Peijie Wang comments

Results 9 comments of


                                            Peijie Wang

acc does't improve while training 100 epochs

I haven't solve this problem

acc does't improve while training 100 epochs

problem solved， set learn rate to 1e-5， then it will be ok

acc does't improve while training 100 epochs

lr=1e-5，and in scheduler.py set max_lr: float = 1e-5, other configs remain unchanged

acc does't improve while training 100 epochs

my best accuracy was 73.6% on base version

> 在load model之前加上这段 llava_model_args 应该可以避免这个warning > > ```python > llava_model_args = { > "multimodal": True, > "attn_implementation": "sdpa", > } > tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map,...

RuntimeError: probability tensor contains either `inf`, `nan` or element < 0

when inference on A6000*4， it occurs error above；while inference on H800 it is OK. And I fing when I run llava-onevision-7B on multipile gpus， it also produce error above， but...

multi-image input

Does taskweaver support multi-image input?

Peijie Wang

acc does't improve while training 100 epochs

acc does't improve while training 100 epochs

acc does't improve while training 100 epochs

acc does't improve while training 100 epochs

vision tower参数问题

vision tower参数问题

最新的2d5模型使用huggingface上的示例下载（model = AutoModel.from_pretrained('internlm/internlm-xcomposer2d5-7b', torch_dtype=torch.bfloat16, trust_remote_code=True).cuda().eval()），不会下载build_mlp.py文件，有没有遇到这个情况的

RuntimeError: probability tensor contains either `inf`, `nan` or element < 0

multi-image input