Readme.md

Workflow of Next Vision Transformer

Ohsome2label(Data preparing)

1. Check the ohsome2label API useful or not, use $ ohsome2label —help

2. Change the save direction and some details of output: name: WWTP workspace: ./BW_WWTP project_time: 2020-05-18 task: object detection

3. Check the size of bboxes between output and the prepared one

4. Use the input like:

   • {'label': 'wastewater', 'key': 'landuse', 'value': 'basin'}

$ ohsome2label --config C:\Users\24634\PycharmProjects\pythonProject\ohsome2label\config\config.yaml vector

II: change the .geojson file

$ ohsome2label label $ ohsome2label image $ ohsome2label visualize -n 10

#to split the train and test data from ohsome2label python tf_record_from_coco.py --label_input=D:\MA\Data\Ohsome\ohsome2label\STUT_WWTP --train_rd_path=D:\MA\Data\Ohsome\ohsome2label\STUT_WWTP\train.json --valid_rd_path=D:\MA\Data\Ohsome\ohsome2label\STUT_WWTP\valid.json

Docker

#I just built a new docker image "vit" in tum node 6

$ docker build -f ViT.dockerfile -t vit .

#I am in the step of setting up the docker env, and training the coco2017 data sets with my NVT model.

$ sudo docker run -it --ipc=host --gpus all -v /home/dev/Next-ViT:/source/Next-Vit -v /usr/local/cuda-11.8:/usr/local/cuda-11.8 -w /source/Next-Vit/detection nextvit1:latest

$ cd detection/

$ PORT=29501 bash dist_train.sh configs/mask_rcnn_nextvit_small_1x.py 1

$ PORT=29501 bash dist_train.sh configs/mask_rcnn_nextvit_base_1x.py 1

$ PORT=29501 bash dist_train.sh configs/mask_rcnn_nextvit_large_1x.py 1

#change the data to my german data, but there is an error about the the nummer of class, so we should change coco.py(fixed by add "classes = ('wastewater', ),")

Work flow

1. Devided the total data sets to 2 big parts, Training samples and Test samples.

2. Split the training samples into training data and validation data

3. Fit the training samples to our 3 inter-structures Next-ViT(small,base,large) There are a comparation between these training results:

• small:

• base:

• large:

4. save the model and make a prediction Perpare a test_samples folder, which have test images and negative images in this folder. And we need a txt file to get all a name list of all the images'names. Use the prediction.py to make a prediction with the above images. The comparations between the different sizes of Next-ViT

And another final visualisation in stuttgart:

5. calculate the confusion matrix There are acutually 3 sizes of the the model, after comparing the Large size works at the best. Thus, in this section, we calculate the confusion matrix based on these large model with different thresholds. The test area located in the stuttgart and the surrounding areas.

References: Next-ViT: https://github.com/bytedance/Next-ViT