0. Sign up
1. Install
2. Prepare the dataset
3. Training
4. Convert PIDNet to _torchfx.pt
5. Model compression with NetsPresso Python Package
6. Fine-tuning the compressed model
7. Evaluation
8. Custom inputs
To get started with the NetsPresso Python package, you will need to sign up at NetsPresso.
Clone repo, including PyTorch >= 1.11, < 2.0.
git clone https://github.com/Nota-NetsPresso/PIDNet_nota # clone
- Download the Cityscapes and CamVid datasets and unzip them in
data/cityscapes
anddata/camvid
dirs. - Check if the paths contained in lists of
data/list
are correct for dataset images.
- Download the images and annotations from Kaggle, where the resolution of images is 960x720 (original);
- Unzip the data and put all the images and all the colored labels into
data/camvid/images/
anddata/camvid/labels
, respectively; - Following the split of train, val and test sets used in SegNet-Tutorial, we have generated the dataset lists in
data/list/camvid/
; - Finished!!! (We have open an issue for everyone who's interested in CamVid to discuss where to download the data and if the split in SegNet-Tutorial is correct. BTW, do not directly use the split in Kaggle, which is wrong and will lead to unnormal high accuracy. We have revised the CamVid content in the paper and you will see the correct results after its announcement.)
- Download the ImageNet pretrained models and put them into
pretrained_models/imagenet/
dir. - For example, train the PIDNet-S on Cityscapes with batch size of 12 on 2 GPUs:
python tools/train.py --cfg configs/cityscapes/pidnet_small_cityscapes.yaml GPUS (0,1) TRAIN.BATCH_SIZE_PER_GPU 6
- Or train the PIDNet-L on Cityscapes using train and val sets simultaneously with batch size of 12 on 4 GPUs:
python tools/train.py --cfg configs/cityscapes/pidnet_large_cityscapes_trainval.yaml GPUS (0,1,2,3) TRAIN.BATCH_SIZE_PER_GPU 3
- Download the finetuned models for Cityscapes and CamVid and put them into
pretrained_models/cityscapes/
andpretrained_models/camvid/
dirs, respectively. - For example, convert the PIDNet-S on Cityscapes val set:
python tools/export_netspresso.py --cfg configs/cityscapes/pidnet_small_cityscapes.yaml \
TEST.MODEL_FILE pretrained_models/cityscapes/PIDNet_S_Cityscapes_val.pt
Executing this code will create 'model_modelfx.pt' and 'model_headfx.pt'.
Upload & compress your 'model_modelfx.pt' by using NetsPresso Python Package
pip install netspresso
First, import the packages and set a NetsPresso username and password.
from netspresso.compressor import ModelCompressor, Task, Framework, CompressionMethod, RecommendationMethod
EMAIL = "YOUR_EMAIL"
PASSWORD = "YOUR_PASSWORD"
compressor = ModelCompressor(email=EMAIL, password=PASSWORD)
Second, upload 'model_modelfx.pt', which is the model converted to torchfx in step 4, with the following code.
# Upload Model
UPLOAD_MODEL_NAME = "pidnet_model"
TASK = Task.SEMANTIC_SEGMENTATION
FRAMEWORK = Framework.PYTORCH
UPLOAD_MODEL_PATH = "./model_modelfx.pt"
INPUT_SHAPES = [{"batch": 1, "channel": 3, "dimension": [1024, 1024]}]
model = compressor.upload_model(
model_name=UPLOAD_MODEL_NAME,
task=TASK,
framework=FRAMEWORK,
file_path=UPLOAD_MODEL_PATH,
input_shapes=INPUT_SHAPES,
)
Finally, you can compress the uploaded model with the desired options through the following code.
