Update README.md

README.md

@@ -16,14 +16,17 @@ tags:
 - text-to-video-search
 - multimodal-embedding
 ---
-# ![](assets/tara-logo.png) TARA: *T*ext *A*dapted *R*etrieval *A*lignment for Nuanced Video Retrieval
+
+# ![](assets/tara-logo.png) TARA: *T*ext *A*dapted *R*etrieval *A*lignment for Nuanced Video Retrieval (ECCV 2026)
 <!-- # <img src="./assets/logo.png" width="24"> TARA: Time-Aware Retrieval Adaptation for Video Understanding -->
 
 This repository contains inference and evaluation code for the TARA model based on the paper:
-[Adapting MLLMs for Nuanced Video Retrieval](https://arxiv.org/abs/2512.13511)
+[Adapting MLLMs for Nuanced Video Retrieval (ECCV 2026)](https://arxiv.org/abs/2512.13511)
 
 <p align="center">
-  <a href="https://bpiyush.github.io/tara-website/" target="_blank">
+  <img src="https://img.shields.io/badge/ECCV-2026-blue" alt="ECCV 2026">
+  &nbsp;&nbsp;&nbsp;
+  <a href="https://www.robots.ox.ac.uk/~vgg/research/tara/" target="_blank">
     <img src="https://img.shields.io/badge/Project-Page-blue" alt="Project Page">
   </a>
   &nbsp;&nbsp;&nbsp;
@@ -35,6 +38,10 @@ This repository contains inference and evaluation code for the TARA model based
     <img src="https://img.shields.io/badge/arXiv-Paper-b31b1b?logo=arxiv&logoColor=white" alt="arXiv">
   </a>
   &nbsp;&nbsp;&nbsp;
+  <a href="https://huggingface.co/bpiyush/TARA" target="_blank">
+    <img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-yellow" alt="Hugging Face Model">
+  </a>
+  &nbsp;&nbsp;&nbsp;
   <a href="https://huggingface.co/datasets/bpiyush/chirality-in-action" target="_blank">
     <img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/dataset-on-hf-md-dark.svg" alt="Dataset on Hugging Face">
   </a>
@@ -56,6 +63,7 @@ This repository contains inference and evaluation code for the TARA model based
   - [Data Preparation](#data-preparation)
   - [Embedding Computation](#embedding-computation)
   - [General evaluation: MMEB-V2 (Meng et al.)](#general-evaluation-mmeb-v2-meng-et-al)
+- [Training](#training)
 - [Citation](#citation)
 - [License](#license)
 
@@ -100,25 +108,56 @@ This will download all model weights (may take a few minutes depending on your c
 
 ### 3. Install Dependencies
 
+> [!TIP]
+> **Prefer `uv` over conda/pip?** You can set up the same environment with [uv](https://github.com/astral-sh/uv):
+>
+> **GPU requirement:** the prebuilt `flash-attn` wheel used below requires an **Ampere-generation GPU or newer** (compute capability ≥ 8.0) — e.g. RTX A4000/A5000/A6000, A100, RTX 30/40-series, H100. Older GPUs (Turing, Volta, T4, V100) are not supported by flash-attn 2.x. Tested and confirmed working on an **NVIDIA RTX A4000**.
+> ```bash
+> # 1. Install uv (if needed)
+> curl -LsSf https://astral.sh/uv/install.sh | sh
+>
+> # 2. Sync the environment (reads pyproject.toml, installs torch/flash-attn/etc.)
+> uv sync
+>
+> # 3. Quick check that CUDA + flash-attn are wired up correctly
+> uv run python -c "
+> import torch, flash_attn
+> print('torch:', torch.__version__, '| CUDA available:', torch.cuda.is_available())
+> print('flash-attn:', flash_attn.__version__)
+> "
+> ```
+> Once synced, prefix any command from this README with `uv run` instead of activating a conda env, e.g.:
+> ```bash
+> uv run python demo_usage.py --model_path /path/to/download/tara
+> ```
+
+**Step 1 — Create and activate a conda environment:**
+```bash
+conda create -n tara python=3.10 -y
+conda activate tara
+```
 
