Is self.image_feature_neck trainable ?

[qzzzxiaosheng]

In the code, self.image_feature_neck seems to be frozen. It is strange for that.

[qzzzxiaosheng]

trainable_list = ["lm_head", "embed_tokens", "mask_decoder", "text_hidden_fcs"]
if args.resize_vision_tower_size != 224:
    trainable_list.append('mm_projector')
    # trainable_list.append('out_mm_projector')
for n, p in model.named_parameters():
    if any(
        [
            x in n
            for x in trainable_list
        ]
    ):
        print("n: ", n, "p.shape: ", p.shape)
        p.requires_grad = True

I think trainable_list should contain the image_feature_neck and prompt encoder, otherwise they are not trainable.