pytorch_models

Self-supervised Models

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:

• weakly_deepcontext
• weakly_learningbymoving
• weakly_objectcentric
• weakly_solvingpuzzle
• weakly_videoorder
• weakly_videotracking

These models have been verified to be equivalent in the forward pass but have discrepancies in the backwards pass (i.e., gradients):

• weakly_audio
• weakly_egomotion

TODO-models:

• weakly_colorization
• weakly_splitbrain

TODO-misc:

• define get_transform to provide the right data transformations for each task.

Model definition and weights were taken from http://netdissect.csail.mit.edu.