Unsupervised deep convolutional neural network model for the ventral visual stream.
We use python=3.7.4.
tensorflow_gpu==1.15.0, tensorflow whose version is bigger than or equal to 1.9.0 and smaller than 2.0.0 should suffice for most of our network training. However, SimCLR training requires tensorflow==1.15.0.
The reproduction is done in three steps: preparing datasets needed for training, training networks, evaluating the trained networks using neural data.
To prepare datasets, see scripts in prepare_datasets.
To run the following two steps, you need to first install this package (like pip install -e ./).
For neural network training, see scripts in unsup_vvs/network_training folder.
For neural data evaluation, see scripts in unsup_vvs/neural_fit folder.
For using pretrained checkpoints to generate responses on images, see README in unsup_vvs/neural_fit folder.