Update README.md
Browse files
README.md
CHANGED
|
@@ -2,91 +2,187 @@
|
|
| 2 |
license: mit
|
| 3 |
language:
|
| 4 |
- en
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
tags:
|
|
|
|
| 6 |
- vqa
|
| 7 |
- visual-question-answering
|
| 8 |
- knowledge-guided
|
|
|
|
|
|
|
|
|
|
| 9 |
- vit
|
| 10 |
- flan-t5
|
|
|
|
| 11 |
- ok-vqa
|
| 12 |
- multimodal
|
| 13 |
- pytorch
|
| 14 |
-
datasets:
|
| 15 |
-
- ok-vqa
|
| 16 |
-
pipeline_tag: visual-question-answering
|
| 17 |
-
library_name: pytorch
|
| 18 |
---
|
| 19 |
|
| 20 |
-
# Knowledge-Guided
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
-
|
| 23 |
|
| 24 |
-
|
| 25 |
|
| 26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
|
| 28 |
| Component | Model |
|
| 29 |
-
|
|
| 30 |
-
| Vision
|
| 31 |
-
|
|
| 32 |
-
|
|
| 33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
-
##
|
| 36 |
|
| 37 |
-
|
| 38 |
-
-
|
| 39 |
-
|
|
|
|
|
|
|
| 40 |
|
| 41 |
-
|
| 42 |
-
- **Strategy:** Late Fusion. Concatenates ViT CLS token with CLIP-encoded guidance vector, then projects to T5.
|
| 43 |
-
- **Knowledge Source:** LLM-generated visual focus + semantic knowledge via CLIP text encoder.
|
| 44 |
|
| 45 |
-
##
|
| 46 |
-
- **Strategy:** Attention Modification. Injects per-layer guidance projections into the top 6 ViT self-attention layers.
|
| 47 |
-
- **Knowledge Source:** LLM-generated visual focus + semantic knowledge via CLIP text encoder.
|
| 48 |
|
| 49 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
```python
|
| 52 |
import torch
|
| 53 |
-
from transformers import ViTModel, AutoTokenizer, AutoModelForSeq2SeqLM
|
| 54 |
|
|
|
|
|
|
|
| 55 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
|
| 57 |
-
|
| 58 |
-
t5_tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-small")
|
| 59 |
-
t5 = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-small").to(device)
|
| 60 |
-
vit = ViTModel.from_pretrained("google/vit-base-patch16-224").to(device)
|
| 61 |
|
| 62 |
-
#
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
|
| 68 |
-
|
| 69 |
|
| 70 |
-
|
| 71 |
-
- **Early Stopping:** Patience = 3-4 epochs
|
| 72 |
-
- **Max Epochs:** 15
|
| 73 |
-
- **Batch Size:** 8
|
| 74 |
-
- **Evaluation Metrics:** Exact Match, Edit Similarity, Token F1, BERTScore
|
| 75 |
|
| 76 |
-
|
| 77 |
|
| 78 |
-
|
| 79 |
|
| 80 |
```bibtex
|
| 81 |
-
@misc{
|
| 82 |
-
title={
|
| 83 |
-
author={Aneerban},
|
| 84 |
-
year={2026},
|
| 85 |
-
|
| 86 |
-
|
|
|
|
|
|
|
| 87 |
}
|
| 88 |
```
|
| 89 |
|
| 90 |
-
##
|
| 91 |
|
| 92 |
-
MIT
|
|
|
|
| 2 |
license: mit
|
| 3 |
language:
|
| 4 |
- en
|
| 5 |
+
base_model:
|
| 6 |
+
- google/vit-base-patch16-224
|
| 7 |
+
- openai/clip-vit-base-patch16
|
| 8 |
+
- google/flan-t5-small
|
| 9 |
+
datasets:
|
| 10 |
+
- ok-vqa
|
| 11 |
+
metrics:
|
| 12 |
+
- exact_match
|
| 13 |
+
- f1
|
| 14 |
+
- bertscore
|
| 15 |
+
pipeline_tag: visual-question-answering
|
| 16 |
+
library_name: pytorch
|
| 17 |
tags:
|
| 18 |
+
- arxiv:2610.17079
|
| 19 |
- vqa
|
| 20 |
- visual-question-answering
|
| 21 |
- knowledge-guided
|
| 22 |
+
- late-fusion
|
| 23 |
+
- attention-injection
|
| 24 |
+
- vision-transformer
|
| 25 |
- vit
|
| 26 |
- flan-t5
|
| 27 |
+
- clip
|
| 28 |
- ok-vqa
|
| 29 |
- multimodal
|
| 30 |
- pytorch
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
---
|
| 32 |
|
| 33 |
+
# Knowledge-Guided Visual Question Answering Models
|
| 34 |
+
|
| 35 |
+
This repository contains trained model checkpoints for **Knowledge-Guided Visual Question Answering (VQA)** on an augmented subset of the OK-VQA dataset.
