Skip to content

Latest commit

 

History

History
335 lines (290 loc) · 13.9 KB

README.md

File metadata and controls

335 lines (290 loc) · 13.9 KB
Raw

Audio-Visual Generalized Few-Shot Learning with Prototype-Based Co-Adaptation

The code repository for "Audio-Visual Generalized Few-Shot Learning with Prototype-Based Co-Adaptation" [paper] in PyTorch. If you use any content of this repo for your work, please cite the following bib entry:

@inproceedings{DBLP:conf/interspeech/Zhang0YZ22,
  author    = {Yi{-}Kai Zhang and
               Da{-}Wei Zhou and
               Han{-}Jia Ye and
               De{-}Chuan Zhan},
  editor    = {Hanseok Ko and
               John H. L. Hansen},
  title     = {Audio-Visual Generalized Few-Shot Learning with Prototype-Based Co-Adaptation},
  booktitle = {Interspeech 2022, 23rd Annual Conference of the International Speech
               Communication Association, Incheon, Korea, 18-22 September 2022},
  pages     = {531--535},
  publisher = {{ISCA}},
  year      = {2022},
  url       = {https://doi.org/10.21437/Interspeech.2022-652},
  doi       = {10.21437/Interspeech.2022-652},
  timestamp = {Tue, 11 Oct 2022 19:11:50 +0200},
  biburl    = {https://dblp.org/rec/conf/interspeech/Zhang0YZ22.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Prototype-based Co-Adaptation with Transformer

Illustration of Proto-CAT. The model transforms the classification space using [公式] based on two kinds of audio-visual prototypes (class centers): (1) the base training categories (color with blue, green, and pink); and (2) the additional novel test categories (color with burning transition). Proto-CAT learns and generalizes on novel test categories from limited labeled examples, maintaining performance on the base training ones. Tsvg includes audio-visual level and category level prototype-based co-adaptation. From left to right, more coverage and more bright colors represent a more reliable classification space.

 

Results

Dataset LRW LRW-1000
Data Source Audio () Video () Audio-Video () Audio-Video ()
Perf. Measures on H-mean H-mean Base Novel H-mean Base Novel H-mean
LSTM-based 32.20 8.00 97.09 23.76 37.22 71.34 0.03 0.07
GRU-based 37.01 10.58 97.44 27.35 41.71 71.34 0.05 0.09
MS-TCN-based 62.29 19.06 80.96 51.28 61.76 71.55 0.33 0.63
MAML 35.49 10.25 40.09 66.70 49.20 29.40 23.21 25.83
BootstrappedMAML 33.75 6.52 35.29 64.20 45.17 28.15 27.98 28.09
ProtoNet 39.95 14.40 96.33 39.23 54.79 69.33 0.76 1.47
MatchingNet 36.76 12.09 94.54 36.57 52.31 68.42 0.95 1.89
MetaOptNet 43.81 19.59 88.20 47.06 60.73 69.01 1.79 3.44
DeepEMD -- 27.02 82.53 16.43 27.02 64.54 0.80 1.56
FEAT 49.90 25.75 96.26 54.52 68.83 71.69 2.62 4.89
DFSL 72.13 42.56 66.10 84.62 73.81 31.68 68.72 42.56
CASTLE 75.48 34.68 73.50 90.20 80.74 11.13 54.07 17.84
Proto-CAT (Ours) 84.18 74.55 93.37 91.20 92.13 49.70 38.27 42.25
Proto-CAT+ (Ours) 93.18 90.16 91.49 54.55 38.16 43.88

Audio-visual generalized few-shot learning classification performance (in %; measured over 10,000 rounds; higher is better) of 5-way 1-shot training tasks on LRW and LRW-1000 datasets. The best result of each scenario is in bold font. The performance measure on both base and novel classes (Base, Novel in the table) is mean accuracy. Harmonic mean (i.e., H-mean) of the above two is a better generalized few-shot learning performance measure.

