Please see logs/log_{task}_{dataset_name}_{model_name} for numbers

4. modes discovery

sh ./scripts/model/pretrain_modes_discovery.sh

5. trajectory customization

sh ./scripts/model/pretrain_trajectory_cus.sh

The dist_e for 10 different seeds can be found in the csv under ./results/endloc_customization/

Train your network on three datasets

(Caution!) it will take long time if you use default parameters

Please review the hyperparameter in ./src/param/param_{dataset_name}.py

It is better to use wandb to track it, or you can add the following line in the beginning of the main function

os.environ['WANDB_MODE'] = 'offline'

MoCap

sh ./scripts/model/train_odmo_mocap.sh

HumanAct12

sh ./scripts/model/train_odmo_humanact12.sh

UESTC

sh ./scripts/model/train_odmo_uestc.sh

Note

Each time of training, it will generate a folder with unique name (we call it model name) under the ckpt/{dataset_name} folder, please keep tracking the most recent ones or you can use wandb to track it

Inference the model

python inference_odmo_best.py {model_name} {sampling strategy} {device}

For example we want to inference mocap_abc using mode_preserve_sampling on cuda:0

python ./src/inference_odmo_best.py mocap_abc MM cuda:0

or by using the conventional sampling

python ./src/inference_odmo_best.py mocap_abc standard cuda:0

Metric based on Classifier

It's a cpu task. we need the folder name of the sampled data in ./sampled_data folder

python ./src/metrics/calculate_metric.py mocap_end_release_MM

If we want to calculate metric on UESTC dataset (it has both train/test set) we can use

python ./src/metrics/calculate_metric.py uestc_end_release_MM train
python ./src/metrics/calculate_metric.py uestc_end_release_MM test

You also can use

./src/metrics/calculate_metric_ext.py

to calculate metric from any specific sampled data

Metric on APD

This is also a CPU task

python src/metrics/calculate_apd.py mocap_end_release_MM

Similarly, we can use the following command for uestc dataset

python ./src/metrics/calculate_apd.py uestc_end_release_MM train
python ./src/metrics/calculate_apd.py uestc_end_release_MM test

Modes interpolation

The modes_discovery_interpolation.py can handle both pretrained model and the customized trained model.

For pretrained model

python ./src/customization/modes_discovery_interpolation.py {dataset_name} pretrained {use_end} {device}

For customized trained model, the model name is the name randomly generated in your ckpt/{dataset} folder

python ./src/customization/modes_discovery_interpolation.py {dataset_name} {model_name} {use_end} {device}

Trajectory customization and metrics

The trajectory_customization.py can handle both pretrained model and the customized trained model.

For pretrained model

python ./src/customization/trajectory_customization.py {dataset_name} pretrained {device}

For customized trained model, the model name is the name randomly generated in your ckpt/{dataset} folder

python ./src/customization/trajectory_customization.py {dataset_name} {model_name} {device}

The dist_e for 10 different seeds can be found in the csv under ./results/endloc_customization/

Plotting the sampled motion

If we want to plot the sampled motion from the inference*.py, we can use the

./src/draw_func/draw_gif_from_np_multi.py

in that file you can specify the datasetname, npz file name, output folder, downsample_rate and n_jobs.