Bug in the implementation when non_analyze=False

[FabienRoger]

In watermarking/watermark_processors/message_model_processor.py
decoded_message, decoded_prob = decoded_probs.max(0) should be decoded_prob, decoded_message = decoded_probs.max(0)

[leanwang326]

Sorry for this. We have fixed this as you suggest.

[zzq73]

Thanks for your work, I want to know what is the form of watermark message input into the model.

[leanwang326]

watermark_processor = WmProcessorMessageModel(message_model=lm_message_model,
tokenizer=tokenizer,
encode_ratio=args.encode_ratio,
max_confidence_lbd=args.max_confidence_lbd,
strategy=args.message_model_strategy,
message=args.message,
top_k=args.top_k,
)
here, in this sentence: message=args.message (w.l.o.g, we assume message is an integer. other kind of information can convert to 0-1 bits and then to an integer)

[zzq73]

谢谢您的解答，但是抱歉，我没太懂0-1比特的含义，是说嵌入的每一个水印message最多包含1比特的信息量吗？比如可以嵌入是否是由模型生成的文本，这种包含是否两个选项的水印，也可以嵌入一个确定的没有选项的信息，比如嵌入时间等，是这样理解吗？请问如果信息量超过1比特，可以怎样处理呢

[leanwang326]

for example, we can encode an integer between 0-1023, then it corresponds to an information of 10 bit

[zzq73]

Thank you for your answer