024_v4/checkpoints/best_model.pt' 52 | if id_constrain: 53 | model_path = 'experiments/celeba/intelli-paint/paint_1024_id-constrain_v2/checkpoints/best_model.pt' 54 | else: 55 | model_path = 'experiments/celeba/intelli-paint/paint_1024_v1/checkpoints/best_model.pt' 56 | 57 | resize_dims = (256,256) 58 | 59 | elif experiment_type == 'cars_encode': 60 | model_path = 'pretrained_models/restyle_e4e_cars_encode.pt' 61 | model_path = 'experiments/cars196/intelli-paint/paint_512_v1/checkpoints/best_model.pt' 62 | resize_dims = (192,256) 63 | 64 | ckpt = torch.load(model_path, map_location='cpu') 65 | opts = ckpt['opts'] 66 | # pprint.pprint(opts) # Display full options used 67 | # update the training options 68 | opts['checkpoint_path'] = model_path 69 | opts['device'] = device 70 | 71 | opts = Namespace(**opts) 72 | net = e4e(opts) 73 | # if experiment_type == 'horse_encode' or experiment_type == 'ffhq_encode': 74 | # net = e4e(opts) 75 | # else: 76 | # net = pSp(opts) 77 | 78 | net.eval() 79 | net = net.to(device) 80 | print('Model successfully loaded!') 81 | 82 | transform = transforms.Compose([ 83 | transforms.Resize(resize_dims), 84 | transforms.ToTensor(), 85 | transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]) 86 | 87 | return net, transform, opts 88 | 89 | 90 | def get_avg_image(net, experiment_type='ffhq'): 91 | avg_image = net(net.latent_avg.unsqueeze(0), 92 | input_code=True, 93 | randomize_noise=False, 94 | return_latents=False, 95 | average_code=True)[0] 96 | avg_image = avg_image.to('cuda').float().detach() 97 | if experiment_type == "cars_encode": 98 | avg_image = avg_image[:, 32:224, :] 99 | return avg_image 100 | 101 | def run_on_batch(inputs, net, opts, avg_image, target_id_feat=None): 102 | y_hat, latent = None, None 103 | #results_batch = {idx: [] for idx in range(inputs.shape[0])} 104 | #results_latent = {idx: [] for idx in range(inputs.shape[0])} 105 | for iter in range(opts.n_iters_per_batch): 106 | if iter == 0: 107 | avg_image_for_batch = avg_image.unsqueeze(0).repeat(inputs.shape[0], 1, 1, 1) 108 | x_input = torch.cat([inputs, avg_image_for_batch], dim=1) 109 | else: 110 | x_input = torch.cat([inputs, y_hat], dim=1) 111 | 112 | y_hat, latent = net.forward(x_input, 113 | target_id_feat=target_id_feat, 114 | latent=latent, 115 | randomize_noise=False, 116 | return_latents=True, 117 | resize=opts.resize_outputs) 118 | 119 | if opts.dataset_type == "cars_encode": 120 | if opts.resize_outputs: 121 | y_hat = y_hat[:, :, 32:224, :] 122 | else: 123 | y_hat = y_hat[:, :, 64:448, :] 124 | 125 | # # store intermediate outputs 126 | # for idx in range(inputs.shape[0]): 127 | # results_batch[idx].append(y_hat[idx]) 128 | # results_latent[idx].append(latent[idx].cpu().numpy()) 129 | 130 | # resize input to 256 before feeding into next iteration 131 | if opts.dataset_type == "cars_encode": 132 | y_hat = torch.nn.AdaptiveAvgPool2d((192, 256))(y_hat) 133 | else: 134 | y_hat = net.face_pool(y_hat) 135 | 136 | return y_hat, latent #results_batch, results_latent 137 | 138 | def predict_image_completion(image, net, transform, opts, preprocess=False, experiment_type='ffhq', resize_dims=(256,256), multi_modal=False, num_multi_output=5, n_iters=5, latent_mask=None ,mix_alpha=None, id_constrain=False, target_id_feat=None): 139 | opts.n_iters_per_batch = n_iters 140 | opts.resize_outputs = False # generate outputs at full resolution 141 | 142 | if preprocess: 143 | image = transform(image).to(device).unsqueeze(0) 144 | 145 | with torch.no_grad(): 146 | avg_image = get_avg_image(net,experiment_type) 147 | images, latents = run_on_batch(image, net, opts, avg_image) 148 | #run_on_batch(transformed_image.unsqueeze(0), net, experiment_type=experiment_type) 149 | #result_images, latent = images[0], latents[0] 150 | 151 | if preprocess: 152 | result_image = tensor2im(result_images[-1]).resize(resize_dims[::-1]) 153 | return images, latents -------------------------------------------------------------------------------- /utils/inference_utils.