v1.1

TensorFlow SavedModel formet weights. Details.

v1.0

Checkpoints of Video-FocalNet in Keras

Checkpoints of Video-FocalNet model in keras. The pretrained weights are ported from official pytorch model. Following are the list of all available model in .h5 format.

Checkpoint Naming Style

For the variation and brevity, the general format is:

# K400  : Kinetics-400
# K600  : Kinetics-600
# SSV2  : Something-Something-V2
# D48   : Driving-48
# ANET  : ActivityNet
dataset = 'K400' # K400, K600, SSV2, D48, ANET
size = 'B' # S, T, B
num_frames = 8
input_size = 224 

>> checkpoint_name = (
   f'TFVideoFocalNet{size}'
   f'{dataset}_'
   f'{num_frames}x{input_size}.h5'
)
>> checkpoint_name 
TFVideoFocalNetB_K400_8x224.h5

Here, size represent the base (B), small (S), and tiny (T) version of video-focalnet. Officially, there are 3 checkponts (B/S/T) for Kinetics-400 (K400) dataset and rest of the dataset have only base (B) checkpoint.

Highlights

Reference from model_config.py

Checkpoints	Usage
FocalNetT_K400 FocalNetS_K400 FocalNetB_K400 FocalNetB_K600 FocalNetB_SSV2 FocalNetB_D48 FocalNetB_ANET	from videofocalnet import VideoFocalNetT from videofocalnet import VideoFocalNetS from videofocalnet import VideoFocalNetB model_name = 'FocalNetT_K400' vfocal_net = VideoFocalNetT(name=model_name) y = vfocal_net (np.ones(shape=(1, 8, 224, 224, 3))) vfocal_net.load_weights('TFVideoFocalNetT_K400_8x224.h5') print(y.shape) # TensorShape([1, 400])