Freeze Model Parameters During Training

  # lock image tower, i.e., Locked Image Tuning (LiT) https://arxiv.org/abs/2111.07991
--lock-image-model \

# lock all weight in image tower, while only train the text tower
--lock-image-partial 'weight' \

# only unlock all weight in image tower, while other params are locked
--lock-image-partial '!weight' --lock-image-model \

# Only train the first layer (transformer block) of the image backbone
--lock-image-partial '!resblocks.0'  --lock-image-model \

# Only unfreeze all bias and norm params, i.e., Bias and Normalization Optimization (BiNor) https://arxiv.org/abs/2203.07190
--lock-image-partial '!bias,!ln,!bn' --lock-text-partial '!bias,!ln' --lock-image-model  --lock-text-model \