AoT_TCAM

[project page] [dataset and pre-process tools]

Torch model for arrow of time prediction in the CVPR 18 paper

D. Wei, J. Lim, A. Zisserman, W. Freeman. "Learning and Using the Arrow of Time." in CVPR 2018.

Demo:

1. training from scratch flow-TCAM model for AoT prediction on UCF101

CUDA_ID=0,1,2;GPU_ID=1,2,3;CPU_N=10;
N_BATCH=32;N_FRAME=20;N_C=2
M_LABEL=2;M_LOSS=0;M_BEND=cudnn
M_DROPOUT=0.5,0.5;M_ARCH=vggbn_tcam_pair2
D_TYPE=2;D_PRE=5;D_NAME=flow%s_%04d.jpg;D_CROP=2;D_TCROP=1;
D_OUT=results/ucf_train/
D_TXT=data/@01_cnn_ta_flow_orig_fb.txt
E_SAVE=5;E_ALL=20;E_ST=1;E_SIZE=5000;E_ITER=3
E_PARAM=5.1

CUDA_VISIBLE_DEVICES=${CUDA_ID} th main_video.lua -GPUs ${GPU_ID} -nDonkeys ${CPU_N} \
    -batchSize ${N_BATCH} -vnF ${N_FRAME} -vnC ${N_C} \
    -lossType ${M_LOSS} -nClasses ${M_LABEL}  -Mdropout ${M_DROPOUT} -netType ${M_ARCH}  -backend ${M_BEND} \
    -cropType ${D_CROP} -TcropType ${D_TCROP} -cache ${D_OUT} -data ${D_TXT} -xType ${D_PRE} -imType ${D_TYPE} -imFormat ${D_NAME} \
    -retrain train -retrainOpt train -paramId ${E_PARAM} \
    -nEpochs ${E_ALL} -epochNumber ${E_ST} -epochSave ${E_SAVE} -epochSize ${E_SIZE} -iter_size ${E_ITER} 2>&1 | tee ${Dout}/log-${N_BATCH}-${N_FRAME}-${N_C}-${M_DROPOUT}-${M_ARCH}-${D_TYPE}-${D_PRE}-${D_CROP}-${D_TCROP}-${E_ALL}-${E_ST}-${E_ITER}.log

Tips:

• to debug in lua, add this line: local dbg = require('util/debugger');dbg()

• to visualize data in matlab, add this line: local matio = require 'matio'; matio.save('test.mat',{t1=data1,t2=data2})

Reference Torch Packages:

1. [multi-gpu trainining]
2. [debugger]
3. [matlab data i/o]

Citation

Please cite our paper if you find it useful for your work:

@inproceedings{wei2018learning,
  title={Learning and Using the Arrow of Time},
  author={Wei, Donglai and Lim, Joseph J and Zisserman, Andrew and Freeman, William T},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={8052--8060},
  year={2018}
}