Skip to content

The authors' implementation of the "Neural Head Reenactment with Latent Pose Descriptors" (CVPR 2020) paper.

License

Notifications You must be signed in to change notification settings

phygitalism/latent-pose-reenactment

 
 

Repository files navigation

Neural Head Reenactment with Latent Pose Descriptors

Burkov, E., Pasechnik, I., Grigorev, A., & Lempitsky V. (2020, June). Neural Head Reenactment with Latent Pose Descriptors. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

See the project page for an overview.

Prerequisites

For fine-tuning a pre-trained model, you'll need an NVIDIA GPU, preferably with 8+ GB VRAM. To train from scratch, we recommend a total of 40+ GB VRAM.

Set up your environment as described here.

Running the pretrained model

  • Collect images of the person to reenact.
  • Run utils/preprocess_dataset.sh to preprocess them. Read inside for instructions.
  • Download the meta-model checkpoint.
  • Run the below to fine-tune the meta-model to your person, first setting the top variables. If you want, also launch a TensorBoard at "$OUTPUT_PATH" to view progress, preferably with the --samples_per_plugin "scalars=1000,images=100" option; mainly check the "images" tab to find out at which iteration the identity gap becomes small enough.
# in this example, your images should be "$DATASET_ROOT/images-cropped/$IDENTITY_NAME/*.jpg"
DATASET_ROOT="/where/is/your/data"
IDENTITY_NAME="identity/name"
MAX_BATCH_SIZE=8             # pick the largest possible, start with 8 and decrease until it fits in VRAM
CHECKPOINT_PATH="/where/is/checkpoint.pth"
OUTPUT_PATH="outputs/"       # a directory for outputs, will be created
RUN_NAME="tony_hawk_take_1"  # give your run a name if you want

# Important. See the note below
TARGET_NUM_ITERATIONS=230

# Don't change these
NUM_IMAGES=`ls -1 "$DATASET_ROOT/images-cropped/$IDENTITY_NAME" | wc -l`
BATCH_SIZE=$((NUM_IMAGES<MAX_BATCH_SIZE ? NUM_IMAGES : MAX_BATCH_SIZE))
ITERATIONS_IN_EPOCH=$(( NUM_IMAGES / BATCH_SIZE ))

mkdir -p $OUTPUT_PATH

python3 train.py \
    --config finetuning-base                 \
    --checkpoint_path "$CHECKPOINT_PATH"     \
    --data_root "$DATASET_ROOT"              \
    --train_split_path "$IDENTITY_NAME"      \
    --batch_size $BATCH_SIZE                 \
    --num_epochs $(( (TARGET_NUM_ITERATIONS + ITERATIONS_IN_EPOCH - 1) / ITERATIONS_IN_EPOCH )) \
    --experiments_dir "$OUTPUT_PATH"         \
    --experiment_name "$RUN_NAME"

Note. TARGET_NUM_ITERATIONS is important, make sure to tune it. Pick too low, underfit and get an identity gap; pick too high, overfit and get poor mimics. I suggest that you start with 125 when NUM_IMAGES=1 and increase with more images, say, to 230 when NUM_IMAGES>30. But your concrete case may be different. If you have a lot of disk space, pass a flag to save checkpoints every so often (e.g. --save_frequency 4 will save a checkpoint every 4 * NUM_IMAGES iterations), then drive (see below how) each of them and thus find the iteration where the best tradeoff happens for your avatar.

  • Take your driving video and crop it with python3 utils/crop_as_in_dataset.py. Run with --help to learn how. Or, equivalently, just reuse utils/preprocess_dataset.sh with COMPUTE_SEGMENTATION=false.
  • Organize the cropped images from the previous step as "<data_root>/images-cropped/<images_path>/*.jpg".
  • Use them to drive your fine-tuned model (the checkpoint is at "$OUTPUT_PATH/$RUN_NAME/checkpoints") with python3 drive.py. Run with --help to learn how.

Training (meta-learning) your own model

You'll need a training configuration (aka config) file. Start with "configs/default.yaml" or just edit that. These files specify various training options which you can find in code as argparse parameters. Any of these options can be specified both in the config file and on the command line (e.g. --batch_size=7), and are resolved as follows (any source here overrides all the preceding ones):

  • argparse defaults — these are specified in the code directly;
  • those saved in a loaded checkpoint (if starting from a checkpoint);
  • your --config file;
  • command line.

The command is

python3 train.py --config=config_name [any extra arguments ...]

Or, with multiple GPUs,

python3 -um torch.distributed.launch --nproc_per_node=<number of GPUs> train.py --config=config_name [any extra arguments ...]

Reference

Consider citing us if you use the code:

@InProceedings{Burkov_2020_CVPR,
author = {Burkov, Egor and Pasechnik, Igor and Grigorev, Artur and Lempitsky, Victor},
title = {Neural Head Reenactment with Latent Pose Descriptors},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

About

The authors' implementation of the "Neural Head Reenactment with Latent Pose Descriptors" (CVPR 2020) paper.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 97.4%
  • Shell 2.6%