/PlacesCNNInference_Deakin

Primary LanguagePython

PlacesCNN Transfer Learning

The process includes 2 main steps:

  1. Extract raw feature of placesCNN after processing through avg_pooling layer (vector dimension = 512).
  2. Train a feedforward neural network to classify new classes based on pre-trained features.

Requirements:

  • Python >= 3.7
  • Torch & TorchVision >= 1.6.0

1. Installation

git clone https://github.com/nvtu/PlacesCNNTransferLearning
cd PlacesCNNTransferLearning
git clone https://github.com/nvtu/places365
pip3 install -r requirements.txt

2. Download SHREC'18 2D Scene dataset

wget http://orca.st.usm.edu/~bli/SceneIBR2018/SceneIBR2018_Dataset.zip

Unzip all file in the zip file.

3. Running Instruction

  1. Configure data and output folder path in config.ini file, for instance:
  • DATA_DIR = /mnt/DATA/nvtu/SceneIBR2018_Dataset
  • PROCESSED_DATA_DIR = /home/nvtu/PlacesCNNTraining/PROCESSED_DATA
  • MODEL_DIR = /home/nvtu/PlacesCNNTraining/MODEL
  1. Extract placesCNN features:
python3 extract_SHREC_placeCNN_features.py
  1. Prepare training data and create h5py file for fast data processing
python3 prepare_placefeat_data.py
  1. Start training
python3 train.py

--> The best model is saved in MODEL_DIR

STANDALONE PLACESCNN RUNNING

Please look at the file places365/run_placesCNN_unified.py in __main__ part to see example of standalone placesCNN running