Skip to content

Latest commit

 

History

History
110 lines (83 loc) · 4.02 KB

training.md

File metadata and controls

110 lines (83 loc) · 4.02 KB
Raw

Training BEDLAM-CLIFF

Prepare data

To download BEDLAM and AGORA training images and ground truth labels, run the following script. Since, the data size is huge it will take a while to finish. Please first register on BEDLAM website.

bash fetch_training_data.sh

Once downloaded, you can uncompress the data in the same directory.

If you have not yet prepare the data for demo run, then you need to follow data preparation instructions in Demo. This will download all the utility files need to run the training.

Finally download the HRNet checkpoint pretrained on COCO images from here and put it in data/ckpt/pretrained.

To use 3DPW as validation dataset during training, you need to download 3DPW images and save them in data/test_images/3DPW/imageFiles

Train

BEDLAM-CLIFF MODEL

python train.py --cfg configs/bedlam_cliff.yaml

BEDLAM-HMR MODEL

python train.py --cfg configs/bedlam_hmr.yaml

Finetune with 3DPW

If you want to finetune the model with 3DPW training data, you can run the following script.

python train.py --cfg configs/bedlam_cliff_3dpw_ft.yaml --resume --ckpt data/ckpt/bedlam_cliff.ckpt

Note that you need to download 3DPW training images from 3DPW and place them in data/real_training_images/3DPW/imageFiles. You can also provide your own image path in train/core/config.py '3dpw-train-smplx': 'data/real_training_images/3DPW'.

Training Hand Model

Prepare data

Register on BEDLAM website and then run the following script:

bash fetch_hand_training_data.sh

Download the HRNet pretrained Imagenet checkpoint from here and place it in data/ckpt/pretrained

Once the data is downloaded, you can train hand model with following command.

python train_hands.py --cfg configs/hands.yaml

Training BEDLAM-CLIFF-X

Prepare data

Register on BEDLAM website and then run the following script:

bash fetch_hand_training_data.sh

BEDLAM-CLIFF-X

For training BEDLAM-CLIFF-X, you can either use the pretrained checkpoint for body and hand model stored in data/ckpt or train them from scratch as described above. Once you have the checkpoints for BEDLAM-CLIFF and hand model, you can train BEDLAM-CLIFF-X

python trainx.py --cfg configs/bedlam_cliff_x.yaml --hand_ckpt data/ckpt/hands_with_agora.ckpt --body_ckpt data/ckpt/bedlam_cliff.ckpt

Training CLIFF model with real images

Prepare data

Register on BEDLAM website and then run the following script:

bash fetch_real_training_data.sh

Also download the training images from respective dataset websites in data/real_training_images COCO (2014 Train Images), 3DPW, MPII, MPI-INF-3DHP, H3.6M. After the download, you should have the following structure in your data folder.

${data}  
|-- real_training_images
|   |-- 3DPW
|   |   |-- imageFiles
|   |-- coco
|   |   |-- train2014
|   |-- mpii
|   |   |-- images
|   |-- h36m
|   |   |-- Images
|   |   |   |-- S1 
|   |   |   |-- S2 ..
|   |-- mpi-inf-3dhp
|   |   |-- S1
|   |   |-- S2 ..

Please note that if you don't want to follow the directory structure, you can also modify the path to the image folder in train/core/config.py

CLIFF

With the following command you could train the CLIFF model with real images.

python train_smpl.py --cfg configs/orig_cliff.yaml

Finetune with 3DPW

If you want to finetune the model with 3DPW training data, you can run the following script.

python train_smpl.py --cfg configs/orig_cliff_3dpw_ft.yaml --resume --ckpt data/ckpt/cliff.ckpt