---
license: apache-2.0
language: vi
library_name: transformers
tags:
- image-captioning
- knowledge-distillation
- blip
- fashion
---

# BLIP-base distilled (Attention Transfer + WSLD) — Fashion-200k captioning

Student model distill từ `Salesforce/blip-image-captioning-large` (Teacher) sang
`Salesforce/blip-image-captioning-base` đã prune decoder (12 → 4 tầng), train trên
Fashion-200k bằng combined loss:

- **Attention Transfer (AT)** ở vision encoder — so khớp attention map giữa Student/Teacher.
- **WSLD (Weighted Soft Label Distillation)** ở decoder — soft loss có trọng số theo token.
- Cross-entropy với ground-truth caption.

## Kiến trúc
- Tham số: 148.4M | Kích thước: 283.0MB
- Decoder giữ lại tầng: [0, 3, 6, 9]

## Kết quả trên test set
{
  "BLEU-4": 0.2458812827818811,
  "BLEU-1": 0.6297635645404903,
  "ROUGE-1": 0.5509984919361472,
  "ROUGE-2": 0.33081984865850167,
  "ROUGE-L": 0.4506778285021464,
  "METEOR": 0.40850410599132086,
  "CIDEr": 0.6211192626331922
}

## Cách dùng
```python
from huggingface_hub import hf_hub_download
from transformers import BlipForConditionalGeneration, BlipProcessor
import torch, json

repo_id = "demo-at-wsld-8/blip-at-wsld-fashion200k"
cfg = json.load(open(hf_hub_download(repo_id, "deploy_config.json"), encoding="utf-8"))
ckpt = hf_hub_download(repo_id, "student_at_wsld.pth")
# xem code dung lai kien truc (prune_decoder_layers, fix_decoder_layer_idx) trong app.py cua Space demo
```

## Code training đầy đủ
Xem notebook Kaggle: https://www.kaggle.com/code/anhvunguyenquynh/at-wsld