srl-ethz/DeformingTemplate

Environment

conda activate cadex the environment of cadex has been exported into this very directory in the file: environment.yml

running validations

for the deformation(change config number inside the file)

python deformValid.py --load

for the reconstruction(change config number inside the file)

python evaluation_ae.py 

running the trainings

example for deformation training

python deformTrainCos5ycbEnd2End.py

example for reconstruction training

python train_car_donut.py

Folders

~/Desktop/CaDeX

~/Desktop/occupancy_flow

~/Desktop/GRNet

~/Desktop/imagesFinal

~/Desktop/imagesProposal

~/Desktop/makeDataset

~/Desktop/visulaize

~/Desktop/ycb

~/Desktop/detectron2

~/Desktop/zed_camera

running cadex

python run.py --config ./configs/dfaust/testing/dfaust_w_pf_test_seen.yaml -f

Cadex timing on my computer:

During testing the batch size is 1 and the number of frames is 17:

Model_base.py runs the val_batch function for just one batch: ([1, 17, 100000, 3])

This contains 5 steps:

Prediction: 0.056

Post_process:1.5974e-5

Dataparallel_postprocess:5.483e-6

Post_process_after_optim:2.474→this contains mesh generation

First mesh

For the rest of the frames :0.365069

The mapping to first frame for all in parallel takes 0.02

There is also some clamping(clamp all vtx to unit cube)

Detach_before_optim:4.4345

So all in all for rest frames it is 0.056+0.3 which is too long(note that there is one for for which its runtime is divided by T so the runtime calculated here is for just one rest frame)

And the reported time for the all of the rest frames in the paper is 0.68(this runtime should not be divided by the number of rest frames since some of the operations on the frames are done in parallel)

running Neural_Diffeomorphic_Flow--NDF

conda activate nmf

CUDA_VISIBLE_DEVICES=0 python generate_training_meshes.py -e '/home/elham/Desktop/Neural_Diffeomorphic_Flow--NDF/pretrained/pancreas_experiments/' --debug --start_id 0 --end_id 10 --octree --keep_normalization

My code timing:

item reading: 7.152557373046875e-07

creating meshes: 0.004199981689453125

sample points: 0.007259368896484375

encode: 0.017279624938964844

decode: 0.05488896369934082

loop time: 5.7220458984375e-05

backpass time: 0.31566452980041504

how long did it take in all: 0.46349287033081055

running the occupancy flow code

conda activate see

python generate.py configs/demo.yaml

cd ~/hdd/occflow/occupancy_flow

timing for first mesh: 0.5334651470184326

rest time: 0.21221256256103516

in order to train on the 6 ycb items with 1000 deforming sequences for each python train.py ./configs/ycbTrain2.yml

Datasets

all 6 ycb objects with one deforming sequence generated for each

/home/elham/srl-nas/elham/watertight/ycb/ycb_mult_5_one_seq

just scissors with a thousand sequences of deformed objects in it

/home/elham/srl-nas/elham/watertight/ycb/ycb_mult_1_thousand_seq

all 6 ycb objects with a 1000 sequences for each /home/elham/hdd/data/ycb/ycb_mult_5_thousand_seq

Environments

Cadex : for this repo(deformTemplate) and the cadex repo

See : for the occflow repo

Nmf : for the Neural_Diffeomorphic_Flow--NDF repo

zed2: for the detectron2 repo and also the zed camera repo

mounting the drives

sshfs -o allow_other eli@129.132.57.251:/hdd/eli ~/hdd

sudo mount -t cifs -o username=eaminmans,domain=D,vers=2.0