Skip to content

Latest commit

 

History

History
133 lines (101 loc) · 4.3 KB

DockerREADME.md

File metadata and controls

133 lines (101 loc) · 4.3 KB

Introduction

How Container Built

  • Use Base Anaconda3 image, install all the modules currently in use.

  • Code is using a branch/PR that supports python3.

  • Image copy overs dlib model + snapshot

  • Copy over code and implementations

Info about docker-anaconda 3

Link to dockerhub
Anaconda distribution is installed into the /opt/conda
Ensures that the default user has the conda command in their path.

List installed packages in Conda

Solution Link Conda is installed in 'pytorchNew' envrionment

Commands

conda activate pytorchNew
conda list -e > requirements.txt
conda create --name <env> --file requirements.txt  # Rebuild Environment

Build + Push Container

docker build -t 16fb/deepheadpose:ZX .
docker push 16fb/deepheadpose:ZX

## Naming Conventions:
:GPU -> GPU Variant
:ZX -> ZX Variant with his code

Testing Container

Run Container

  • --gpus all to enable GPU
  • -it run interactively
  • -v mount new video to container
  • conda activate pytorch activate anaconda envrionment
docker run -it --gpus all -v ${PWD}/toMount:/home/deep-head-pose/mount 16fb/deepheadpose:GPU

conda activate pytorch

Run using old model

Use I assume use "/" because its Linux fs

python code/test_on_video_dlib.py --snapshot models/hopenet_robust_alpha1.pkl --face_model dlib/mmod_human_face_detector.dat  --video conan-cruise.gif --fps 15 --n_frames 10

Use New Model:

python code/test_on_video_dlib.py --snapshot models/mysnap_epoch_29.pkl --face_model dlib/mmod_human_face_detector.dat  --video conan-cruise.gif --fps 15 --n_frames 100

Use New Model: On new mounted Video:

python code/test_on_video_dlib.py --snapshot models/mysnap_epoch_29.pkl --face_model dlib/mmod_human_face_detector.dat  --video mount/Kamala.gif --fps 15 --n_frames 100

How to run on different image

Ideally using bind mounts:
-v <Source Directory>:<Container Directory>
Place .gif into "toMount" directory, then bind mount into container as "mount/".
-> Reference the new gif for data.

Bind /toMount directory to WORKDIR/mount

So files can be read and written to and from Host
-v /toMount:/home/deep-head-pose/mount

Extract final output from container to host directory

Copy the output video directory into mount/video
cp -r /home/deep-head-pose/output/video /home/deep-head-pose/mount/video

Import + Export Image to .tar file

Theres no progress bar:

Save to tar file

  • docker save --output <FileName> <ImageName>
  • docker save --output deepheadpose 16fb/deepheadpose:latest docker save --output deepheadpose 16fb/deepheadpose:ZX

Load from tar file

  • docker load --input <FileName>
  • docker load --input deepheadpose

Testing without docker

Testing base video deep-head-pose

conda activate pytorchNew
python code/test_on_video_dlib.py --snapshot models/mysnap_epoch_29.pkl --face_model dlib/mmod_human_face_detector.dat --video conan-cruise.gif --fps 15 --n_frames 100

Notes:
Its quite slow for video

Testing ZX Video deep-head-pose

conda activate pytorchNew
python code/i.py --snapshot models/mysnap_epoch_29.pkl --face_model dlib/mmod_human_face_detector.dat --video conan-cruise.gif --fps 24 --n_frames 100

Notes:
Had to set face model to take in from args.
Still quite slow.
It does show obtained tensors.

Testing ZX Webcam deep-head-pose

conda activate pytorchNew
python code/i_webcam.py --snapshot models/mysnap_epoch_29.pkl --face_model dlib/mmod_human_face_detector.dat --video conan-cruise.gif --fps 15 --n_frames 100

Notes:
Very Very Very Slow.

Removed the wait key at the end of i_webcam.py.
no dice.

Attempts at making it run faster.

test env, which wants use latest version of libs

conda create --name test conda activate test conda install python pytorch torchvision torchaudio cudatoolkit=11.0 -c pytorch conda install opencv matplotlib pandas scipy scikit-image cmake dlib -c conda-forge

ZX anaconda

conda create --name ZX --file conda.txt conda install torchvision

Thers no torchvision... what? no ski-image