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | def get_average_image(net, opts): 5 | avg_image = net(net.latent_avg.unsqueeze(0), 6 | input_code=True, 7 | randomize_noise=False, 8 | return_latents=False, 9 | average_code=True)[0] 10 | avg_image = avg_image.to('cuda').float().detach() 11 | if opts.dataset_type == "cars_encode": 12 | avg_image = avg_image[:, 32:224, :] 13 | return avg_image 14 | 15 | 16 | def run_on_batch(inputs, net, opts, avg_image, target_id_feat=None): 17 | y_hat, latent = None, None 18 | results_batch = {idx: [] for idx in range(inputs.shape[0])} 19 | results_latent = {idx: [] for idx in range(inputs.shape[0])} 20 | for iter in range(opts.n_iters_per_batch): 21 | if iter == 0: 22 | avg_image_for_batch = avg_image.unsqueeze(0).repeat(inputs.shape[0], 1, 1, 1) 23 | x_input = torch.cat([inputs, avg_image_for_batch, inputs, avg_image_for_batch], dim=1) 24 | else: 25 | x_input = torch.cat([inputs, y_hat, inputs, y_hat], dim=1) 26 | 27 | y_hat, latent = net.forward(x_input, 28 | target_id_feat=target_id_feat, 29 | latent=latent, 30 | randomize_noise=False, 31 | return_latents=True, 32 | resize=opts.resize_outputs) 33 | 34 | if opts.dataset_type == "cars_encode": 35 | if opts.resize_outputs: 36 | y_hat = y_hat[:, :, 32:224, :] 37 | else: 38 | y_hat = y_hat[:, :, 64:448, :] 39 | 40 | # store intermediate outputs 41 | for idx in range(inputs.shape[0]): 42 | results_batch[idx].append(y_hat[idx]) 43 | results_latent[idx].append(latent[idx].cpu().numpy()) 44 | 45 | # resize input to 256 before feeding into next iteration 46 | if opts.dataset_type == "cars_encode": 47 | y_hat = torch.nn.AdaptiveAvgPool2d((192, 256))(y_hat) 48 | else: 49 | y_hat = net.face_pool(y_hat) 50 | 51 | return results_batch, results_latent 52 | -------------------------------------------------------------------------------- /utils/model_utils.py: -------------------------------------------------------------------------------- 1 | # specify the encoder types for pSp and e4e - this is mainly used for the inference scripts 2 | ENCODER_TYPES = { 3 | 'pSp': ['GradualStyleEncoder', 'ResNetGradualStyleEncoder', 'BackboneEncoder', 'ResNetBackboneEncoder'], 4 | 'e4e': ['ProgressiveBackboneEncoder', 'ResNetProgressiveBackboneEncoder'] 5 | } 6 | 7 | RESNET_MAPPING = { 8 | 'layer1.0': 'body.0', 9 | 'layer1.1': 'body.1', 10 | 'layer1.2': 'body.2', 11 | 'layer2.0': 'body.3', 12 | 'layer2.1': 'body.4', 13 | 'layer2.2': 'body.5', 14 | 'layer2.3': 'body.6', 15 | 'layer3.0': 'body.7', 16 | 'layer3.1': 'body.8', 17 | 'layer3.2': 'body.9', 18 | 'layer3.3': 'body.10', 19 | 'layer3.4': 'body.11', 20 | 'layer3.5': 'body.12', 21 | 'layer4.0': 'body.13', 22 | 'layer4.1': 'body.14', 23 | 'layer4.2': 'body.15', 24 | } 25 | -------------------------------------------------------------------------------- /utils/train_utils.py: -------------------------------------------------------------------------------- 1 | 2 | def aggregate_loss_dict(agg_loss_dict): 3 | mean_vals = {} 4 | for output in agg_loss_dict: 5 | for key in output: 6 | mean_vals[key] = mean_vals.setdefault(key, []) + [output[key]] 7 | for key in mean_vals: 8 | if len(mean_vals[key]) > 0: 9 | mean_vals[key] = sum(mean_vals[key]) / len(mean_vals[key]) 10 | else: 11 | print('{} has no value'.format(key)) 12 | mean_vals[key] = 0 13 | return mean_vals 14 | --------------------------------------------------------------------------------