I am Haoxing Chen. Now, I'm an AI researcher at Tiansuan Lab, Ant Group.
- e-mail: [email protected]
- google scholar: https://scholar.google.com/citations?hl=zh-CN&pli=1&user=BnS7HzAAAAAJ
- homepage: https://chenhaoxing.github.io/
This is the implementation of paper 'Hierarchical Dynamic Image Harmonization' (ACM MM'2023, Oral).
License: MIT License
I am Haoxing Chen. Now, I'm an AI researcher at Tiansuan Lab, Ant Group.
Hi author, I'm trying your harmonization method. But when I was processing the data set, I did not find IHD_train.txt and IHD_test.txt. During the training process, the data set could not be preprocessed and an error was reported.
Could you please upload the two files?
Thank you very much for your reply
hi, i wonder to test this model with my pics, do you have pretrained model to be released? thx
Hello do you plan to release the Trained weights?
Given an image, the harmonized quality is better other methods. But for video, the consistency is poor, how to solve it?
只有训练集验证集的代码,没有给测试集的代码吗?
I have tried muliple times to train your model but I constantly hit stall around this values MSE:20.184, PSNR:39.458, fMSE:217.668, SSIM:0.992 which, while still being quite good, are not the one described in the paper.
I would like to see the configuration you used to train it. This is mine:
----------------- Options ---------------
batch_size: 90 [default: 12]
beta1: 0.9
checkpoints_dir: ./checkpoints
continue_train: False
crop_size: 512 [default: 256]
d_lr_ratio: 1.0
dataset_mode: iharmony4
dataset_root: ./data/harmony [default: ./datasets/IH256/]
display_env: main
display_freq: 500
display_id: 1
display_port: 8097
display_server: http://localhost
display_winsize: 256
epoch: latest
epoch_count: 1
g_lr_ratio: 1.0
gan_mode: vanilla
gp_ratio: 1.0
gpu_ids: 0,1 [default: 0]
init_gain: 0.02
init_type: normal
input_nc: 3
isTrain: True [default: None]
is_train: True
lambda_L1: 1.0
lambda_a: 1.0
lambda_v: 1.0
load_iter: 0
load_size: 256
lr: 0.001
lr_decay_iters: 100
lr_policy: target_decay
max_dataset_size: inf
model: hdnet
n_layers_D: 3
name: armonizzatoremm [default: experiment_train]
ndf: 64
netD: basic
netG: hdnet
ngf: 32
niter: 120
niter_decay: 0
no_dropout: False
no_html: False
normD: instance
normG: RAIN
num_threads: 15 [default: 32]
output_nc: 3
phase: train
pool_size: 0
preprocess: none
print_freq: 300
save_by_iter: False
save_epoch_freq: 1
save_latest_freq: 5000
serial_batches: False
suffix:
update_html_freq: 500
verbose: False
----------------- End -------------------
对于1080P的图片输入有时会出现大量白色噪点,应该怎么办呢
Hi,
I'm very interested in your work, and I'm also very grateful for you sharing your code. While reproducing your code, I encountered two issues:
1)it seems that there's no requirement.txt;
2)the input images can only be 512x512 or 256x256, is that correct?
Can I provide a pre trained model? I need to use it to harmonize local images.
Hi,
I find that when computing struct similarity in evaluation, unless channel_axis is set, there is an exception. Its a small fix, but I wonder if I'm doing something wrong otherwise everyone will get this error for RGB images.
Thanks!
Hi, I followed the steps in the documentation to complete the training and testing.
The dataset which I used is iHarmony4, all images are resized to 256 * 256, but the quantitative results are inconsistent with your paper:
| average MSE | average PSNR
my result | 16.463 | 40.805
paper | 23.42 | 38.58
This result confuses me. Is there anything I missed? Thanks a lot !
When can the authors release the pre-trained model?
In the part of Training and validation, --save_dir and --device are Invalid. If the authors have time, you can check this part.
Hope everything goes well!
Hello dear developers,
any update on the new code?
Thanks
没有找到requirements.txt,是不是忘记放上去了呢?
Just for anyone interested in training this model. Do not use multiple gpus, it shows massive performance dump.
For instance, I was getting 1.49s/i with two gpus at batch size 12, and with one gpu I get 5.53it/s
在preprocess_iharmony4.py中并没有把调整后的图像数据写入新创建的数据集iHarmony4Resized,而是直接在原数据集中修改。我之前运行这个py文件时一直都是iHarmony4Resized文件夹是空的,后来用断点调试发现了这个问题,建议作者对原代码修改一下
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.