woshidandan / image-aesthetics-and-quality-assessment Goto Github PK

这边按照默认option的参数从头训练了20个epcoh，lcc只有0.10多，是需要载入预训练权重吗？backbone是你们自己设计的，你们应该也是直接在AVA上训练的吧？

感谢您的杰出工作！
想请问您，论文中提到的多个数据集的train\test\val split 怎样获取？

非常感谢！

您好，我在做一个基于clip搜索生成图片的搜索引擎，发现有很多图片相关性很高但是质量不高，为了数据冷启动，想加一个质量因子。

我想用 Image-Aesthetics-Assessment ， 但是发现 get_score_one_image函数定义是空的，怎么基于训练好的模型推理单张图片，能给一个完整的演示代码吗？（我对写网站比较熟悉，转过一些模型到ONNX，但是对AI开发不太熟悉）

Thank you for your excellent work. Would you mind providing the inference code?

Hi, thanks for opening source your nice work. When will the paper be released?

thanks for your contribution.
I'm meeting a problem in AVA dataset. When I read the pre-trained weights and validate on the AVA dataset, the resulting metrics are incorrectly high.

The code to read the weights is shown below, I made some modifications since it doesn't run directly on my environment.

`def start_train(opt):
train_loader, val_loader, test_loader = create_data_part(opt)
args, config = parse_option()
print(f"Creating model:{config.MODEL.TYPE}/{config.MODEL.NAME}")
model = build_model(config)

if os.path.exists(config.MODEL.RESUME):
    print(config.MODEL.RESUME)
    checkpoint = torch.load(config.MODEL.RESUME)
    pre_weights = torch.load(config.MODEL.RESUME)
    # pre_weights = checkpoint['model']
    pre_dict = {}
    for k, v in pre_weights.items():
        if "cls_head" not in k:
            pre_dict[k] = v
    model.load_state_dict(pre_dict, strict=False)
    #直接读取
    model.load_state_dict(torch.load(opt['path_to_model_weight'], map_location='cuda:0'))
model.to('cuda')`

I want to know if you have also encountered this situation. I am a beginner and the modified code may have some errors. I hope to get some suggestions from you.

请问图片是应该直接resize到224x224，还是应该保持长宽比不变然后剪裁一下

图中aesthetic_score是laion美学分

感谢你的工作！请问，模型指标上贴出了640的效果要远远好于224的，但是你提供的权重文件中没有发现640的权重文件，配置文件也不含有640的配置文件，能否提供下？谢谢！

dat_base_in1k_224.pth

如题