In models/self_supervised.py, we provide a stand-alone file that provides model definitions for self-supervised models converted from Caffe to PyTorch.
model = models.self_supervised.alexnet(task_name='weakly_solvingpuzzle', pretrained=True)
The following models have been converted and fully verified to be equivalent:
These models have been verified to be equivalent in the forward pass but have discrepancies in the backwards pass (i.e., gradients):
get_transformto provide the right data transformations for each task.
Model definition and weights were taken from http://netdissect.csail.mit.edu.