pytorch-video-feature-extractor

Extract video features from pre-trained CNNs.

Environment

• Ubuntu 16.04
• CUDA 9.0
• cuDNN 7.3.1
• Nvidia Geforce GTX Titan Xp 12GB
• Python 3.6.8

Supported CNNs

• AlexNet
• GoogleNet
• VGGNet
• DenseNet
• MobileNet
• ResNet
• ShuffleNet
• SqueezeNet
• InceptionNet (v3 & v4)

How to use

Step 1. Install python packages specified in requirements.txt

$ python -m venv .env
$ source .env/bin/activate
(.env) $ pip install --upgrade pip
(.env) $ pip install -r requirements.txt

Step 2. Prepare videos and put them in to a directory.

For example, for my case

|-- datasets
    |-- MSVD
        |-- C9LHk0AJI7U_196_205.avi
        |-- D1tTBncIsm8_248_254.avi
        ...

Step 3. Extract CNN features from videos.

For example, to extract VGG19 features from 'MSVD' dataset,

(.env) $ python main.py \
    --video_dpath datasets/MSVD \
    --model vgg19_bn \
    --batch_size 25 \
    --stride 5 \
    --out features/MSVD_VGG19.hdf5

, or you can refer to the script files in the scripts directory.

References

• AlexNet: https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py
• GoogleNet: https://github.com/pytorch/vision/blob/master/torchvision/models/googlenet.py
• VGGNet: https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py
• DenseNet: https://github.com/pytorch/vision/blob/master/torchvision/models/densenet.py
• MobileNet: https://github.com/pytorch/vision/blob/master/torchvision/models/mobilenet.py
• ResNet: https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
• ShuffleNet: https://github.com/pytorch/vision/blob/master/torchvision/models/shufflenetv2.py
• SqueezeNet: https://github.com/pytorch/vision/blob/master/torchvision/models/squeezenet.py
• InceptionNet-v3: https://github.com/pytorch/vision/blob/master/torchvision/models/inception.py
• InceptionNet-v4: https://github.com/Cadene/pretrained-models.pytorch/blob/master/pretrainedmodels/models/inceptionv4.py

Acknowlegement

I got the codes and pre-trained models from pytorch/vision, and Cadene/pretrained-models.pytorch. Many thanks!