# Recommendation Compression
COMPRESSED_MODEL_NAME = "test_l2norm"
COMPRESSION_METHOD = CompressionMethod.PR_L2
RECOMMENDATION_METHOD = RecommendationMethod.SLAMP
RECOMMENDATION_RATIO = 0.6
OUTPUT_PATH = "./compressed_pidnet.pt"
compressed_model = compressor.recommendation_compression(
model_id=model.model_id,
model_name=COMPRESSED_MODEL_NAME,
compression_method=COMPRESSION_METHOD,
recommendation_method=RECOMMENDATION_METHOD,
recommendation_ratio=RECOMMENDATION_RATIO,
output_path=OUTPUT_PATH,
)
Click to check 'Full upload & compress code'
pip install netspresso
from netspresso.compressor import ModelCompressor, Task, Framework, CompressionMethod, RecommendationMethod
EMAIL = "YOUR_EMAIL"
PASSWORD = "YOUR_PASSWORD"
compressor = ModelCompressor(email=EMAIL, password=PASSWORD)
# Upload Model
UPLOAD_MODEL_NAME = "pidnet_model"
TASK = Task.SEMANTIC_SEGMENTATION
FRAMEWORK = Framework.PYTORCH
UPLOAD_MODEL_PATH = "./model_modelfx.pt"
INPUT_SHAPES = [{"batch": 1, "channel": 3, "dimension": [1024, 1024]}]
model = compressor.upload_model(
model_name=UPLOAD_MODEL_NAME,
task=TASK,
framework=FRAMEWORK,
file_path=UPLOAD_MODEL_PATH,
input_layers=INPUT_SHAPES,
)
# Recommendation Compression
COMPRESSED_MODEL_NAME = "test_l2norm"
COMPRESSION_METHOD = CompressionMethod.PR_L2
RECOMMENDATION_METHOD = RecommendationMethod.SLAMP
RECOMMENDATION_RATIO = 0.6
OUTPUT_PATH = "./compressed_pidnet.pt"
compressed_model = compressor.recommendation_compression(
model_id=model.model_id,
model_name=COMPRESSED_MODEL_NAME,
compression_method=COMPRESSION_METHOD,
recommendation_method=RECOMMENDATION_METHOD,
recommendation_ratio=RECOMMENDATION_RATIO,
output_path=OUTPUT_PATH,
)
More commands can be found in the official NetsPresso Python Package docs: https://nota-netspresso.github.io/PyNetsPresso-docs
Alternatively, you can do the same as above through the GUI on our website: https://console.netspresso.ai/models
After compression, retraining is necessary. You can retrain with the following code.
Along with the --netspresso option, you need to put the path of the compressed model in the --model option and the path of model_headfx.pt that came out while converting to torchfx in the --head option.
python tools/train.py --netspresso --model model_model.pt --head model_head.pt --cfg configs/cityscapes/pidnet_small_cityscapes.yaml
If you want to perform additional compression, compress x_model_model_pt from training as in Step 5.
In the above command, put the path of the newly compressed model in the --model option. In the --head option, you need to change it to the path of x_model_head_pt that came out through retraining.
After retraining, files like x_model_model.pt and x_model_head.pt come out. You can run the model through eval.py.
python tools/eval.py --netspresso --model x_model_model.pt --head x_model_head.pt
Also, you can put all your images in samples/
and then run the command below using compressed and retrained model for image format of .png:
python tools/custom.py --netspresso --model x_model_model.pt --t '.png'
You can use the compressed model however you like!
Join our Discussion Forum for providing feedback or sharing your use cases, and if you want to talk more with Nota, please contact us here.
Or you can also do it via email(contact@nota.ai) or phone(+82 2-555-8659)!
This is the official repository for our recent work: PIDNet (PDF)
Comparison of inference speed and accuracy for real-time models on test set of Cityscapes.
- Towards Real-time Applications: PIDNet could be directly used for the real-time applications, such as autonomous vehicle and medical imaging.
- A Novel Three-branch Network: Addtional boundary branch is introduced to two-branch network to mimic the PID controller architecture and remedy the overshoot issue of previous models.
- More Accurate and Faster: PIDNet-S presents 78.6% mIOU with speed of 93.2 FPS on Cityscapes test set and 80.1% mIOU with speed of 153.7 FPS on CamVid test set. Also, PIDNet-L becomes the most accurate one (80.6% mIOU) among all the real-time networks for Cityscapes.
- This paper was accepted by CVPR 2023, new version and associated materials will be available soon! (Apr/06/2023)
- Fixed the data bug for Camvid and the new version of arXiv preprint will be available on Jun 13th. (Jun/09/2022)
- Our paper was marked as state of the art in Papers with Code. (Jun/06/2022)
- Our paper was submitted to arXiv for public access. (Jun/04/2022)
- The training and testing codes and trained models for PIDNet are available here. (Jun/03/2022)
A demo of the segmentation performance of our proposed PIDNets: Original video (left) and predictions of PIDNet-S (middle) and PIDNet-L (right)
Cityscapes Stuttgart demo video #1
Cityscapes Stuttgart demo video #2
An overview of the basic architecture of our proposed Proportional-Integral-Derivative Network (PIDNet).
Instantiation of the PIDNet for semantic segmentation.
For simple reproduction, we provide the ImageNet pretrained models here.
Model (ImageNet) | PIDNet-S | PIDNet-M | PIDNet-L |
---|---|---|---|
Link | download | download | download |
Also, the finetuned models on Cityscapes and Camvid are available for direct application in road scene parsing.
