Skip to content

Latest commit

 

History

History
100 lines (72 loc) · 3.32 KB

unsupervised.md

File metadata and controls

100 lines (72 loc) · 3.32 KB
Raw

3D Unsupervised Learning

Unsupervised pretraining

  1. Download the ShapeNetCore.v1.zip from here and place it in the folder tensorflow/script/dataset/midnet_data. Then run the following command to preprocess the data and make the tfrecords:

    cd  tensorflow/data
python midnet_data.py --run shapenet_create_tfrecords \
                      --scanner <The path of the virtual_scanner>

  2. For convenience, we also provide the tfrecords we made. Download the data and pretrained weights with the following command.

    python midnet_data.py --run download_data

  3. Run the following command to train the network.

    cd  tensorflow/script
python run_mid.py --config configs/mid_hrnet_d6.yaml

Finetune on ModelNet40

Follow the instructions here to download the ModelNet40 and preprocess the data. Make sure you can train the HRNet with random initialization here.

Then run the following script to finetune the network with the pretrained weights we provided. If you would like to finetune the network with your own pretrained weights, you can simply provide the checkpoint via the command parameter --ckpt.

python run_cls_cmd.py --alias m40_finetune --mode finetune

In our paper, we also do experiments MidNet(Fix) in which the backbone network is fixed and only one linear classifier is trained. Run the following command.

python run_linear_cls_cmd.py --alias m40_linear

Finetune on ShapeNetPart

Follow the instructions here to train the HRNet with random initialization.

Then run the following script to finetune the network with the pretrained weights we provided. If you would like to finetune the network with your own pretrained weights, you can simply provide the checkpoint via the command parameter --ckpt.

python run_seg_shapenet_cmd.py --alias shapenet_finetune --mode finetune

In our paper, we also do experiments MidNet(Fix) in which the backbone network is fixed and only 2 FC layers are trained. Run the following command.

python run_seg_shapenet_cmd.py --alias shapenet_2fc --mode 2fc

Finetune on PartNet

Follow the instructions here to download the PartNet and preprocess the data.

Change the working directory to tensorflow/script. Run the following script to finetune the network on PartNet with the pretrained weights we provided. Compared with a random initialization, the IoU increases from 58.4 to 60.8. If you would like to finetune the network with your own pretrained weights, you can simply provide the checkpoint via the command parameter --ckpt.

python run_seg_partnet_cmd.py --alias partnet_finetune --mode finetune

In our paper, we also do experiments MidNet(Fix) in which the backbone network is fixed and only the last two FC layers are trained. Run the following command to reproduce the results.

python run_seg_partnet_cmd.py --alias partnet_fix --mode fix

The trained weights and logs can also be downloaded (6.6G) here.