Nguyen Hai Duong ([email protected])
Prlab - Chonnam National University

• In Windows 7, type cd /d path to change working directory
• In Windows 10, type cd path instead

1. Windows 7/10 Pro x64
2. Anaconda 4.2 Python 3.5
3. TensorFlow 1.1.0 with Keras interface
4. CUDA 8.0
5. cuDNN v5
6. OpenCV
7. Libraries: natsort, pydot, sklearn, libplotmat (using pip install ...), and graphviz (after installation, add C:\Program Files (x86)\Graphviz2.38\bin to SYSTEM PATH!)

1. Close all cmd windows
2. Open cmd as administrator again
3. Enter conda create -n tensorflow python=3.5
4. Enter activate tensorflow
5. Enter pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/windows/gpu/tensorflow_gpu-1.1.0-cp35-cp35m-win_amd64.whl
6. Validate installation by entering python -c "import tensorflow as tf; print(tf.__version__)"

1. Enter conda install -c https://conda.binstar.org/menpo opencv3
2. Validate installation by entering python -c "import cv2; print(cv2.__version__)"

1. (Recommended) Download and extract already extracted face images from \\168.131.152.92\emotion\Duong\AFEW_7_2017\face_images to a folder (D:\AFEW2017 for example)

• afew2017_aligned_extracted_faces_cascade_cnn.rar contains ALIGNED face images:
  
• afew2017_aligned_extracted_faces_cascade_cnn.rar contains UNALIGNED face images:
  

2. (Skip this step in case you've already downloaded images from option 1) Manually extract face images:

• Download and extract original AFEW2017 dataset from \\168.131.152.92\emotion\Duong\AFEW_7_2017\2017_EmotiW.zip to a folder, such as D:\AFEW2017
• Download video converter from \\168.131.152.92\emotion\Duong\AFEW_7_2017\avi_to_mp4.py and save it in D:\AFEW2017\Train_AFEW and D:\AFEW2017\Val_AFEW
• Open cmd:
  • Type cd /d D:\AFEW2017\Train_AFEW, and enter python avi_to_mp4.py
  • Type cd /d D:\AFEW2017\Val_AFEW, and enter python avi_to_mp4.py
• Install mxnet
  • Download prebuildbase_win10_x64_vc14.7z and 20170718_mxnet_x64_vc14_gpu.7z from \\168.131.152.92\emotion\Duong\AFEW_7_2017\mxnet
  • Extract prebuildbase_win10_x64_vc14.7z to a folder, such as D:\mxnet
  • Extract and overwrite 20170718_mxnet_x64_vc14_gpu.7z to the above folder
  • Run setupenv.cmd as Administrator
  • Open cmd as Administrator
  • Type cd /d D:\mxnet\python
  • Enter python setup.py install
  • Press python -c "import mxnet; print(mxnet.__version__)" to validate installation
• Download mtcnn_detector.py, and afew2017_face_extraction_from_mp4s.ipynb from this repo and save them in a folder (called E2017)
• Download Cascade CNN model from \\168.131.152.92\emotion\Duong\AFEW_7_2017\mxnet\cascade_cnn.rar and extract all the contents into E2017\model
• Open cmd, type cd /d E2017
• Enter jupyter notebook to run Jupyter IDE
• Open afew2017_face_extraction_from_mp4s.ipynb
• Modify PATH = 'E:/EmotiW2017/Val_AFEW/' + video_path, and PATH2 = 'E:/EmotiW2017/lstm/Val_AFEW/' + video_path based on the extracted AFEW2017 dataset path on your computer
• Run All Cell (Cell > Run All Cell)
• Change Val_AFEW to Train_AFEW in PATH and PATH2. Run all cell again to extract faces from the training set

Note that: the extracted images will be saved in PATH2

1. Download afew2017_cnn_lstm_with_data_processing.ipynb from this repo and save it to a folder (called E2017)
2. Open cmd, type cd /d E2017
3. Enter jupyter notebook to run Jupyter IDE
4. Open afew2017_cnn_lstm_with_data_processing.ipynb
5. The training results are already done in In [56]-In [59]
6. In order to retrain:

• modify PATH = 'E:/EmotiW2017/lstm/original/Train_AFEW', and PATH = 'E:/EmotiW2017/lstm/original/Val_AFEW' based on the extracted face images path on your computer
• Run All Cell (Cell > Run All Cell)

Backup M:\CNU\EmotiW\2016\Face_detection\caffe_python_mtcnn-master