Update README.md

README.md

@@ -6,12 +6,48 @@ tags:
 - contrastive learning
 ---
 
-FLAIR Model
+**FLAIR Model**
 
 Authors: [Rui Xiao](https://www.eml-munich.de/people/rui-xiao), [Sanghwan Kim](https://kim-sanghwan.github.io/), [Mariana-Iuliana Georgescu](https://lilygeorgescu.github.io/), [Zeynep Akata](https://www.eml-munich.de/people/zeynep-akata), [Stephan Alaniz](https://www.eml-munich.de/people/stephan-alaniz)
 
 FLAIR was introduced in the paper [FLAIR: VLM with Fine-grained Language-informed Image Representations](https://arxiv.org/abs/2412.03561). Based on ViT-B-16 Model from [OpenCLIP](https://github.com/mlfoundations/open_clip), FLAIR features text-conditioned attention pooling at the end of its vision transformer. Pre-trained on MLLM-recaptioned datasets from [DreamLIP](https://huggingface.co/datasets/qidouxiong619/dreamlip_long_captions), FALIR achieves strong performance in tasks such as zero-shot image-text retrieval and zero-shot segmentation.
 
+**Usage**
+We offer the detailed usage in our [Github repo](https://github.com/ExplainableML/flair). Example Usage:
+
+```python
+import flair
+from PIL import Image
+import torch
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+print(f"Using device: {device}")
+
+pretrained = flair.download_weights_from_hf(model_repo='xiaorui638/flair', filename='flair-cc3m-recap.pt')
+model, _, preprocess = flair.create_model_and_transforms('ViT-B-16-FLAIR', pretrained=pretrained)
+
+model.to(device)
+model.eval()
+
+tokenizer = flair.get_tokenizer('ViT-B-16-FLAIR')
+
+image = preprocess(Image.open("../assets/puppy.jpg")).unsqueeze(0).to(device)
+
+text = tokenizer(["In the image, a small white puppy with black ears and eyes is the main subject", # ground-truth caption
+                  "The white door behind the puppy is closed, and there's a window on the right side of the door", # ground-truth caption
+                  "A red ladybug is surrounded by green glass beads", # non-ground-truth caption
+                  "Dominating the scene is a white desk, positioned against a white brick wall"]).to(device) # non-ground-truth caption
+
+with torch.no_grad(), torch.cuda.amp.autocast():
+    flair_logits = model.get_logits(image=image, text=text)
+    clip_logits = model.get_logits_as_clip(image=image, text=text)
+
+    print("logits get using flair's way:", flair_logits) # [4.4062,  6.9531, -20.5000, -18.1719]
+    print("logits get using clip's way:", clip_logits) # [12.4609, 15.6797, -3.8535, -0.2281]
+```
+
+
+