Apparatus and method for training image classification model
An apparatus for training an image classification model according to an embodiment disclosed includes a first trainer that trains a model body and a first head through supervised learning based on a labeled data set subjected to type 1 labeling, a second trainer that trains the model body, the first head, and a second head through multi-task learning based on the labeled data set and an unlabeled data set, and a third trainer that trains a plurality of third heads through supervised learning based on the labeled data set subjected to type 2 labeling while freezing the model body.
1. A method for training an image classification model performed by a computing device including one or more processors and a memory for storing one or more programs executed by the one or more processors, the method comprising:
a first training step of training a model body and a first head through supervised learning based on a labeled data set subjected to type 1 labeling;
a second training step of training the model body, the first head, and a second head through multi-task learning based on the labeled data set and an unlabeled data set; and
a third training step of training a plurality of third heads through supervised learning based on the labeled data set subjected to type 2 labeling while freezing the model body,
wherein the model body extracts feature vector for input data, and each of the first head, the second head, and the third head generates a classification result based on the feature vector.
2. The method of claim 1 , wherein, in the type 1 labeling,
when image data included in the labeled data set has a single label, the image data having the single label is classified into a class corresponding to the single label; and
when the image data included in the labeled data set has multiple labels, the image data having the multiple labels is classified into a new class corresponding to the multiple labels themselves.
3. The method of claim 1 , wherein, in the second training step, the model body and the first head are trained through supervised learning again based on the labeled data set, and the model body and the second head are trained through self-supervised learning.
4. The method of claim 3 , wherein
in the second training step, the model body and the first head are trained through supervised learning again by reflecting changes to one or more hyper parameters obtained from a user.
5. The method of claim 3 , wherein, in the second training step, the model body and the second head are trained through self-supervised learning based on derived data transformed from image data included in the labeled data set.
6. The method of claim 1 , wherein, in the second training step, the model body and the first head are trained through unsupervised learning based on the unlabeled data set.
7. The method of claim 1 , wherein, in the type 2 labeling, each image data included in the labeled data set is subjected to binary classification according to whether or not each image data included in the labeled data set has a label corresponding to a specific class.
8. The method of claim 1 , wherein, in the third training step, based on the labeled data set classified into a plurality of classes according to the type 2 labeling, each of the plurality of third heads corresponding to each of the plurality of classes is trained through supervised learning.
9. An apparatus for training an image classification model comprising:
a first trainer that trains a model body and a first head through supervised learning based on a labeled data set subjected to type 1 labeling;
a second trainer that trains the model body, the first head, and a second head through multi-task learning based on the labeled data set and an unlabeled data set; and
a third trainer that trains a plurality of third heads through supervised learning based on the labeled data set subjected to type 2 labeling while freezing the model body,
wherein the model body extracts feature vector for input data, and each of the first head, the second head, and the third head generates a classification result based on the feature vector.
10. The apparatus of claim 9 , wherein, in the type 1 labeling,
when image data included in the labeled data set has a single label, the image data having the single label is classified into a class corresponding to the single label; and
when the image data included in the labeled data set has multiple labels, the image data having the multiple labels is classified into a new class corresponding to the multiple labels themselves.
11. The apparatus of claim 9 , wherein the second trainer trains the model body and the first head through supervised learning again based on the labeled data set, and trains the model body and the second head through self-supervised learning.
12. The apparatus of claim 11 , wherein the second trainer trains the model body and the first head through supervised learning again by reflecting changes to one or more hyper parameters obtained from a user.
13. The apparatus of claim 11 , wherein the second trainer trains the model body and the second head through self-supervised learning based on derived data transformed from image data included in the labeled data set.
14. The apparatus of claim 9 , wherein the second trainer trains the model body and the first head through unsupervised learning based on the unlabeled data set.
15. The apparatus of claim 9 , wherein, in the type 2 labeling, each image data included in the labeled data set is subjected to binary classification according to whether or not each image data included in the labeled data set has a label corresponding to a specific class.
16. The apparatus of claim 9 , wherein the third trainer, based on the labeled data set classified into a plurality of classes according to the type 2 labeling, trains each of the plurality of third heads corresponding to each of the plurality of classes through supervised learning.