Human Abnormal Behaviors Detection Using X3D

Some detection results on CCTV_FIGHTS and RWF-2000 test sets

To install Horovod with NCCL support on the cluster:

1.Activate the python virtual environment

2.Use pip to install CMake

3.Use module load to load gcc8/8.4.0 and nccl2-cuda10.2-gcc/2.7.8 and openmpi-geib-cuda10.2-gcc/4.0.5 (which are already installed on the cluster)

4.Since gcc8 and NCCL are already installed on the cluster, run this command (make sure to specify your path to python virtual environment):

HOROVOD_WITH_GLOO=1 HOROVOD_WITH_MPI=1 HOROVOD_WITH_PYTORCH=1 HOROVOD_WITHOUT_TENSORFLOW=1 HOROVOD_WITHOUT_MXNET=1 HOROVOD_CMAKE={Path to python virtual environment}/bin/cmake HOROVOD_GPU_OPERATIONS=NCCL HOROVOD_NCCL_HOME=/cm/shared/apps/nccl2-cuda10.2-gcc/2.7.8 pip install --no-cache-dir horovod[pytorch]

To check if horovod is installed with NCCL correctly, use:

horovodrun --check-build

if it is installed correctly, the information should be shown as below:

Available Frameworks:
    [ ] TensorFlow
    [X] PyTorch
    [ ] MXNet

Available Controllers:
    [X] MPI
    [X] Gloo

Available Tensor Operations:
    [X] NCCL
    [ ] DDL
    [ ] CCL
    [X] MPI
    [X] Gloo