 

Prerequisites

Environment

Please refer to requirements.txt and run:

pip install -r requirement.txt

Dataset

  • Use preprocessed data (suggested):

    LRW and LRW-1000 forbid directly share the preprocessed data.

  • Use raw data and do preprocess:

    Download LRW Dataset and unzip, like,

    /your data_path set in .sh file
├── lipread_mp4
│   ├── [ALL CLASS FOLDER]
│   ├── ...

    Run prepare_lrw_audio.py and prepare_lrw_video.py to preprocess data on video and audio modality, respectively. Please modify the data path in the above preprocessing file in advance.

    Similarly, Download LRW-1000 dataset and unzip. Run prepare_lrw1000_audio.py and prepare_lrw1000_video.py to preprocess it.

Pretrained Weights

We provide pretrained weights on LRW and LRW-1000 dataset. Download from Google Drive or Baidu Yun(password: 3ad2) and put them as:

/your init_weights set in .sh file
├── Conv1dResNetGRU_LRW-pre.pth
├── Conv3dResNetLSTM_LRW-pre.pth
├── Conv1dResNetGRU_LRW1000-pre.pth
├── Conv3dResNetLSTM_LRW1000-pre.pth

 

How to Train Proto-CAT

For LRW dataset, fine-tune the parameters in run/protocat_lrw.sh, and run:

cd ./Proto-CAT/run
bash protocat_lrw.sh

Similarly, run bash protocat_lrw1000.sh for dataset LRW-1000.

Run bash protocat_plus_lrw.sh / bash protocat_plus_lrw1000.sh to train Proto-CAT+.

How to Reproduce the Result of Proto-CAT

Download the trained models from Google Drive or Baidu Yun(password: swzd) and run:

bash test_protocat_lrw.sh

Run bash test_protocat_lrw1000.sh, bash test_protocat_plus_lrw.sh, or bash test_protocat_plus_lrw1000.sh to evaluate other models.

 

Code Structures

Proto-CAT's entry function is in main.py. It calls the manager Trainer in models/train.py that contains the main training logic. In Trainer, prepare_handle.prepare_dataloader combined with train_prepare_batch inputs and preprocesses generalized few-shot style data. fit_handle controls forward and backward propagation. callbacks deals with the behaviors at each stage.

Arguments

All parameters are defined in models/utils.py. We list the main ones below:

  • do_train, do_test: Store-true switch for whether to train or test.
  • data_path: Data directory to be set.
  • model_save_path: Optimal model save directory to be set.
  • init_weights: Pretrained weights to be set.
  • dataset: Option for the dataset.
  • model_class: Option for the top model.
  • backend_type: Option list for the backend type.
  • train_way, val_way, test_way, train_shot, val_shot, test_shot, train_query, val_query, test_query: Tasks setting of generalized few-shot learning.
  • gfsl_train, gfsl_test: Switch for whether train or test in generalized few-shot learning way, i.e., whether additional base class data is included.
  • mm_list: Participating modalities.
  • lr_scheduler: List of learning rate scheduler.
  • loss_fn: Option for the loss function.
  • max_epoch: Maximum training epoch.
  • episodes_per_train_epoch, episodes_per_val_epoch, episodes_per_test_epoch: Number of sampled episodes per epoch.
  • num_tasks: Number of tasks per episode.
  • meta_batch_size: Batch size of each task.
  • test_model_filepath: Trained weights .pth file path when testing a model.
  • gpu: Multi-GPU option like --gpu 0,1,2,3.
  • logger_filename: Logger file save directory.
  • time_str: Token for each run, and will generate by itself if empty.
  • acc_per_class: Switch for whether to measure the accuracy of each class with base, novel, and harmonic mean.
  • verbose, epoch_verbose: Switch for whether to output message or output progress bar.
  • torch_seed, cuda_seed, np_seed, random_seed: Seeds of random number generation.

 

Acknowledgment

We thank the following repos providing helpful components/functions in our work.