CarlCloudWang comments

Results 6 comments of


                                            CarlCloudWang

how use the multi-GPU inference ，can you give me some demo code

全tm伸手党😅

How Do I Perform Offline Inference? After a Model Is Downloaded to the Local Computer, the Model Cannot Be Executed?

https://huggingface.co/q-future/one-align/tree/main

About the number of input video frames

sample some frames per second... def load_video(video_file, skip=1): from decord import VideoReader vr = VideoReader(video_file) # Get video frame rate fps = vr.get_avg_fps() # Calculate frame indices for 1fps frame_indices...

When I try to use koniq-10 for training, I encounter the problem that the loss is always 0.

是不是用了预训练模型，那loss肯定小到不显示

RuntimeError: Error(s) in loading state_dict for MPLUGOwl2LlamaForCausalLM: size mismatch for norm.weight: copying a param with shape torch.Size([768]) from checkpoint, the shape in current model is torch.Size([4096]). You may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method.

你试试直接把模型拆开来自己写个__init__加载模型，从scorer.py看，实际上这模型就三个部分： tokenizer = AutoTokenizer.from_pretrained(model_path, model_path=model_path, use_fast=False, cache_dir=cache_dir) model = MPLUGOwl2LlamaForCausalLM.from_pretrained(model_path, model_path=model_path, local_files_only=True, cache_dir=cache_dir, low_cpu_mem_usage=True, device_map="auto") image_processor = CLIPImageProcessor.from_pretrained(model_path) 在Modeling_llama.py的结尾，作者很巧妙地直接把llama2的函数动态加载了，所以随便加载权重。

Why do we need expand2sqaure in scorers?

In my survey, I found that no expandsqure can promote the capicity of whole model... u can only use preprocessing from CLIP_ImageProcessor, and this one can cause real improvement.