Muzhi Zhu

icon indicating a website link https://z-mu-z.github.io

icon company ZJU @aim-uofa icon location Hangzhou icon location PhD student in ZJU

Results 15 comments of Muzhi Zhu

Comparison with GroundedDINO

+1

The code for training.

Hello, our method is training-free, we rely on the pre-trained models SAM and DINOv2.

Bidirectional Matching

In forward matching, the dimension of the points in Pr is 2D. That is, each point has x and y coordinates representing its position in the image space and corresponds...

how to train the model and reproduce the results？

Hi, thanks for your interest. We are cleaning up the code and will release it as soon as it is ready.

微盘显示受理人已离职？

+1

💡 [REQUEST] Qwen-VL 在 4卡3090（24GB），可否实现LoRA 微调呢？

好像可以，我在4090上试过

NameError: name 'str2optimizer32bit' is not defined

python -m bitsandbytes XGPU-lite: L-229:Client configuration: use_uma:1, compute_schedule_mode:4, need_launch_kernel_admission:0, time_slice_spin_or_cv:1, enable_heart_beat:0, enable_monitor:0. XGPU-lite: L-163:func: cuInit, pid: 22366, tid: 22366, flags: 0 XGPU-lite: L-163:func: cuInit, pid: 22366, tid: 22366, flags: 0...

SegNext require many clicks when I run demo on DAVIS test image

Thanks for your reply. BTW, is this demo support box input?

input_features values are all -1.5 in modeling_qwen2_5_omni.py when using lora sft Qwen2.5Omni

@Kuangdd01 Thanks again! One follow-up question: Is adding the fake audio input strictly necessary? It seems like, based on the official inference code, it's also possible to not insert an...

input_features values are all -1.5 in modeling_qwen2_5_omni.py when using lora sft Qwen2.5Omni

Thanks for the explanation! I’ll check the outputs with and without fake inputs if I have time and share the results here.