Update README.md

README.md

@@ -13,13 +13,11 @@ datasets:
 - genecis
 pipeline_tag: feature-extraction
 library_name: pytorch
-base_model:
-- openai/clip-vit-large-patch14
 ---
 
 # Pretrain like Your Inference: Masked Tuning Improves Zero-Shot Composed Image Retrieval
 
-This repository contains the official pre-trained and tuned model weights for **PLI (Pretrain like Your Inference)**, accepted at **ICME 2025**.
+This repository contains the official pre-trained and tuned model weights (CLIP-ViT-L/14 backbone) for **PLI (Pretrain like Your Inference)**, accepted at **ICME 2025**.
 
 [![GitHub](https://img.shields.io/badge/GitHub-Code-blue?logo=github)](https://github.com/Chen-Junyang-cn/PLI)
 [![arXiv](https://img.shields.io/badge/arXiv-2311.07622-b31b1b.svg?logo=arxiv)](https://arxiv.org/abs/2311.07622)
@@ -55,15 +53,14 @@ import clip
 from huggingface_hub import hf_hub_download
 
 # 1. Download the weights from Hugging Face
-# Replace 'Chen-Junyang/PLI' with your actual Hugging Face repo ID
 checkpoint_path = hf_hub_download(
-    repo_id="Chen-Junyang/PLI", 
-    filename="pli_clip_resnet50.pt"  # Your weight filename
+    repo_id="jayong/PLI-CLIP-VIT-L-14", 
+    filename="best.pth"
 )
 
-# 2. Initialize the base CLIP model (e.g., RN50 or ViT-B/32)
+# 2. Initialize the base CLIP model (ViT-L/14)
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model, preprocess = clip.load("RN50", device=device)
+model, preprocess = clip.load("ViT-L/14", device=device)
 
 # 3. Load the tuned weights
 state_dict = torch.load(checkpoint_path, map_location=device)
@@ -74,32 +71,6 @@ print("PLI model weights loaded successfully!")
 
 ---
 
-## 📊 Evaluation Results
-
-PLI achieves state-of-the-art performance on four major ZS-CIR benchmark datasets:
-
-### 1. FashionIQ
-Evaluated on the validation set of FashionIQ.
-
-| Category | R@10 | R@50 |
-| :--- | :---: | :---: |
-| Dress | - | - |
-| Shirt | - | - |
-| Toptee | - | - |
-| **Average** | **-** | **-** |
-
-### 2. CIRR
-Evaluated on the validation set of CIRR.
-
-| Method | R@1 | R@5 | R@10 | R@50 |
-| :--- | :---: | :---: | :---: | :---: |
-| CLIP (RN50) | - | - | - | - |
-| **PLI (Ours)** | **-** | **-** | **-** | **-** |
-
-*Note: Please update the tables above with the exact values from your paper (Table 1, Table 2, etc.).*
-
----
-
 ## 🛠️ Method Overview
 
 ```
@@ -121,15 +92,17 @@ Evaluated on the validation set of CIRR.
 If you find our work or weights useful in your research, please consider citing our paper:
 
 ```bibtex
-@article{chen2023pretrain,
-  title={Pretrain like Your Inference: Masked Tuning Improves Zero-Shot Composed Image Retrieval},
+@inproceedings{chen2025pretrain,
+  title={Pretrain like your inference: Masked tuning improves zero-shot composed image retrieval},
   author={Chen, Junyang and Lai, Hanjiang},
-  journal={arXiv preprint arXiv:2311.07622},
-  year={2023}
+  booktitle={2025 IEEE International Conference on Multimedia and Expo (ICME)},
+  pages={1--6},
+  year={2025},
+  organization={IEEE}
 }
 ```
 
 ---
 
 ## 📭 Contact / Feedback
-For questions or feedback, please raise an issue on our [GitHub Repository](https://github.com/Chen-Junyang-cn/PLI) or contact `Junyang Chen`.
+For questions or feedback, please raise an issue on our [GitHub Repository](https://github.com/Chen-Junyang-cn/PLI).