Closed alexrey88 closed 4 years ago
The idea is to implement simCLR in the code: train the base encoder with simCLR, and then fine-tune the weights of the encoder with our main task (the masker). See if we get better results than with current deeplabv2 encoder.
Reminder : add art option for pretraining, and move simclr loss computation to another function get_pretraining_loss
The idea is to implement simCLR in the code: train the base encoder with simCLR, and then fine-tune the weights of the encoder with our main task (the masker). See if we get better results than with current deeplabv2 encoder.