hzhang57 / something-something-v2-baseline

Contains code to get you started with a baseline on version 2 of "something-something" dataset

something-something-v2-baseline

Contains code to get you started with a baseline on version 2 of "something-something" dataset

• Paper: https://arxiv.org/abs/1706.04261
• Data and Leaderboard: https://20bn.com/datasets/something-something/v2

Performance of pre-trained model on validation set:

Model	top-1	top-5
model3D_1	49.88%	78.82%
model3D_1_224	47.67%	77.35%
model3D_1 with left-right augmentation and fps jitter	51.33%	80.46%

Prerequisites

• Python 3.x
• PyTorch: 0.4.0 (conda installation preferred - ref https://pytorch.org/)
• torchvision
• matplotlib
• skvideo (scikit-video)
• ffmpeg
• opencv-python
• sh
• PyAV (conda install av -c conda-forge)

Setting up

Download the dataset

The dataset is provided in the form of videos in webm format using VP9 encoding, occupying a total size of 19.4 GB. The videos are in landscape format with height (the shorter side) of 240px at 12 frames/sec.

• Follow instructions on the data page
• Download the json files to fetch annotations of the data

Modify config file to include the above paths

In configuration file (located at configs/config_model1.json), modify the

• path to data: data_folder
• path to JSONs: json_data_train, json_data_val, json_data_test

How to train from scratch?

Run: CUDA_VISIBLE_DEVICES=0,1 python train.py -c configs/config_model1.json -g 0,1 --use_cuda

where,

• CUDA_VISIBLE_DEVICES: environment variable to specify GPU ids to use. (Note: uses all gpus if not specified)

Hyperparameters

Please refer to config file at: configs/config_model1.json

• batch_size: 30 - change this to fill your GPU memory (Note: should be a multiple of number of gpus used)
• num_workers: 5: number of parallel processes to fetch and pre-process data (increase to max possible CPU cores you have to get better GPU utilisation)
• lr: 0.008 - increase it if you happen to increase the batch size
• clip_size: 72 - number of frames in a video sample as input to the model (which at default 12 fps covers 6 secs)
• step_size_train: 1 - factor by which FPS is reduced (so a step size of 2 would mean an fps of 6)
• input_spatial_size: 84 - dimension of each frame in input is scaled and cropped to 84x84, but you can use the ubiquitous frame size of 224x224, since the data is provided with height of 240px in landscape format
• column_units: 512: desired number of units in feature space for each sample

How to use a pre-trained model?

• We provide a vanilla implementation of VGG-styled 3D-CNN with 11 layers of 3D convolutions. Please refer here: model3D_1.py

• Use the notebook to get predictions from this model

Test model and get submission file on test data

Modify path to model file in checkpoint variable of config file

CUDA_VISIBLE_DEVICES=0,1 python train.py -c configs/config_model1.json -g 0,1 -r -e --use_cuda

The options used here are:

• -r: to resume an already trained model
• -e: to evaluate the model on test data

Grad-CAM

Use the notebook to visualize saliency maps of any example from validation set

Commonsense score

Use the notebook to fetch commonsense score using contrastive groups list in directory assets/

For more details, please refer: https://openreview.net/pdf?id=rkX9Z_kwf

LICENSE

Most code is copyright (c) 2018 Twenty Billion Neurons GmbH under an MIT Licence. See the file LICENSE for details. Some code snippets have been taken from Keras (see LICENSE_keras) and the PyTorch (see LICENSE_pytorch). See comments in the source code for details.

References

[1] Goyal et al. ‘The “something something” video database for learning and evaluating visual common sense.’ arXiv preprint arXiv:1706.04261 (2017). In ICCV 2017.

[2] https://github.com/jacobgil/pytorch-grad-cam