Collection shoaib6174/video_swin_transformer/1

Collection of Video Swin Transformers feature extractor models.

Overview

This collection contains different Video Swin Transformer [1] models. The original model weights are provided from [2]. There were ported to Keras models (tf.keras.Model) and then serialized as TensorFlow SavedModels. The porting steps are available in [3].

About the models

These models can be directly used to extract features from videos. These models are accompanied by Colab Notebooks with fine-tuning steps for action-recognition task and video-classification.

The table below provides a performance summary:

model_name	pre-train dataset	fine-tune dataset	acc@1(%)	acc@5(%)
swin_tiny_patch244_window877_kinetics400_1k	ImageNet-1K	Kinetics 400(1k	78.8	93.6
swin_small_patch244_window877_kinetics400_1k	ImageNet-1K	Kinetics 400(1k)	80.6	94.5
swin_base_patch244_window877_kinetics400_1k	ImageNet-1K	Kinetics 400(1k)	80.6	96.6
swin_base_patch244_window877_kinetics400_22k	ImageNet-12K	Kinetics 400(1k)	82.7	95.5
swin_base_patch244_window877_kinetics600_22k	ImageNet-1K	Kinetics 600(1k)	84.0	96.5
swin_base_patch244_window1677_sthv2	Kinetics 400	Something-Something V2	69.6	92.7

These scores for all the models are taken from [2].

Video Swin Transformer Feature extractors Models

Notes

The input shape for these models are [None, 3, 32, 224, 224] representing [batch_size, channels, frames, height, width]. To create models with different input shape use this notebook.

References

[1] Video Swin Transformer Ze et al. [2] Video Swin Transformers GitHub [3] GSOC-22-Video-Swin-Transformers GitHub