The huggingface-transformers implementation of Fine-grained Late-interaction Multi-modal Retriever.
The official implementation is at here.
The details of the model and checkpoints can be found here.
The details for reproducing the datasets and evaluation in the paper can be found here.
- [03/09/2024] We have uploaded the images used in the M2KR benchmark here .
- [10/08/2024] We received many requests regarding adding multilingual abilities to PreFLMR. We announce that we are now training the Chinese version of PreFLMR and will release it very soon. Stay tuned!
- [05/06/2024] 🔥🔥🔥We made some updates to the implementation
- Added an evaluation script that reproduces the results in the PreFLMR paper here
- Added the updated benchmark results with the transformer implementation here
- Added an example script to fine-tune PreFLMR on a custom retrieval dataset here
- IMPORTANT: fixed the OVEN data splits in the M2KR benchmark, and updated each entry with a fixed instruction to ensure the evaluation result is not affected by random sampling of instructions. Please delete your local cache and download the dataset again.
- FLMR
| Model | WIT Recall@10 | IGLUE Recall@1 | KVQA Recall@5 | MSMARCO Recall@5 | OVEN Recall@5 | LLaVA Recall@1 | EVQA Recall@5 | EVQA Pseudo Recall@5 | OKVQA Recall@5 | OKVQA Pseudo Recall@5 | Infoseek Recall@5 | Infoseek Pseudo Recall@5 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LinWeizheDragon/PreFLMR_ViT-G🤗 | 0.619 | 0.718 | 0.419 | 0.783 | 0.643 | 0.726 | 0.625 | 0.721 | 0.302 | 0.674 | 0.392 | 0.577 |
| LinWeizheDragon/PreFLMR_ViT-L🤗 | 0.605 | 0.699 | 0.440 | 0.779 | 0.608 | 0.729 | 0.609 | 0.708 | 0.314 | 0.690 | 0.374 | 0.578 |
| LinWeizheDragon/PreFLMR_ViT-B🤗 | 0.427 | 0.574 | 0.294 | 0.786 | 0.468 | 0.673 | 0.550 | 0.663 | 0.272 | 0.658 | 0.260 | 0.496 |
Note: We converted the checkpoints from PyTorch to Huggingface-transformers, whose benchmark results differ from the numbers reported in the original paper slightly. You can reproduce the results in the above paper by referring to the instructions in this document.
Create virtualenv:
conda create -n FLMR python=3.10 -y
conda activate FLMR
Install Pytorch:
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
Install faiss
conda install -c pytorch -c nvidia faiss-gpu=1.7.4 mkl=2021 blas=1.0=mkl
Test if faiss generate error
python -c "import faiss"
Install FLMR
git clone https://github.com/LinWeizheDragon/FLMR.git
cd FLMR
pip install -e .
Install ColBERT engine
cd third_party/ColBERT
pip install -e .
Install other dependencies
pip install ujson gitpython easydict ninja datasets transformers
-
Load pre-trained models
import os import torch import pandas as pd import numpy as np from torchvision.transforms import ToPILImage from transformers import AutoImageProcessor from flmr import index_custom_collection from flmr import FLMRQueryEncoderTokenizer, FLMRContextEncoderTokenizer, FLMRModelForRetrieval # load models checkpoint_path = "LinWeizheDragon/PreFLMR_ViT-G" image_processor_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k" query_tokenizer = FLMRQueryEncoderTokenizer.from_pretrained(checkpoint_path, subfolder="query_tokenizer") context_tokenizer = FLMRContextEncoderTokenizer.from_pretrained( checkpoint_path, subfolder="context_tokenizer" ) model = FLMRModelForRetrieval.from_pretrained( checkpoint_path, query_tokenizer=query_tokenizer, context_tokenizer=context_tokenizer, ) image_processor = AutoImageProcessor.from_pretrained(image_processor_name)
-
Create document collections
