感谢您的工作！有一些困惑请教！

[rtfgithub]

在代码中，第一阶段的训练中image encoder是冻结的，可学习的text tokens和和text encoder是可学习的。这和论文里描述的只有text tokens是可学习的，image encoder和text encoder是冻结的不匹配呀。

[DRACOyu]

hello，stage1，text encoder is fix，
def make_optimizer_1stage(cfg, model):
params = []
keys = []
for key, value in model.named_parameters():
if "prompt_learner" in key:
lr = cfg.SOLVER.STAGE1.BASE_LR
weight_decay = cfg.SOLVER.STAGE1.WEIGHT_DECAY
params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}]
keys += [key]
if cfg.SOLVER.STAGE1.OPTIMIZER_NAME == 'SGD':
optimizer = getattr(torch.optim, cfg.SOLVER.STAGE1.OPTIMIZER_NAME)(params, momentum=cfg.SOLVER.STAGE1.MOMENTUM)
elif cfg.SOLVER.STAGE1.OPTIMIZER_NAME == 'AdamW':
optimizer = torch.optim.AdamW(params, lr=cfg.SOLVER.STAGE1.BASE_LR, weight_decay=cfg.SOLVER.STAGE1.WEIGHT_DECAY)
else:
optimizer = getattr(torch.optim, cfg.SOLVER.STAGE1.OPTIMIZER_NAME)(params)
return optimizer

[cyberblue1]

updated image encoder and text encoder do not need to match , it only takes updated image encoder to test.