X-PLUG / mPLUG-Owl

mPLUG-Owl: The Powerful Multi-modal Large Language Model Family
https://www.modelscope.cn/studios/damo/mPLUG-Owl
MIT License
2.25k stars 171 forks source link

Inference result is different from the demo. #133

Open ssiq opened 1 year ago

ssiq commented 1 year ago

I downloaded the checkpoint from huggingface https://huggingface.co/MAGAer13/mplug-owl-llama-7b and https://huggingface.co/MAGAer13/mplug-owl-llama-7b-ft. With the following generate config, I fail to reproduce the demo result. generate_kwargs = dict( temperature=1, top_p=0.9, top_k=3, num_beams=1, no_repeat_ngram_size=2, do_sample=False, early_stopping=True, length_penalty=1, max_length=512 ) The input is the moday image in the examples directory. The result of demo is : The meme is funny because it features a cute French bulldog lying on the floor with the caption "Monday. Just...Mondoy." The dog's expression and body language convey a sense of exhaustion and resignation, which is often associated with Mondays. The phrase "just...monday" is a play on words, as it implies that Monday is just another day, but it can also be interpreted as a sarcastic comment on how the day feels. This mildly humorous and relatable image is likely to resonate with many people who have experienced the same feelings on a Monday morning.

The result of mplug-owl-llama-7b is: The meme is funny because it depicts a tired French bulldog lying on the floor with the caption "Monday. Just...Mondoy." The dog's expression and posture convey a sense of exhaustion and resignation, which is often associated with Mondays. The use of the word "mondoy" instead of "money" or "monkey" adds a humorous twist to the mood, making it more relatable and entertaining for the viewer.

The result of mplug-owl-llama-7b-ft is: The meme is funny because it features a sleepy French bulldog lying on the floor with the caption "Monday. Just...M-O-N-D-A-Y." The dog's expression and the wordplay on "monday" create a lighthearted and relatable moment for people who have experienced the feeling of dread that often accompanies the start of a new workweek. The image captures the essence of the common sentiment that Mondays can be a bit of an uphill battle, and it'll be just another day soon enough.

What is the reason for my failure on reproducing the demo result?

=================================update=====================================

I downlaoded a new checkpoint from the demo space https://huggingface.co/spaces/MAGAer13/mPLUG-Owl/tree/main, and found it is a vedio version checkpoint, but using the above input, its result is also different from the demo output.