This repository is mainly built upon Pytorch and Pytorch-Lightning. We wish to maintain a collections of scalable video transformer benchmarks, and discuss the training recipes of how to train a big video transformer model.

Now, we implement the TimeSformer, ViViT and MaskFeat. And we have pre-trained the TimeSformer-B, ViViT-B and MaskFeat on Kinetics400/600, but still can't guarantee the performance reported in the paper. However, we find some relevant hyper-parameters which may help us to reach the target performance.

1. We have fixed serval known issues and now can build script to pretrain MViT-B with MaskFeat or finetune MViT-B/TimeSformer-B/ViViT-B on K400.
2. We have reimplemented the methods of hog extraction and hog prediction in MaskFeat which are currently more efficient to pretrain.
3. Note that if someone want to train TimeSformer-B or ViViT-B with current repo, they need to carefully adjust the learning rate and weight decay for a better performance. For example, you can can choose 0.005 for peak learning rate and 0.0001 for weight decay by default.

1. Difference
2. TODO
3. Setup
4. Usage
5. Result
6. Acknowledge
7. Contribution

In order to share the basic divided spatial-temporal attention module to different video transformer, we make some changes in the following apart.

We split the position embedding from R(n^t*h*w×d) mentioned in the ViViT paper into R(n^h*w×d) and R(n^t×d) to stay the same as TimeSformer.

In order to make clear whether to add the class_token into the module forward computation, we only compute the interaction between class_token and query when the current layer is the last layer (except FFN) of each transformer block.

• Tokenization: the token embedding filter can be chosen either Conv2D or Conv3D, and the initializing weights of Conv3D filters from Conv2D can be replicated along temporal dimension and averaging them or initialized with zeros along the temporal positions except at the center t/2.
• Temporal MSA module weights: one can choose to copy the weights from spatial MSA module or initialize all weights with zeros.
• Initialize from the MAE pre-trained model provided by ZhiLiang, where the class_token that does not appear in the MAE pre-train model is initialized from truncated normal distribution.
• Initialize from the ViT pre-trained model can be found here.

• [√] add more TimeSformer and ViViT variants pre-trained weights.
  • A larger version and other operation types.
• [√] add linear prob and finetune recipe.
  • Make available to transfer the pre-trained model to downstream task.
• add more scalable Video Transformer benchmarks.
  • We will mainly focus on the data-efficient models.
• add more robust objective functions.
  • Pre-train the model through the dominated self-supervised methods, e.g Mask Image Modeling.

pip install -r requirements.txt

# path to Kinetics400 train set and val set
TRAIN_DATA_PATH='/path/to/Kinetics400/train_list.txt'
VAL_DATA_PATH='/path/to/Kinetics400/val_list.txt'
# path to root directory
ROOT_DIR='/path/to/work_space'
# path to pretrain weights
PRETRAIN_WEIGHTS='/path/to/weights'

# pretrain mvit using maskfeat
python model_pretrain.py \
	-lr 8e-4 -epoch 300 -batch_size 16 -num_workers 8 -frame_interval 4 -num_frames 16 -num_class 400 \
	-root_dir $ROOT_DIR -train_data_path $TRAIN_DATA_PATH

# finetune mvit with maskfeat pretrain weights
python model_pretrain.py \
	-lr 0.005 -epoch 200 -batch_size 8 -num_workers 4 -num_frames 16 -frame_interval 4 -num_class 400 \
	-arch 'mvit' -optim_type 'adamw' -lr_schedule 'cosine' -objective 'supervised' -mixup True \
	-auto_augment 'rand_aug' -root_dir $ROOT_DIR -train_data_path $TRAIN_DATA_PATH \
	-val_data_path $VAL_DATA_PATH -pretrain_pth $PRETRAIN_WEIGHTS

# finetune timesformer with imagenet pretrain weights
python model_pretrain.py \
	-lr 0.005 -epoch 30 -batch_size 8 -num_workers 4 -num_frames 8 -frame_interval 32 -num_class 400 \
	-arch 'timesformer' -attention_type 'divided_space_time' -optim_type 'sgd' -lr_schedule 'cosine' \
	-objective 'supervised' -root_dir $ROOT_DIR -train_data_path $TRAIN_DATA_PATH \
	-val_data_path $VAL_DATA_PATH -pretrain_pth $PRETRAIN_WEIGHTS -weights_from 'imagenet'

# finetune vivit with imagenet pretrain weights
python model_pretrain.py \
	-lr 0.005 -epoch 30 -batch_size 8 -num_workers 4 -num_frames 16 -frame_interval 16 -num_class 400 \
	-arch 'vivit' -attention_type 'fact_encoder' -optim_type 'sgd' -lr_schedule 'cosine' \
	-objective 'supervised' -root_dir $ROOT_DIR -train_data_path $TRAIN_DATA_PATH \
	-val_data_path $VAL_DATA_PATH -pretrain_pth $PRETRAIN_WEIGHTS -weights_from 'imagenet'

The minimal folder structure will look like as belows.

root_dir
├── results
│   ├── experiment_tag
│   │   ├── ckpt
│   │   ├── log

For each column, we show the masked input(left), HOG predictions(middle) and original video frame(right).

Here, we show the extracted attention map of a random frame sampled from the demo video.

tips: frame_interval and data augment counts for the validation accuracy.

tips: Improve the frame_interval will drop a lot on time performance.

1.speed up training recipes:

• More GPU device.
• pin_memory=True.
• Avoid CPU->GPU Device transfer (such as .item(), .numpy(), .cpu() operations on tensor or log to disk).

2.get_batch first means that we firstly read all frames through the video reader, and then get the target slice of frames, so it largely slow down the data-loading speed.

this repo is built on top of Pytorch-Lightning, pytorchvideo, skimage, decord and kornia. I also learn many code designs from MMaction2. I thank the authors for releasing their code.

I look forward to seeing one can provide some ideas about the repo, please feel free to report it in the issue, or even better, submit a pull request.

And your star is my motivation, thank u~