num_items = 100 feature_dim = 1664 passage_contents = [f"This is test sentence {i}" for i in range(num_items)] # Option 1. text-only documents custom_collection = passage_contents # Option 2. multi-modal documents with pre-extracted image features # passage_image_features = np.random.rand(num_items, feature_dim) # custom_collection = [ # (passage_content, passage_image_feature, None) for passage_content, passage_image_feature in zip(passage_contents, passage_image_features) # ] # Option 3. multi-modal documents with images # random_images = torch.randn(num_items, 3, 224, 224) # to_img = ToPILImage() # if not os.path.exists("./test_images"): # os.makedirs("./test_images") # for i, image in enumerate(random_images): # image = to_img(image) # image.save(os.path.join("./test_images", "{}.jpg".format(i))) # image_paths = [os.path.join("./test_images", "{}.jpg".format(i)) for i in range(num_items)] # custom_collection = [ # (passage_content, None, image_path) # for passage_content, image_path in zip(passage_contents, image_paths) # ]
-
Run indexing on the custom collection
index_custom_collection( custom_collection=custom_collection, model=model, index_root_path=".", index_experiment_name="test_experiment", index_name="test_index", nbits=8, # number of bits in compression doc_maxlen=512, # maximum allowed document length overwrite=True, # whether to overwrite existing indices use_gpu=False, # whether to enable GPU indexing indexing_batch_size=64, model_temp_folder="tmp", nranks=1, # number of GPUs used in indexing )
-
Create toy query data
num_queries = 2 query_instructions = [f"instruction {i}" for i in range(num_queries)] query_texts = [f"{query_instructions[i]} : query {i}" for i in range(num_queries)] query_images = torch.zeros(num_queries, 3, 224, 224) query_encoding = query_tokenizer(query_texts) query_pixel_values = image_processor(query_images, return_tensors="pt")['pixel_values']
-
Obtain query embeddings with model
inputs = dict( input_ids=query_encoding['input_ids'], attention_mask=query_encoding['attention_mask'], pixel_values=query_pixel_values, ) # Run model query encoding res = model.query(**inputs) queries = {i: query_texts[i] for i in range(num_queries)} query_embeddings = res.late_interaction_output
-
Search the collection
from flmr import search_custom_collection, create_searcher # initiate a searcher searcher = create_searcher( index_root_path=".", index_experiment_name="test_experiment", index_name="test_index", nbits=8, # number of bits in compression use_gpu=True, # whether to enable GPU searching ) # Search the custom collection ranking = search_custom_collection( searcher=searcher, queries=queries, query_embeddings=query_embeddings, num_document_to_retrieve=5, # how many documents to retrieve for each query ) # Analyse retrieved documents ranking_dict = ranking.todict() for i in range(num_queries): print(f"Query {i} retrieved documents:") retrieved_docs = ranking_dict[i] retrieved_docs_indices = [doc[0] for doc in retrieved_docs] retrieved_doc_scores = [doc[2] for doc in retrieved_docs] retrieved_doc_texts = [passage_contents[doc_idx] for doc_idx in retrieved_docs_indices] data = { "Confidence": retrieved_doc_scores, "Content": retrieved_doc_texts, } df = pd.DataFrame.from_dict(data) print(df)
import torch
from flmr import FLMRQueryEncoderTokenizer, FLMRContextEncoderTokenizer, FLMRModelForRetrieval
checkpoint_path = "LinWeizheDragon/PreFLMR_ViT-L"
image_processor_name = "openai/clip-vit-large-patch14"
query_tokenizer = FLMRQueryEncoderTokenizer.from_pretrained(checkpoint_path, subfolder="query_tokenizer")
context_tokenizer = FLMRContextEncoderTokenizer.from_pretrained(checkpoint_path, subfolder="context_tokenizer")
model = FLMRModelForRetrieval.from_pretrained(checkpoint_path,
query_tokenizer=query_tokenizer,
context_tokenizer=context_tokenizer,
)
Q_encoding = query_tokenizer(["Using the provided image, obtain documents that address the subsequent question: What is the capital of France?", "Extract documents linked to the question provided in conjunction with the image: What is the capital of China?"])
D_encoding = context_tokenizer(["Paris is the capital of France.", "Beijing is the capital of China.",
"Paris is the capital of France.", "Beijing is the capital of China."])