-* Create/activate the conda env (skip if you already have it):
-   ```bash
-   conda create -n tara python=3.10 -y
-   conda activate tara
-   ```
-* Install CUDA 12.1 PyTorch wheels (adjust the index URL if you need a different CUDA/CPU build):
-   ```bash
-   pip install --index-url https://download.pytorch.org/whl/cu121 \
-     torch==2.5.1+cu121 torchvision==0.20.1+cu121 torchaudio==2.5.1+cu121
-   ```
-* Install the remaining model dependencies:
-   ```bash
-   pip install -r requirements.txt
-   ```
-* (Optional) Verify the install:
-   ```bash
-   python -c "import torch, transformers; print(torch.cuda.is_available(), transformers.__version__)"
-   ```
+**Step 2 — Install CUDA-specific wheels** (adjust the index URL if you need a different CUDA build):
+```bash
+pip install --index-url https://download.pytorch.org/whl/cu121 \
+  torch==2.5.1+cu121 torchvision==0.20.1+cu121 torchaudio==2.5.1+cu121
+```
+
+**Step 3 — Install `flash-attn`** (must be built against the torch version installed above):
+```bash
+pip install flash-attn==2.8.3 --no-build-isolation
+```
+
+**Step 4 — Install TARA and all remaining dependencies:**
+```bash
+pip install -e /path/to/tara
+```
+where `/path/to/tara` is the root of this cloned repository. If you are already inside it, use `.` instead.
+
+**(Optional) Verify the install:**
+```bash
+python -c "import torch, transformers, modeling_tara; print(torch.cuda.is_available(), transformers.__version__)"
+```
 
 
 ## Quick Start
@@ -278,25 +317,82 @@ python evals/compute_metrics.py \
 
 We evaluate on the video classification and video retrieval tasks in MMEB-V2 to demonstrate the generalizability of TARA.
 
-TODO
+First, compute video embeddings for MMEB-V2:
+
+```bash
+# Video classification tracks
+python evals/compute_video_embeddings_mmeb.py \
+  --model_path /path/to/download/tara \
+  --model_name tara_7b \
+  --task cls
+
+# Video retrieval tracks
+python evals/compute_video_embeddings_mmeb.py \
+  --model_path /path/to/download/tara \
+  --model_name tara_7b \
+  --task ret
+```
+
+Then compute text embeddings:
+
+```bash
+# Classification text pools
+python evals/compute_text_embeddings_mmeb.py \
+  --model_path /path/to/download/tara \
+  --model_name tara_7b \
+  --task cls
+
+# Retrieval text pools
+python evals/compute_text_embeddings_mmeb.py \
+  --model_path /path/to/download/tara \
+  --model_name tara_7b \
+  --task ret
+```
+
+Finally, compute MMEB-V2 metrics:
+
+```bash
+python evals/compute_metrics_mmebv2.py \
+  --model_path /path/to/download/tara \
+  --model_name tara_7b \
+  --task all
+```
+
+By default, embeddings and metrics are saved to:
+
+```bash
+/path/to/download/tara/embs/tara_7b_video_embeddings_mmebv2_video_cls.pt
+/path/to/download/tara/embs/tara_7b_video_embeddings_mmebv2_video_ret.pt
+/path/to/download/tara/embs/tara_7b_text_embeddings_mmebv2_text_cls.pt
+/path/to/download/tara/embs/tara_7b_text_embeddings_mmebv2_text_ret.pt
+/path/to/download/tara/embs/metrics_tara_7b_mmebv2.json
+```
+
+If your MMEB-V2 data/config paths are different, override:
+`--data_root`, `--video_cls_cfg`, `--video_ret_cfg`, `--video_ret_root`, `--save_dir`, and `--feat_dir`.
+
+
+## Training
+
+See [TRAIN.md](TRAIN.md) for the full training pipeline (data format, fine-tuning, and weight merging).
 
 ## Citation
 
 If you use this model, please cite:
 ```bibtex
-@article{tara2025,
+@article{bagad2026tara,
   title={Adapting MLLMs for Nuanced Video Retrieval},
-  author={Piyush Bagad and Andrew Zisserman},
-  year={2025}
-  journal={arXiv preprint arXiv:2512.13511}
+  author={Bagad, Piyush and Zisserman, Andrew},
+  journal={European Conference on Computer Vision (ECCV)},
+  year={2026}
 }
 ```
 
 ```bibtex
-@article{bagad2025chirality,
+@inproceedings{bagad2025chirality,
   title={Chirality in Action: Time-Aware Video Representation Learning by Latent Straightening},
   author={Bagad, Piyush and Zisserman, Andrew},
-  journal={arXiv preprint arXiv:2509.08502},
+  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
   year={2025}
 }
 ```