Hello everyone,\nI wonder how pseudo labelling works as far as i know pseudo labelling is training model on dataset and then test it on test data after testing we labelled test data images with objects detected with certain threshold and then mix it up with train data and train second model (pls correct if i am wrong) if this is pseudo labelling then it should be depend upon size of test data.
", "messages": [ { "id": "921453", "postDate": "07/09/2020 09:51:49", "content": "Hello everyone,\nI wonder how pseudo labelling works as far as i know pseudo labelling is training model on dataset and then test it on test data after testing we labelled test data images with objects detected with certain threshold and then mix it up with train data and train second model (pls correct if i am wrong) if this is pseudo labelling then it should be depend upon size of test data.
", "rawMarkdown": "Hello everyone,\nI wonder how pseudo labelling works as far as i know pseudo labelling is training model on dataset and then test it on test data after testing we labelled test data images with objects detected with certain threshold and then mix it up with train data and train second model (pls correct if i am wrong) if this is pseudo labelling then it should be depend upon size of test data.", "votes": null }, { "id": "921567", "postDate": "07/09/2020 11:42:57", "content": "Yes, correct. When the test dataset is small, the pseudo labelling would not lead to significant improvement. Small is pretty fuzzy :) in this context but in meaning of having few examples compared with the structure complexity and training dataset size.
", "rawMarkdown": "Yes, correct. When the test dataset is small, the pseudo labelling would not lead to significant improvement. Small is pretty fuzzy :) in this context but in meaning of having few examples compared with the structure complexity and training dataset size.", "votes": null }, { "id": "921966", "postDate": "07/09/2020 17:30:49", "content": "Intuitively it does not make much sense to me. Like, you're learning from what you predicted, what difference does that really make right? I think mostly what it does is regularizes your model more than anything. Everyone's been able to increase their score using pseudo-labeling so I guess it does work even if the test set is relatively small.
", "rawMarkdown": "Intuitively it does not make much sense to me. Like, you're learning from what you predicted, what difference does that really make right? I think mostly what it does is regularizes your model more than anything. Everyone's been able to increase their score using pseudo-labeling so I guess it does work even if the test set is relatively small.", "votes": null }, { "id": "922244", "postDate": "07/10/2020 00:37:23", "content": "See this notebook for pseudo labeling - https://www.kaggle.com/nvnnghia/yolov5-pseudo-labeling
", "rawMarkdown": "See this notebook for pseudo labeling - https://www.kaggle.com/nvnnghia/yolov5-pseudo-labeling", "votes": null } ], "comments": [ { "id": 921567, "author_name": "waelkh", "author_url": "", "post_date": "07/09/2020 11:42:57", "content": "Yes, correct. When the test dataset is small, the pseudo labelling would not lead to significant improvement. Small is pretty fuzzy :) in this context but in meaning of having few examples compared with the structure complexity and training dataset size.
", "votes": null, "replies": [] }, { "id": 921966, "author_name": "jonykarki", "author_url": "", "post_date": "07/09/2020 17:30:49", "content": "Intuitively it does not make much sense to me. Like, you're learning from what you predicted, what difference does that really make right? I think mostly what it does is regularizes your model more than anything. Everyone's been able to increase their score using pseudo-labeling so I guess it does work even if the test set is relatively small.
", "votes": null, "replies": [] }, { "id": 922244, "author_name": "rashmibanthia", "author_url": "", "post_date": "07/10/2020 00:37:23", "content": "See this notebook for pseudo labeling - https://www.kaggle.com/nvnnghia/yolov5-pseudo-labeling
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "921453": "Hello everyone,\nI wonder how pseudo labelling works as far as i know pseudo labelling is training model on dataset and then test it on test data after testing we labelled test data images with objects detected with certain threshold and then mix it up with train data and train second model (pls correct if i am wrong) if this is pseudo labelling then it should be depend upon size of test data.", "921567": "Yes, correct. When the test dataset is small, the pseudo labelling would not lead to significant improvement. Small is pretty fuzzy :) in this context but in meaning of having few examples compared with the structure complexity and training dataset size.", "921966": "Intuitively it does not make much sense to me. Like, you're learning from what you predicted, what difference does that really make right? I think mostly what it does is regularizes your model more than anything. Everyone's been able to increase their score using pseudo-labeling so I guess it does work even if the test set is relatively small.", "922244": "See this notebook for pseudo labeling - https://www.kaggle.com/nvnnghia/yolov5-pseudo-labeling" }, "source": "meta" }