Q_pixel_values = torch.zeros(2, 3, 224, 224)
inputs = dict(
query_input_ids=Q_encoding['input_ids'],
query_attention_mask=Q_encoding['attention_mask'],
query_pixel_values=Q_pixel_values,
context_input_ids=D_encoding['input_ids'],
context_attention_mask=D_encoding['attention_mask'],
use_in_batch_negatives=True,
)
res = model.forward(**inputs)
print(res)Note that the examples in this code block are only for demonstration purposes. They show that the pre-trained model gives higher scores to correct documents. In real training, you always need to pass in the documents in the order "positive doc for query1, negative doc1 for query1, negative doc2 for query1, ..., positive doc for query2, negative doc1 for query2, negative doc2 for query2, ...". You may want to read the later section which provides an example finetuning script.
pip install transformers
from transformers import AutoConfig, AutoModel, AutoImageProcessor, AutoTokenizer
import torch
checkpoint_path = "LinWeizheDragon/PreFLMR_ViT-L"
image_processor_name = "openai/clip-vit-large-patch14"
query_tokenizer = AutoTokenizer.from_pretrained(checkpoint_path, subfolder="query_tokenizer", trust_remote_code=True)
context_tokenizer = AutoTokenizer.from_pretrained(checkpoint_path, subfolder="context_tokenizer", trust_remote_code=True)
model = AutoModel.from_pretrained(checkpoint_path,
query_tokenizer=query_tokenizer,
context_tokenizer=context_tokenizer,
trust_remote_code=True,
)
image_processor = AutoImageProcessor.from_pretrained(image_processor_name)We provide two scripts to show how the pretrained models can be used in evaluation:
examples/example_use_flmr.py: an example script to evaluate FLMR (with 10 ROIs) on OK-VQA.examples/example_use_preflmr.py: an example script to evaluate PreFLMR on E-VQA.
cd examples/Download KBVQA_data from here and unzip the image folders. The ROI/captioning/object detection results have been included.
Run the following command (remove --run_indexing if you have already run indexing once):
python example_use_flmr.py \
--use_gpu --run_indexing \
--index_root_path "." \
--index_name OKVQA_GS\
--experiment_name OKVQA_GS \
--indexing_batch_size 64 \
--image_root_dir /path/to/KBVQA_data/ok-vqa/ \
--dataset_path BByrneLab/OKVQA_FLMR_preprocessed_data \
--passage_dataset_path BByrneLab/OKVQA_FLMR_preprocessed_GoogleSearch_passages \
--use_split test \
--nbits 8 \
--Ks 1 5 10 20 50 100 \
--checkpoint_path LinWeizheDragon/FLMR \
--image_processor_name openai/clip-vit-base-patch32 \
--query_batch_size 8 \
--num_ROIs 9 \You can download the E-VQA images from https://github.com/google-research/google-research/tree/master/encyclopedic_vqa. We will add a dataset link here soon.
cd examples/Run the following command (remove --run_indexing if you have already run indexing once):
python example_use_preflmr.py \
--use_gpu --run_indexing \
--index_root_path "." \
--index_name EVQA_PreFLMR_ViT-G \
--experiment_name EVQA \
--indexing_batch_size 64 \
--image_root_dir /rds/project/rds-hirYTW1FQIw/shared_space/vqa_data/KBVQA_data/EVQA/eval_image/ \
--dataset_hf_path BByrneLab/multi_task_multi_modal_knowledge_retrieval_benchmark_M2KR \
--dataset EVQA \
--use_split test \
--nbits 8 \
--Ks 1 5 10 20 50 100 500 \
--checkpoint_path LinWeizheDragon/PreFLMR_ViT-G \
--image_processor_name laion/CLIP-ViT-bigG-14-laion2B-39B-b160k \
--query_batch_size 8 \
--compute_pseudo_recall \Here, we upload all the M2KR datasets into one HF dataset BByrneLab/multi_task_multi_modal_knowledge_retrieval_benchmark_M2KR with different datasets as subset.
To reproduce results of the other datasets in the paper, you can change the --dataset to OKVQA, KVQA, LLaVA, OVEN, Infoseek, WIT, IGLUE and EVQA.
Updates:
- Enable
--compute_pseudo_recallto compute pseudo recall for datasets like EVQA/OKVQA/Infoseek - Enable
--Ks 1 5 10 20 50 100 500: max(Ks) needs to be 500 to match the performance reported in the PreFLMR paper.