|
| 36 |
+
|
| 37 |
+
๐ **Research paper:** [Rethinking LLM-Guided Visual Question Answering: Does Deeper Knowledge Integration Really Help?](https://arxiv.org/abs/2610.17079)
|
| 38 |
|
| 39 |
+
The paper investigates whether deeper integration of LLM-generated knowledge improves VQA performance under data-scarce fine-tuning conditions. We compare decision-level late fusion with attention-matrix injection using a shared ViT + FLAN-T5 framework.
|
| 40 |
|
| 41 |
+
## Research Highlights
|
| 42 |
|
| 43 |
+
- Compare shallow, decision-level late fusion against deep attention injection.
|
| 44 |
+
- Generate structured guidance with Llama 3.3 70B.
|
| 45 |
+
- Encode guidance using a CLIP text encoder.
|
| 46 |
+
- Expand the filtered training set from 3,144 to approximately 36,000 samples through question paraphrasing.
|
| 47 |
+
- Evaluate the baseline, late fusion, and attention injection using a consistent experimental setup.
|
| 48 |
+
|
| 49 |
+
The results show that **late fusion performs better than attention injection in this data-scarce experimental setting**. This does not imply that deeper integration is always inferior.
|
| 50 |
+
|
| 51 |
+
## Architecture
|
| 52 |
|
| 53 |
| Component | Model |
|
| 54 |
+
|---|---|
|
| 55 |
+
| Vision encoder | `google/vit-base-patch16-224` |
|
| 56 |
+
| Guidance text encoder | `openai/clip-vit-base-patch16` |
|
| 57 |
+
| Answer generation model | `google/flan-t5-small` |
|
| 58 |
+
| Benchmark | OK-VQA |
|
| 59 |
+
| LLM used to generate guidance | Llama 3.3 70B |
|
| 60 |
+
|
| 61 |
+
The pipeline generates structured guidance containing visual focus concepts, distractor elements to ignore, and external knowledge statements. A CLIP text encoder represents this guidance as an embedding used by the downstream VQA model.
|
| 62 |
+
|
| 63 |
+
### Integration strategies
|
| 64 |
+
|
| 65 |
+
**1. Baseline โ No Guidance**
|
| 66 |
+
|
| 67 |
+
Uses visual features and question representations without external LLM-generated guidance.
|
| 68 |
+
|
| 69 |
+
**2. Late Fusion โ Shallow Integration**
|
| 70 |
+
|
| 71 |
+
Combines the global ViT representation with the encoded guidance immediately before answer decoding. The visual feature extraction itself is left unchanged.
|
| 72 |
+
|
| 73 |
+
**3. Attention Injection โ Deep Integration**
|
| 74 |
+
|
| 75 |
+
Injects guidance into the query and key projections of the last six ViT layers, modifying visual attention.
|
| 76 |
|
| 77 |
+
## Available Checkpoints
|
| 78 |
|
| 79 |
+
| Checkpoint | Strategy | Approximate file size |
|
| 80 |
+
|---|---|---:|
|
| 81 |
+
| `best_baseline_model.pth` | No guidance | 309 MB |
|
| 82 |
+
| `best_fusion_model.pth` | Late fusion | 313 MB |
|
| 83 |
+
| `best_attention_model.pth` | Attention injection | 325 MB |
|
| 84 |
|
| 85 |
+
These files contain trained checkpoint weights. Loading them requires the matching model architecture and compatible pretrained components.
|
|
|
|
|
|
|
| 86 |
|
| 87 |
+
## Evaluation Results
|
|
|
|
|
|
|
| 88 |
|
| 89 |
+
The following scores are reported in the paper for the validation split of the augmented, top-500-filtered OK-VQA subset. **These are not official full-benchmark OK-VQA leaderboard scores.**
|
| 90 |
+
|
| 91 |
+
| Strategy | Exact Match (%) | Edit Similarity (%) | Token F1 (%) | BERTScore |
|
| 92 |
+
|---|---:|---:|---:|---:|
|
| 93 |
+
| Baseline โ No Guidance | 29.44 | 40.57 | 30.97 | 91.21 |
|
| 94 |
+
| Late Fusion | **39.44** | **48.97** | **41.34** | **92.14** |
|
| 95 |
+
| Attention Injection | 31.80 | 42.66 | 33.33 | 91.00 |
|
| 96 |
+
|
| 97 |
+
### Main finding
|
| 98 |
+
|
| 99 |
+
Late fusion achieves **39.44% exact match**, a 10.00-percentage-point improvement over the baseline. Attention injection achieves 31.80%, a 2.36-point improvement over the baseline.
|
| 100 |
+
|
| 101 |
+
These results suggest that, with limited fine-tuning data and largely frozen pretrained components, simpler knowledge integration can be more effective than modifying visual attention.
|
| 102 |
+
|
| 103 |
+
## Training Details
|
| 104 |
+
|
| 105 |
+
- **Dataset:** Filtered OK-VQA training data with the top 500 answer classes.
