Bristol-Myers-Squibb-Molecular-Translation

This respository contains my code for competition in kaggle.

7th Place Solution for Bristol-Myers Squibb – Molecular Translation

Team: Mr_KnowNothing, Shivam Gupta, Phaedrus, Nischay Dhankhar, atfujita

• All models(Team)
  Public LB: 0.60(7th)
  Private LB: 0.60(7th)

The full picture of our solution is here

Note: This repository contains only my models and only train script.

• My models(3 Models averaging)
  Public LB: 0.66
  Private LB: 0.66

Note: This repogitory is based on hengck23's great assets. Please check here for details

My Models

Vit based model

• Encoder: vit_deit_base_distilled_patch16_384
• Decoder: TransformerDecoder
• Loss: LabelSmoothingLoss
• Augumentation: RandomScale, Cutout

There are 2 Vit based models.
The second was re-training by strengthening Noize Injection and Augmentation.

Vit model1

• Public LB: 0.77 (With Normalize)
• Private LB: 0.78 (With Normalize)

Vit model2

• Public LB: 0.76 (With Normalize)
• Private LB: 0.77 (With Normalize)

Swin Transformer based model

• Encoder: swin_base_patch4_window12_384_in22k
• Decoder: TransformerDecoder
• Loss: LabelSmoothingLoss
• Augumentation: RandomScale, Cutout

Swin model

• Public LB: 0.91 (With Normalize)
• Private LB: 0.92 (With Normalize)