docta-ai / docta Goto Github PK

Hi,

I saw there are different setups forsample_size for hoc_cfg and detect_cfg in different experiments.
Some experiments use the same values: 35000 for both for hoc_cfg and detect_cfg in docta_cifar10.py .
Other use different values: 15000 for hoc_cfg and 350000 detect_cfg in docta_rare_pattern_cifar10.ipynb.

I tried running docta_cifar10.py with different sample_size values and I get different results with every different setup.

I do not quite understand how sample_size influences the results and how should I select the sample_size for a dataset.
Thank you.

Hi.
Thank you for this interesting package.

I was looking at the examples in the tools, more precisely at docta_rare_pattern.py and using it with the following command:
%run ./tools/docta_rare_pattern.py --feature_type 'embedding' --suffix 'c1m_subset' , following the demo in the notebook docta_rare_pattern_clothes.ipynb.

In this case the dataset is created using:
dataset = Customize_Image_Folder(root=cfg.data_root, transform=None).

Looking at the Customize_Image_Folder(), given that transform =None the image will be transformed using the following function:

if transform is None:
            self.transform = transforms.Compose([
                transforms.Resize(256),
                transforms.CenterCrop(224),
                transforms.ToTensor(),
                transforms.Normalize((0.6959, 0.6537, 0.6371), (0.3113, 0.3192, 0.3214)),
            ])

and followed up by:

feature_all.append(tmp_feature.permute(1, 2, 0).numpy().astype(np.uint8))  

But then in the docta_rare_pattern.py file the pre_processor.encode_feature() is used to get the image embeddings but also customized the data (performs another normalization step before extracting the embeddings) with:
dataset_list += [CustomizedDataset(feature=self.dataset.feature, label=self.dataset.label, preprocess=preprocess)], where the preprocess function is retrieved from get_encoder():
model_embedding, _, preprocess = open_clip.create_model_and_transforms(self.cfg.embedding_model)

The CLIP preprocessing function seems to be this one:

Compose(
    Resize(size=224, interpolation=bicubic, max_size=None, antialias=warn)
    CenterCrop(size=(224, 224))
    <function _convert_to_rgb at 0x7ff56ae71f30>
    ToTensor()
    Normalize(mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711])
)

I am wondering why the images are normalized two times: one within the Customized_Image_Folder() and a second time with the preprocessing function from open_clip.
Are there any steps that I am missing? I cannot fully understand why the double normalization & preprocessing.
Could you please explain why this strategy is used in the method?

I am greatly thankful for any hints and insights that you can provide.
Thank you for your time.

Later edit: I also noticed this in docta_cifar10.py where the data is loaded with Cifar10_noisy(cfg, train=True) which has a transform function defined and then the pre_process.encode_feature() is called which customized the data with another preprocessing step.

I apologize for the long post.

你好，感谢你的工作。我在实际使用的时候对一些代码中没有解释参数有些疑惑，比如already_2nn以及num_rounds
如果我想对40W+的43分类数据进行docta，你有建议的参数设置吗？

hoc_cfg = dict(
max_step = 1501,
T0 = None,
p0 = None,
lr = 0.1,
num_rounds = 50,
sample_size = 280000,
already_2nn = False,
device = 'cuda'
)

detect_cfg = dict(
num_epoch = 51,
sample_size = 280000,
k = 43,
name = 'simifeat',
method = 'rank'
)