Can I apply Custom Dataset?

[LJKwang]

Hello, I'm trying to learn by applying a custom dataset with unsupervised learning without labeling, is it possible? If It's possible, could you tell me how?

[jhgan00]

Hello, IRT basically requires a labeled dataset. It would be more appropriate to refer to MAE or SimCLR(MAE trains ViT in a self-supervised manner). These methods may require a relatively large scale of data and large batch sizes.

• SimCLR

• MAE Pre-training

[LJKwang]

I have an additional question!

1. I'd like to see multiple renking lists like the test code here, can I apply the same for MAE or simCLR?

2. SOP DATASET doesn't have a label but only has a class id, so isn't this a non-supervised learning?

Thank you for reading and responding!

[jhgan00]

1. Yes, you can extract compressed representations of images using trained models and retrieve similar images in the embedding space based on similarity measures like cosine similarity.

2. Could you please clarify the term 'label'? The class IDs of the SoP dataset are a type of labels. https://paperswithcode.com/dataset/stanford-online-products.

[LJKwang]

I'm sorry to keep asking you. It's hard because it's my first time applying artificial intelligence.

args = torch.load("./outputs/cub200/args.pth") state_dict = torch.load("./outputs/sop/encoder.pth", map_location="cpu")

Of the two codes above, I found that "encoder.pth" is the result of learning. But, I used "args.pth" that the author posted. When customizing, won't it have a big impact?

[jhgan00]

The settings defined in args are primarily parameters that only affect the training phase. If you are only running inference using a trained model, there should be no issues.

However, if you intend to continue training the model, you might need to adjust the hyperparameters that are defined in args to achieve better performance.

[LJKwang]

Thank you very much!

[jhgan00]

Please feel free to leave a comment if you need more support