A PyTorch implementation of my thesis with the same name.
- Install from pip
pip install rmn
# or
git clone [email protected]:phamquiluan/ResidualMaskingNetwork.git
cd ResidualMaskingNetwork
pip install -e .
- Run infenrece by the following Python scripts
from rmn import RMN
m = RMN()
m.video_demo()
# or
image = cv2.imread("some-image-path.png")
results = m.detect_emotion_for_single_frame(image)
print(results)
image = m.draw(image, results)
cv2.imwrite("output.png", image)
- Model file: download (this checkpoint is trained on VEMO dataset, locate it at
./saved/checkpoints/
directory) - Download 2 files: prototxt, and res10_300x300_ssd for face detection OpenCV. Locate at current directory or checking file path with
ssd_infer.py
file.
python ssd_infer.py
- Recent Update
- Benchmarking on FER2013
- Benchmarking on ImageNet
- Installation
- Download datasets
- Training on FER2013
- Training on ImageNet
- Evaluation results
- Download dissertation and slide
- [05/05/2021] Release ver 2, add colab
- [27/02/2021] Add paper
- [14/01/2021] Packaging Project and publish
rmn
on Pypi - [27/02/2020] Update Tensorboard visualizations and Overleaf source
- [22/02/2020] Test-time augmentation implementation.
- [21/02/2020] Imagenet training code and trained weights released.
- [21/02/2020] Imagenet evaluation results released.
- [10/01/2020] Checking demo stuff and training procedure works on another machine
- [09/01/2020] First time upload
We benchmark our code thoroughly on two datasets: FER2013 and VEMO. Below are the results and trained weights:
Model | Accuracy |
---|---|
VGG19 | 70.80 |
EfficientNet_b2b | 70.80 |
Googlenet | 71.97 |
Resnet34 | 72.42 |
Inception_v3 | 72.72 |
Bam_Resnet50 | 73.14 |
Densenet121 | 73.16 |
Resnet152 | 73.22 |
Cbam_Resnet50 | 73.39 |
ResMaskingNet | 74.14 |
ResMaskingNet + 6 | 76.82 |
Results in VEMO dataset could be found in my thesis or slide (attached below)
We also benchmark our model on ImageNet dataset.
Model | Top-1 Accuracy | Top-5 Accuracy |
---|---|---|
Resnet34 | 72.59 | 90.92 |
CBAM Resnet34 | 73.77 | 91.72 |
ResidualMaskingNetwork | 74.16 | 91.91 |
- Install PyTorch by selecting your environment on the website and running the appropriate command.
- Clone this repository and install package prerequisites below.
- Then download the dataset by following the instructions below.
- Python 3.6+
- requirements.txt
- FER2013 Dataset (locate it in
saved/data/fer2013
likesaved/data/fer2013/train.csv
) - ImageNet 1K Dataset (ensure it can be loaded by torchvision.datasets.Imagenet)
- To train network, you need to specify model name and other hyperparameters in config file (located at configs/*) then ensure it is loaded in main file, then run training procedure by simply run main file, for example:
python main_fer.py # Example for fer2013_config.json file
- The best checkpoints will chosen at term of best validation accuracy, located at
saved/checkpoints
- The TensorBoard training logs are located at
saved/logs
, to open it, usetensorboard --logdir saved/logs/
- By default, it will train
alexnet
model, you can switch to another model by editconfigs/fer2013\_config.json
file (toresnet18
orcbam\_resnet50
or my networkresmasking\_dropout1
.
To perform training resnet34 on 4 V100 GPUs on a single machine:
python ./main_imagenet.py -a resnet34 --dist-url 'tcp://127.0.0.1:12345' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0
For student, who takes care of font family of confusion matrix and would like to write things in LaTeX, below is an example for generating a striking confusion matrix.
(Read this article for more information, there will be some bugs if you blindly run the code without reading).
python cm_cbam.py
I used no-weighted sum avarage ensemble method to fusing 7 different models together, to reproduce results, you need to do some steps:
- Download all needed trained weights and located on
./saved/checkpoints/
directory. Link to download can be found on Benchmarking section. - Edit file
gen_results
and run it to generate result offline for each model. - Run
gen_ensemble.py
file to generate accuracy for example methods.
DVC has been used to track various data processing and model training stages. This ensures both repeatability and posterity, as the nature of processing is recorded as are the results. The actual data is stored in an S3-compliant object store which isn't publicly available. Artifacts can be made available upon request to [email protected].
- Dissertation PDF (in Vietnamese)
- Dissertation Overleaf Source
- Presentation slide PDF (in English) with full appendix
- Presentation slide Overleaf Source
- Paper
We have accumulated the following to-do list, which we hope to complete in the near future
- Still to come:
- Upload all models and training code.
- Test time augmentation.
- GPU-Parallel.
- Pretrained model.
- Demo and inference code.
- Imagenet trained and pretrained weights.
- GradCAM visualization and Pooling method for visualize activations.
- Centerloss Visualizations.
Note: Unfortunately, I am currently join a full-time job and research on another topic, so I'll do my best to keep things up to date, but no guarantees. That being said, thanks to everyone for your continued help and feedback as it is really appreciated. I will try to address everything as soon as possible.
- Same as in dissertation.
L. Pham, H. Vu, T. A. Tran, "Facial Expression Recognition Using Residual Masking Network", IEEE 25th International Conference on Pattern Recognition, 2020, 4513-4519. Milan -Italia.
@inproceedings{luanresmaskingnet2020,
title={Facial Expression Recognition using Residual Masking Network},
author={Luan, Pham and Huynh, Vu and Tuan Anh, Tran},
booktitle={IEEE 25th International Conference on Pattern Recognition},
pages={4513--4519},
year={2020}
}