Model Architecture

Train GANimator

data.yaml
main.yaml
model.yaml
optim.yaml
scheduling.yaml
util.yaml

Data

GANimator uses a single training sample and this is the reason we use a loader that repeatedly loads the same .npz file for both the train and the validation processes:

data.yaml

# @package _global_

npz_filename: ???
train_data_iter: 15000

data:
  train:
    iterator:
      datasets:
        repeated_npz:
          filename: ${npz_filename}
          length: ${train_data_iter}

  val:
    loader:
      batch_size: 1
    iterator:
      datasets:
        repeated_npz:
          filename: ${npz_filename}
          length: 200

Model Architecture

The model consists of 4 stages, with the first 3 stages comprising 2 generator and 2 discriminator, while the last stage includes 1 generator and 1 discriminator. The overall structure of the stage is the following:

model.yaml

generator_sX: #combined generators for stages 1-3
  noise0: [noise_level_0]
  noise1: [noise_level_1]
  generated: [prev]
  _out_: [fake]

discriminator_sX: #combined discriminator for stages 1-3
  fake0:
    - fake.stage0
    - motion_data_level_x
  fake1:
    - fake.stage1
    - motion_data_level_y
  _out_:
    - fake_score
    - real_score

The generation part of the stage receives an amplified noisy input for the 1st generator (noise0) and another one for the 2nd generator (noise1), the generated motion features of the previous stage (generated) and returns 2 generated (fake) motion features - one per generator. The corresponding discrimination part receives the 2 generated features (fake0, fake1) and the corresponding real ones (motion_data_level_x, motion_data_level_y) and returns a score for both.

Gradual Learning

One major difference with the original GANimator implementation is the ability to train all strages sequentially as an end-to-end process. This is achieved by exploiting moai’s scheduling functionality:

metrics.yaml

# @package _global_
_moai_:
  _execution_:
    _schedule_:
      - _epoch_: 1
        _modifications_:
          forward_mode:
            modules:
              generator_s1: eval_nograd
        _fit_:
          _stages_:
            process_disc:
              _optimizer_: optim_disc_stage_1
              _flows_:
                - horizontal_input_mapping
                - stage_1
                - generator_s1
                - stage_2
                - generator_s2
                - disc_prediscrimination
                - discriminator_s2
            process_sample:
              _optimizer_: optim_gen_stage_1
              _flows_:
                - horizontal_input_mapping
                - stage_1
                - generator_s1
                - stage_2
                - generator_s2
                - sample_prediscrimination
                - discriminator_s2
            process_reco:
              _optimizer_: optim_gen_stage_1
              _flows_:
                - horizontal_input_mapping
                - stage_1
                - stage_1_reco
                - generator_s1
                - stage_2
                - stage_2_reco
                - generator_s2
                - reco_prediscrimination
        _val_:
          _datasets_:
            repeated_npz:
              _flows_:
                - stage_1_val
                - generator_s1
                - stage_2_val
                - generator_s2
                - val_prediscrimination
                - val_prediscrimination_post

The YAML snippet depicts the flow for training the 2nd stage given that generator(s) s1 are trained and frozen. Note that we use _epoch_ to indicate the stage number (i.e., _epoch_: 0 corresponds to the 1st stage).

Logging & Visualization

The used loss functions and validation metrics are logged to our local ClearML server, while the generated motion features are converted to a Mixamo skeleton representation and are visualilzed in the Rerun viewer.

util.yaml

engine:
  modules:
    clearml:
      project_name: GANimator
      task_name: ${experiment.name}
      tags: [train]
    rerun:
      annotations:
        parents:
          mixamo:
            [-1, 0, 1, 2, 3, 4, 0, 6, 7, 8, 9, 0, 11, 12, 13, 14, 13, 16, 17, 18, 13, 20, 21, 22]
        labels:
          mixamo:
            - "Hips" # 0
            - "LeftUpLeg" # 1
            - "LeftLeg" # 2
            - "LeftFoot" # 3
            - "LeftToeBase" # 4
            - "LeftToe_End" # 5
            - "RightUpLeg" # 6
            - "RightLeg" # 7
            - "RightFoot" # 8
            - "RightToeBase" # 9
            - "RightToe_End" # 10
            - "Spine" # 11
            - "Spine1" # 12
            - "Spine2" # 13
            - "Neck" # 14
            - "Head" # 15
            - "LeftShoulder" # 16
            - "LeftArm" # 17
            - "LeftForeArm" # 18
            - "LeftHand" # 19
            - "RightShoulder" # 20
            - "RightArm" # 21
            - "RightForeArm" # 22
            - "RightHand" # 23

To visualize the skeleton all we need is the joint positions and the Mixamo kinematic tree information.

To train the GANimator model, one should run the following command:

python -m moai run fit conf/run/main.yaml npz_filename=%input_npz_file_name%

Last updated on September 19, 2024

Data Preprocessing Motion Generation