ETRI_Fashion-How_Season-5_Task3

This is the repository for ETRI_Fashion-How Season 5 task 3 by VCL.

1. Goal

The goal of this task is to develop a model that ranks three outfit candidates based on user conversations. In this process, the model sequentially learns a total of six tasks following an incremental learning setting, where it is crucial to retain knowledge from previous tasks without forgetting.

Keyword: Task-Free Incremental Learning

2. Conditions

• Augmentation : Yes
• Ensemble : No
• Memory : No
• Multimodal : Yes
• External Data : No
• External Model : Yes
• Minimum model performance: 0.7
• Use of previous model features: No

3. Metric

WKT(Weighted Kenall's Tau)

example

Real rank	Pred. rank	WKT
2,1,0	2,1,0	1.0000
2,1,0	2,0,1	0.5455
2,1,0	1,2,0	0.1818
2,1,0	1,0,2	-0.3636
2,1,0	0,2,1	-0.3636
2,1,0	0,1,2	-1.0000

4. Model structure

5. Setup

• Python : 3.8.19
• CUDA : 11.3
• Library : requirements.txt

Training with validation

bash run_example.sh

Validation Scores

Task	Task1_val	Task2_val	Task3_val	Task4_val	Task5_val	Task6_val	Mean
1	0.8348						0.8348
2	0.7455	0.7632					0.7544
3	0.8064	0.6097	0.7768				0.7310
4	0.8210	0.6912	0.5578	0.8133			0.7208
5	0.7809	0.6996	0.4921	0.7255	0.7434		0.6883
6	0.8131	0.6377	0.5774	0.7299	0.6906	0.7261	0.6958

Final Score

(after quantization)

0.7063361

Server: 16.8MB

My PC: 17.6 MB