README.md

PSGNets

Models for converting image-like inputs into hierarchical graph-like scene representations

Installation Instructions:

0. Create a python2 virtual environment (highly recommended)

   cd [WORKING_DIR]
   virtualenv env -p python
   source env/bin/activate
   

1. Clone and install the psgnets package:

   git clone [email protected]:neuroailab/PSGNets.git
   cd PSGNets && pip install -e .
   

2. Check installation by training a ResNet18 on ImageNet categorization [Note: this uses the TFutils training script]:

   cd psgnets/trainval
   python train_psgnet_tfutils.py --gpus [X] --config_path ./training_configs/resnet_config.py --exp_id [MY_EXP_ID] --batch_size 256 --save_dir ./training_configs/tfutils_params --port [PORT_TO_MONGODB] [--data_dir [PATH_TO_IMAGENET]]
   

where [X] is the number of a free GPU on the node you're on, [MY_EXP_ID] is a name for your training experiment, [PORT_TO_MONGODB] is a valid locahost port hosting a mongodb, and [PATH_TO_IMAGENET] is an optional path where Imagenet data are stored. If you don't pass an argument to --data_dir, the model will train from a default directory specified in psgnets/data/imagenet_data.py.