2. Fine-tuning SOTA video models on your own dataset

Fine-tuning is an important way to obtain good video models on your own data when you don’t have large annotated dataset or don’t have the computing resources to train a model from scratch for your use case. In this tutorial, we provide a simple unified solution. The only thing you need to prepare is a text file containing the information of your videos (e.g., the path to your videos), we will take care of the rest. You can start fine-tuning from many popular pre-trained models (e.g., I3D, R2+1D, SlowFast and TPN) using a single command line.

Custom DataLoader

The first and only thing you need to prepare is the data annotation files train.txt and val.txt. We provide a general dataloader for you to use on your own dataset. Your data can be stored in any hierarchy, and the train.txt should look like:

video_001.mp4 200 0
video_001.mp4 200 0
video_002.mp4 300 0
video_003.mp4 100 1
video_004.mp4 400 2
video_100.mp4 200 10

As you can see, there are three items in each line, separated by spaces. The first item is the path to your training videos, e.g., video_001.mp4. The second item is the number of frames in each video. But you can put any number here because our video loader will compute the number of frames again automatically during training. The third item is the label of that video, e.g., 0. val.txt looks the same as train.txt in terms of format.

Once you prepare the train.txt and val.txt, you are good to go. In this tutorial, we will use I3D model and Something-something-v2 dataset as an example. Suppose you have Something-something-v2 dataset and you don’t want to train an I3D model from scratch. First, prepare the data anotation files as mentioned above. Second, follow this configuration file i3d_resnet50_v1_custom.yaml. Specifically, you just need to change the data paths and number of classes in that yaml file.

TRAIN_ANNO_PATH: '/home/ubuntu/data/sthsthv2/sthsthv2_train.txt'
VAL_ANNO_PATH: '/home/ubuntu/data/sthsthv2/sthsthv2_val.txt'
TRAIN_DATA_PATH: '/home/ubuntu/data/sthsthv2/20bn-something-something-v2/'
VAL_DATA_PATH:  '/home/ubuntu/data/sthsthv2/20bn-something-something-v2/'

If you want to tune other parameters, it is also easy to do. Change the learning rate, batch size, clip lenght according to your use cases. Usually a small learning rate is preferred since the model initialization is decent.

We also support finetuning on other models, e.g., resnet50_v1b_custom.yaml, slowfast_4x16_resnet50_custom.yaml, tpn_resnet50_f32s2_custom.yaml, r2plus1d_v1_resnet50_custom.yaml, i3d_slow_resnet50_f32s2_custom.yaml. Try fine-tuning these SOTA video models on your own dataset and see how it goes.

If you would like to extract good video features on your datasets, feel free to read the next tutorial on feature extraction.

Total running time of the script: ( 0 minutes 0.000 seconds)

Gallery generated by Sphinx-Gallery