|
| 106 |
+
- **Data augmentation:** LLM-based question paraphrasing.
|
| 107 |
+
- **Original filtered training set:** 3,144 samples.
|
| 108 |
+
- **Augmented training set:** Approximately 36,000 samples.
|
| 109 |
+
- **Optimizer:** Adam.
|
| 110 |
+
- **Learning rate:** `3e-5`.
|
| 111 |
+
- **Batch size:** 8โ16.
|
| 112 |
+
- **Maximum epochs:** 15.
|
| 113 |
+
- **Early stopping:** Validation exact match, patience 3โ4.
|
| 114 |
+
- **Evaluation metrics:** Exact Match, Edit Similarity, Token F1, and BERTScore.
|
| 115 |
+
|
| 116 |
+
The ViT and CLIP encoders remain frozen except for the explicitly modified and trainable components of the attention-injection strategy.
|
| 117 |
+
|
| 118 |
+
## Usage
|
| 119 |
+
|
| 120 |
+
You can download a checkpoint from this repository with `huggingface_hub`:
|
| 121 |
+
|
| 122 |
+
```python
|
| 123 |
+
from huggingface_hub import hf_hub_download
|
| 124 |
+
|
| 125 |
+
checkpoint_path = hf_hub_download(
|
| 126 |
+
repo_id="rishii100/knowledge-guided-vqa-models",
|
| 127 |
+
filename="best_fusion_model.pth",
|
| 128 |
+
)
|
| 129 |
+
|
| 130 |
+
print("Checkpoint downloaded to:", checkpoint_path)
|
| 131 |
+
```
|
| 132 |
+
|
| 133 |
+
To run inference, instantiate the matching model class and load the checkpoint:
|
| 134 |
|
| 135 |
```python
|
| 136 |
import torch
|
|
|
|
| 137 |
|
| 138 |
+
# YourLateFusionModel is a placeholder for the exact model class
|
| 139 |
+
# used during training. It must match the checkpoint's architecture.
|
| 140 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 141 |
+
model = YourLateFusionModel().to(device)
|
| 142 |
+
|
| 143 |
+
state_dict = torch.load(
|
| 144 |
+
checkpoint_path,
|
| 145 |
+
map_location=device,
|
| 146 |
+
weights_only=True,
|
| 147 |
+
)
|
| 148 |
+
model.load_state_dict(state_dict)
|
| 149 |
+
model.eval()
|
| 150 |
+
```
|
| 151 |
|
| 152 |
+
**Important:** `YourLateFusionModel` is a placeholder, not an included implementation. The model class, forward pass, tokenizer configuration, and pretrained components must match the training implementation. The checkpoint alone is not a complete, standalone inference pipeline.
|
|
|
|
|
|
|
|
|
|
| 153 |
|
| 154 |
+
## Limitations
|
| 155 |
+
|
| 156 |
+
- Results are based on the experimental subset and validation protocol described in the paper.
|
| 157 |
+
- Performance may differ on the full official OK-VQA benchmark or under different training-data regimes.
|
| 158 |
+
- Attention injection can help some question categories while harming others.
|
| 159 |
+
- The checkpoint files do not, by themselves, provide the complete training pipeline or guarantee reproducibility without the corresponding implementation and configuration.
|
| 160 |
+
- The pretrained components retain their respective upstream licenses and terms.
|
| 161 |
+
|
| 162 |
+
## Paper and Citation
|
| 163 |
+
|
| 164 |
+
**Title:** Rethinking LLM-Guided Visual Question Answering: Does Deeper Knowledge Integration Really Help?
|
| 165 |
|
| 166 |
+
**Authors:** Aneerban Saha, Vijay Kumar, Sameer Madhav D., Priyam Srivastava, and Dhruv Bansal.
|
| 167 |
|
| 168 |
+
**Conference:** Accepted at ADCIS 2026.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
|
| 170 |
+
**arXiv:** [2610.17079](https://arxiv.org/abs/2610.17079)
|
| 171 |
|
| 172 |
+
### BibTeX
|
| 173 |
|
| 174 |
```bibtex
|
| 175 |
+
@misc{saha2026rethinkingvqa,
|
| 176 |
+
title = {Rethinking LLM-Guided Visual Question Answering: Does Deeper Knowledge Integration Really Help?},
|
| 177 |
+
author = {Saha, Aneerban and Kumar, Vijay and Madhav D., Sameer and Srivastava, Priyam and Bansal, Dhruv},
|
| 178 |
+
year = {2026},
|
| 179 |
+
eprint = {2610.17079},
|
| 180 |
+
archivePrefix = {arXiv},
|
| 181 |
+
primaryClass = {cs.CV},
|
| 182 |
+
url = {https://arxiv.org/abs/2610.17079}
|
| 183 |
}
|
| 184 |
```
|
| 185 |
|
| 186 |
+
## License
|
| 187 |
|
| 188 |
+
This repository is marked as MIT licensed.
|