Model (Cityscapes) | Val (% mIOU) | Test (% mIOU) | FPS |
---|---|---|---|
PIDNet-S | 78.8 | 78.6 | 93.2 |
PIDNet-M | 79.9 | 79.8 | 42.2 |
PIDNet-L | 80.9 | 80.6 | 31.1 |
Model (CamVid) | Val (% mIOU) | Test (% mIOU) | FPS |
---|---|---|---|
PIDNet-S | - | 80.1 | 153.7 |
PIDNet-M | - | 82.0 | 85.6 |
This implementation is based on HRNet-Semantic-Segmentation. Please refer to their repository for installation and dataset preparation. The inference speed is tested on single RTX 3090 using the method introduced by SwiftNet. No third-party acceleration lib is used, so you can try TensorRT or other approaches for faster speed.
- Download the Cityscapes and CamVid datasets and unzip them in
data/cityscapes
anddata/camvid
dirs. - Check if the paths contained in lists of
data/list
are correct for dataset images.
- Download the images and annotations from Kaggle, where the resolution of images is 960x720 (original);
- Unzip the data and put all the images and all the colored labels into
data/camvid/images/
anddata/camvid/labels
, respectively; - Following the split of train, val and test sets used in SegNet-Tutorial, we have generated the dataset lists in
data/list/camvid/
; - Finished!!! (We have open an issue for everyone who's interested in CamVid to discuss where to download the data and if the split in SegNet-Tutorial is correct. BTW, do not directly use the split in Kaggle, which is wrong and will lead to unnormal high accuracy. We have revised the CamVid content in the paper and you will see the correct results after its announcement.)
- Download the ImageNet pretrained models and put them into
pretrained_models/imagenet/
dir. - For example, train the PIDNet-S on Cityscapes with batch size of 12 on 2 GPUs:
python tools/train.py --cfg configs/cityscapes/pidnet_small_cityscapes.yaml GPUS (0,1) TRAIN.BATCH_SIZE_PER_GPU 6
- Or train the PIDNet-L on Cityscapes using train and val sets simultaneously with batch size of 12 on 4 GPUs:
python tools/train.py --cfg configs/cityscapes/pidnet_large_cityscapes_trainval.yaml GPUS (0,1,2,3) TRAIN.BATCH_SIZE_PER_GPU 3
- Download the finetuned models for Cityscapes and CamVid and put them into
pretrained_models/cityscapes/
andpretrained_models/camvid/
dirs, respectively. - For example, evaluate the PIDNet-S on Cityscapes val set:
python tools/eval.py --cfg configs/cityscapes/pidnet_small_cityscapes.yaml \
TEST.MODEL_FILE pretrained_models/cityscapes/PIDNet_S_Cityscapes_val.pt
- Or, evaluate the PIDNet-M on CamVid test set:
python tools/eval.py --cfg configs/camvid/pidnet_medium_camvid.yaml \
TEST.MODEL_FILE pretrained_models/camvid/PIDNet_M_Camvid_Test.pt \
DATASET.TEST_SET list/camvid/test.lst
- Generate the testing results of PIDNet-L on Cityscapes test set:
python tools/eval.py --cfg configs/cityscapes/pidnet_large_cityscapes_trainval.yaml \
TEST.MODEL_FILE pretrained_models/cityscapes/PIDNet_L_Cityscapes_test.pt \
DATASET.TEST_SET list/cityscapes/test.lst
- Measure the inference speed of PIDNet-S for Cityscapes:
python models/speed/pidnet_speed.py --a 'pidnet-s' --c 19 --r 1024 2048
- Measure the inference speed of PIDNet-M for CamVid:
python models/speed/pidnet_speed.py --a 'pidnet-m' --c 11 --r 720 960
- Put all your images in
samples/
and then run the command below using Cityscapes pretrained PIDNet-L for image format of .png:
python tools/custom.py --a 'pidnet-l' --p '../pretrained_models/cityscapes/PIDNet_L_Cityscapes_test.pt' --t '.png'
If you think this implementation is useful for your work, please cite our paper:
@misc{xu2022pidnet,
title={PIDNet: A Real-time Semantic Segmentation Network Inspired from PID Controller},
author={Jiacong Xu and Zixiang Xiong and Shankar P. Bhattacharyya},
year={2022},
eprint={2206.02066},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
- Our implementation is modified based on HRNet-Semantic-Segmentation.
- Latency measurement code is borrowed from the DDRNet.
- Thanks for their nice contribution.