Testing - Training

VLOG

Training from scratch

Feel free to train the model by yourself using the following script:

# Path to the VLOG dataset
VLOG=$VLOG # update with you own path

# Training - generic command
./training_vlog.sh <MY-RESUME>

# Training - my command
my_resume=/home/fbaradel/log_eccv18
./training_vlog.sh $my_resume

where <MY-RESUME> is the path to your resume. First you will train the object head (10 epochs) and then you will train the full model (10 epochs).

Testing

We release weights of our model pretrained on VLOG: link. Move the checkpoint to a resume directory.

# Pythonpath
PYTHONPATH=.

# Generic command
python main.py --root <LOC-VLOG-DATA> --resume <PATH-YOUR-RESUME> \
--blocks 2D_2D_2D_2.5D \
--object-head 2D \
--add-background \
--train-set train+val \
--arch orn_two_heads \
--depth 50 \
-t 4 \
-b 16 \
--cuda \
--dataset vlog \
--heads object+context \
-j 4 \
-e 

# My command
python main.py --root $VLOG --resume /home/fbaradel/logdir/eccv18_rebuttal/vlog/two_heads/object_coco_50 \
--blocks 2D_2D_2D_2.5D \
--object-head 2D \
--add-background \
--train-set train+val \
--arch orn_two_heads \
--depth 50 \
-t 4 \
-b 16 \
--cuda \
--dataset vlog \
--heads object+context \
-j 4 \
-e 

where <PATH-YOUR-RESUME> is the location of the directory where the pretrained model is located.