Change the image root paths in examples/evaluate_all.sh and execute:
cd examples
bash evaluate_all.shObtain the report by:
python report.pyYou will need to install pytorch-lightning:
pip install pytorch-lightning==2.1.0
python example_finetune_preflmr.py \
--image_root_dir /path/to/EVQA/images/ \
--dataset_hf_path BByrneLab/multi_task_multi_modal_knowledge_retrieval_benchmark_M2KR \
--dataset EVQA \
--freeze_vit \
--log_with_wandb \
--model_save_path saved_models \
--checkpoint_path LinWeizheDragon/PreFLMR_ViT-G \
--image_processor_name laion/CLIP-ViT-bigG-14-laion2B-39B-b160k \
--batch_size 8 \
--accumulate_grad_batches 8 \
--valid_batch_size 16 \
--test_batch_size 64 \
--mode train \
--max_epochs 99999999 \
--learning_rate 0.000005 \
--warmup_steps 100 \
--accelerator auto \
--devices auto \
--strategy ddp_find_unused_parameters_true \
--num_sanity_val_steps 2 \
--precision bf16 \
--val_check_interval 2000 \
--save_top_k -1 \python example_use_preflmr.py \
--use_gpu --run_indexing \
--index_root_path "." \
--index_name EVQA_PreFLMR_ViT-G_finetuned_model_step_10156 \
--experiment_name EVQA \
--indexing_batch_size 64 \
--image_root_dir /path/to/EVQA/images/ \
--dataset_hf_path BByrneLab/multi_task_multi_modal_knowledge_retrieval_benchmark_M2KR \
--dataset EVQA \
--use_split test \
--nbits 8 \
--num_gpus 1 \
--Ks 1 5 10 20 50 100 500 \
--checkpoint_path saved_models/model_step_10156 \
--image_processor_name laion/CLIP-ViT-bigG-14-laion2B-39B-b160k \
--query_batch_size 8 \By running the above script, we are able to obtain the following finetuning performance:
| Step | Pseudo Recall@5 on EVQA |
|---|---|
| 2500 | 73.6 |
| 10000 | 73.55 |
| 12000 | 74.21 |
| 14000 | 73.73 |
(Checkpoints with low validation losses were picked and tested, run on 2 A100 GPUs)
The FLMR model is implemented following the documentation style of transformers. You can find detailed documentation in the modeling files.
If our work helped your research, please kindly cite our paper for FLMR and PreFLMR.
@inproceedings{
lin2023finegrained,
title={Fine-grained Late-interaction Multi-modal Retrieval for Retrieval Augmented Visual Question Answering},
author={Weizhe Lin and Jinghong Chen and Jingbiao Mei and Alexandru Coca and Bill Byrne},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=IWWWulAX7g}
}
@inproceedings{lin-etal-2024-preflmr,
title = "{P}re{FLMR}: Scaling Up Fine-Grained Late-Interaction Multi-modal Retrievers",
author = "Lin, Weizhe and
Mei, Jingbiao and
Chen, Jinghong and
Byrne, Bill",
editor = "Ku, Lun-Wei and
Martins, Andre and
Srikumar, Vivek",
booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = aug,
year = "2024",
address = "Bangkok, Thailand",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.acl-long.289",
pages = "5294--5316",
abstract = "Large Multimodal Models (LMMs) excel in natural language and visual understanding but are challenged by exacting tasks such as Knowledge-based Visual Question Answering (KB-VQA) which involve the retrieval of relevant information from document collections to use in shaping answers to questions. We present an extensive training and evaluation framework, M2KR, for KB-VQA. M2KR contains a collection of vision and language tasks which we have incorporated into a single suite of benchmark tasks for training and evaluating general-purpose multi-modal retrievers. We use M2KR to develop PreFLMR, a pre-trained version of the recently developed Fine-grained Late-interaction Multi-modal Retriever (FLMR) approach to KB-VQA, and we report new state-of-the-art results across a range of tasks. We also present investigations into the scaling behaviors of PreFLMR intended to be useful in future developments in general-purpose multi-modal retrievers.",
}