Environment

python 3.9 cuda 118 recommended
pip install -r requirements.txt

Train/Test/Demo

python VLMClassifier.py

Architecture

Results on Validation Data

Nr.	VLM	Method	Features	Acc_3↑	Acc_2↑	Acc_1↑
0	CLIP	ZeroShot	Image	6%	25%	70%
1	CLIP	ZeroShot	Image, Text	5%	32%	76%
2	CLIP	Adapter	Image	34%	72%	96%
3	CLIP	Adapter	Image, Text	37%	78%	98%
4	BLIP	ZeroShot	Image	5%	33%	80%
5	BLIP	ZeroShot	Image, Text	4%	28%	74%
6	BLIP	Adapter	Image	42%	76%	97%
7	BLIP	Adapter	Image, Text	43%	80%	98%

• Acc_3 (all 3 directions correct)
• Acc_2 (at least 2 directions correct)
• Acc_1 (at least 1 direction correct)

Manual Evaluation CLIP

Manual Evaluation BLIP

Demo "single image inference"