nicoledjy / dl_proj

dl_proj

• AutoEncoder pretraining

  Use folder autoencoder

  python AE_pretrain_new.py
  

  The AE_pretrain_new.py is the version we decided to use. It mainly go through the images one by one in [3, 256, 306] format. The previous version AE_pretrain.py sew 6 images together as in [3, 256x3, 306x2] format, which didn't yield a good result.

• Roadmap prediction:

  Use folder hrnet

  • --batch-size, default=2
  • --epochs, default=10
  • --lr, default=1e-4
  • --weight-decay, default=1e-4
  • --data-dir, default='../data'
  • --out-file, default='HRNET_RM_model.pt'

  python train_HRNet_RoadMap.py
  

• Bounding Box prediction:

  Use folder yolo

  python trainYolo_withPretrain.py 
  python trainYolo.py 
  

• Object detection without per-training:

  Use folder without_pretrain, arguments for the main.py are the following:

  • --batch-size, default=2
  • --epochs, default=10
  • --lr, default=1e-4
  • --weight-decay, default=1e-4

  python main.py
  

• PIRL pre-training:

  Use folder pretrain, run pirl_train.py(path change may be required) , and main arguments are the following:

  • --num-scene, number of scenes used for the pertaining, default=106
  • --model-type, default=res18
  • --batch-size, default=128
  • --epochs, default=100
  • --lr, default=0.01
  • --count-negatives, default=6400 (need to be half the size of images used)

  python pirl_train.py --num-scene 1 --model-type res50 --batch-size 2 --lr 0.1 --count-negatives 200
  

• Object detection with Resnet50 backbone architecture:

  • use folder pretrain_obj
  • No argument parser added yet, need to set pretrain_res=True and add the path in main.py

  python main.py
  

• Visualization:

  In folder visualization

  Jupyter Notebooks that contain visualization code.