README.md
April 6, 2025 · View on GitHub
Performance
A. MLLMs Evaluation Results
To evaluate MLCD’s performance within multimodal large language models (MLLMs), we replaced the CLIP model in LLaVA-NeXT with the MLCD model. We paired this with the Qwen2.5-7B language model. For reproducibility, we utilized the LLaVA-Pretrain dataset for pre-training and the LLaVA-NeXT-Data for structured fine-tuning. The evaluation results confirm that the MLCD model performs exceptionally well across multiple benchmarks, underscoring its effectiveness in MLLMs.
| Vision Tower | RoPE2D | ChartQA | DocVQA | InfoVQA | OCRBench | MMMU |
|---|---|---|---|---|---|---|
| CLIP (ViT-L-14-336px) | × | 66.52 | 75.21 | 38.88 | 525.00 | 44.20 |
| SigLIP (ViT-SO400M-384px) | × | 69.28 | 76.71 | 41.38 | 554.00 | 46.78 |
| DFN5B (ViT-H-14-378px) | × | 64.36 | 70.87 | 38.59 | 473.00 | 48.00 |
| HF:MLCD (ViT-L-14-336px) | × | 67.84 | 76.46 | 43.48 | 531.00 | 44.30 |
| HF:MLCD (ViT-bigG-14-336px) | √ | 71.07 | 79.63 | 44.38 | 572.00 | 46.78 |
| HF:MLCD (ViT-bigG-14-448px) | √ | 73.80 | 83.34 | 46.59 | 582.00 | 46.00 |
| Vision Tower | MLCD (ViT_L_14_336px) | CLIP (ViT_L_14_336px) |
|---|---|---|
| LLM | Qwen2.5-7B | Qwen2.5-7B |
| AI2D | 76.98 | 73.15 |
| GQA | 64.17 | 63.31 |
| ScienceQA-Img | 78.09 | 76.35 |
| InfoVQA-Val | 43.48 | 38.88 |
| MMBenchCN-Dev | 74.83 | 72.51 |
| MMBenchEN-Dev | 76.37 | 74.57 |
| SeedBench | 68.20 | 66.80 |
| SeedBench-Img | 73.75 | 72.72 |
| MMStar | 50.98 | 48.98 |
| MMMU | 44.30 | 44.20 |
| POPE | 88.69 | 88.83 |
| ChartQA | 67.84 | 66.52 |
| DocVQA-Val | 76.46 | 75.21 |
| TextVQA-Val | 61.69 | 62.47 |
| OCRBench | 531 | 525 |
| MME(cognition) | 432 | 384 |
| MME(perception) | 1598 | 1512 |
B. Linear Probe Evaluation Results
This table presents the results of linear probe evaluations comparing CLIP and MLCD models on the ViT_L_14_336px architecture across various datasets. The linear probe test freezes the pre-trained model's weights and trains a linear classifier on top to assess how well the model's representations generalize to different tasks.
The results of the ImageNet linear probe are as follows:
| Model Name | ImageNet Linear Probe | Hugging Face |
|---|---|---|
| MLCD-ViT-B-32-224px | 79.1 | HF:MLCD-ViT-B-32-224px |
| MLCD-ViT-L-14-336px | 86.3 | HF:MLCD-ViT-L-14-336px |
| MLCD-ViT-bigG-14-224px | 87.1 | HF:MLCD-ViT-bigG-14-224px |
| Dataset | MLCD (ViT_L_14_336px) | CLIP (ViT_L_14_336px) |
|---|---|---|
| Food101 | 96.21 | 95.90 |
| CIFAR-10 | 99.36 | 97.90 |
| CIFAR-100 | 93.69 | 87.40 |
| Birdsnap | 88.18 | 79.90 |
| SUN397 | 87.96 | 82.20 |
| Stanford Cars | 95.16 | 91.50 |
| FGVC Aircraft | 86.38 | 71.60 |
| Describable Textures Dataset | 86.70 | 83.00 |
| Oxford-IIIT Pets | 96.27 | 95.10 |
| Caltech-101 | 97.92 | 96.00 |
| Flowers102 | 99.58 | 99.20 |
| ImageNet | 86.10 | 85.40 |
convert pytorch2huggingface
python convert_vit_bigG_14_rope2d_to_hf.py \
--pytorch_dump_folder_path mlcd-vit-bigG-patch14-336 \
--checkpoint_path MLCD_ViT_bigG_14_336px_pytorch.pt \
--image_size 336