Awesome
CO2A
This repository contains the official implementation for CO2A from Turrisi da Costa, V. G.; Zara, G.; Rota, P.; Santos, T. O.; Sebe, N.; Murino, V.; Ricci, E. - "Dual Contrastive Domain Adaptation for Video Action Recognition" in Proc. of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022.
Mixamo dataset
The current version of the Mixamo dataset and the respective Kinetic subset can be downloaded here.
Repository structure
The repository is structured in the following way:
- :file_folder:
src: source code - :large_blue_diamond:
bash_scripts: scripts to reproduce results in the paper
Requirements
numpy
torch
torchvision
pytorch-lightning
pytorch-lightning-bolts
Pillow
tqdm
wandb
Required python modules can be installed by running
pip3 install -r requirements.txt
Usage
Experiments can be launched by running the scripts in bash_scripts
The framework assumes the required pretrained weights files to be stored in a directory called pretrained, in the repository root.
You can download it with
wget --no-check-certificate 'https://docs.google.com/uc?export=download&id=1uPvokqwmWlLf3ihUF4spYNEPecDMFA5A' -O pretrained/rgb_imagenet.pt
Or in the the following link.
The script can be run with additional flags; we report here the most important ones:
--name ----> name of wandb run - STRING
--project ----> name of wandb project - STRING
--epochs ----> number of epochs - INT - default: 300
--optimizer ----> optimizer - STRING - choices: ["sgd", "adam"] - default: "adam"
--weight_decay ----> weight decay value - FLOAT - default: 1e-6
--scheduler ----> scheduler type - STRING - choices: ["reduce", "cosine", "step", "exponential", "none"] - default: "reduce"
--n_frames ----> number of frames to use for each clip - INT - default: 16
--n_clips ----> number of clips to use - INT - default: 4
--frame_size ----> size of one frame - INT - default: 64
--batch_size ----> batch size - INT - default: 32
--num_workers ----> number of parallel jobs - INT - default: 4
--video_dropout ----> droupout at video-level layer - FLOAT - default: 0.5
--ce_loss_weight ----> cross-entropy loss weight - FLOAT - default: 1.0
--ce_loss_target_weight ----> cross-entropy loss weight for target - FLOAT - default: 0.0
--nce_loss_target_aug_based_weight ----> augmentation based nce weight - FLOAT - default: 0.0
--nce_loss_source_label_based_weight ----> source label based nce weight - FLOAT - default: 0.0
--nce_loss_target_label_based_weight ----> target label based nce weight - FLOAT - default: 0.0
--nce_loss_inter_domain_weight ----> inter domain loss weight - FLOAT - default: 0.0
--nce_loss_target_clip_aug_based_weight ----> clip augmentation nce weight - FLOAT - default: 0.0
--consistency_loss_weight ----> consistency loss weight - FLOAT - default: 0.0
--consistency_threshold ----> threshold value to use for consistency - FLOAT - default: 0.5
--temperature ----> temperature for the losses - FLOAT - 0.2
--selection_factor ----> factor for selecting instances according to confidence - INT - default: 6
--complete_nce_weight ----> complete nce weight - FLOAT - default: 0.0
--target_2_augs ----> whether to apply the 2 target augmentations
--target_augmentations ----> list of target augmentations - LIST - default: []
--supervised_labels ----> use ground-truth target labels for contrastive loss
--source_only ----> whether to train on source only (disable DA)
--pretrained ----> use imagenet+kinetics pre-training
--imagenet_pretrained ----> use imagenet pretraining
--mixamo14_pretrained ----> use mixamo pretraining
--clip_aggregation ----> clip aggregation strategy - STRING - choices: ["avg", "lstm", "lstm_weights", "mlp", "mlp_weights", "multihead_attention", "multihead_attention_att_only"]
--gpus ----> list of gpus to use - INT
--distributed_backend ----> type of distribute backend - STRING - choices: ["dp", "ddp"] - default: "ddp"
--wandb ----> whether to log on wandb