app.py

import gradio as gr
import pickle
import numpy as np
import glob
import tqdm
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel, set_seed
from peft import PeftModel
import logging
import os
import json
import spaces
import ir_datasets
import pytrec_eval
from huggingface_hub import login
import transformers
import peft
import faiss
import sys
from collections import defaultdict

set_seed(42)

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Authenticate with HF_TOKEN
login(token=os.environ['HF_TOKEN'])

# Global variables
CUR_MODEL = "Samaya-AI/Promptriever-Llama2-v1"
BASE_MODEL = "meta-llama/Llama-2-7b-hf"
tokenizer = None
model = None
retrievers = {}
corpus_lookups = {}
queries = {}
q_lookups = {}
qrels = {}
query2qid = {}
datasets = ["scifact"]
current_dataset = "scifact"
faiss_index = None

def log_system_info():
    logger.info("System Information:")
    logger.info(f"Python version: {sys.version}")
    
    logger.info("\nPackage Versions:")
    logger.info(f"torch: {torch.__version__}")
    logger.info(f"transformers: {transformers.__version__}")
    logger.info(f"peft: {peft.__version__}")
    logger.info(f"faiss: {faiss.__version__}")
    logger.info(f"gradio: {gr.__version__}")
    logger.info(f"ir_datasets: {ir_datasets.__version__}")
    
    if torch.cuda.is_available():
        logger.info(f"\nCUDA Information:")
        logger.info(f"CUDA available: Yes")
        logger.info(f"CUDA version: {torch.version.cuda}")
        logger.info(f"cuDNN version: {torch.backends.cudnn.version()}")
        logger.info(f"Number of GPUs: {torch.cuda.device_count()}")
        for i in range(torch.cuda.device_count()):
            logger.info(f"GPU {i}: {torch.cuda.get_device_name(i)}")
    else:
        logger.info("\nCUDA Information:")
        logger.info("CUDA available: No")

log_system_info()

def pool(last_hidden_states, attention_mask, pool_type="last"):
    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)

    if pool_type == "last":
        left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
        if left_padding:
            emb = last_hidden[:, -1]
        else:
            sequence_lengths = attention_mask.sum(dim=1) - 1
            batch_size = last_hidden.shape[0]
            emb = last_hidden[torch.arange(batch_size, device=last_hidden.device), sequence_lengths]
    else:
        raise ValueError(f"pool_type {pool_type} not supported")

    return emb

def create_batch_dict(tokenizer, input_texts, always_add_eos="last", max_length=512):
    batch_dict = tokenizer(
        input_texts,
        max_length=max_length - 1,
        return_token_type_ids=False,
        return_attention_mask=False,
        padding=False,
        truncation=True
    )

    if always_add_eos == "last":
        batch_dict['input_ids'] = [input_ids + [tokenizer.eos_token_id] for input_ids in batch_dict['input_ids']]

    return tokenizer.pad(
        batch_dict,
        padding=True,
        pad_to_multiple_of=8,
        return_attention_mask=True,
        return_tensors="pt",
    )

class RepLlamaModel:
    def __init__(self, model_name_or_path):
        self.base_model = "meta-llama/Llama-2-7b-hf"
        self.tokenizer = AutoTokenizer.from_pretrained(self.base_model)
        self.tokenizer.model_max_length = 2048
        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
        self.tokenizer.pad_token = self.tokenizer.eos_token
        self.tokenizer.padding_side = "right"

        self.model = self.get_model(model_name_or_path)
        self.model.config.max_length = 2048

    def get_model(self, peft_model_name):
        base_model = AutoModel.from_pretrained(self.base_model)
        model = PeftModel.from_pretrained(base_model, peft_model_name)
        model = model.merge_and_unload()
        model.eval()
        return model

    def encode(self, texts, batch_size=48, **kwargs):
        self.model = self.model.cuda()
        all_embeddings = []
        for i in range(0, len(texts), batch_size):
            batch_texts = texts[i:i+batch_size]
            
            batch_dict = create_batch_dict(self.tokenizer, batch_texts, always_add_eos="last")
            batch_dict = {key: value.cuda() for key, value in batch_dict.items()}

            with torch.cuda.amp.autocast():
                with torch.no_grad():
                    outputs = self.model(**batch_dict)
                    embeddings = pool(outputs.last_hidden_state, batch_dict['attention_mask'], 'last')
                    embeddings = F.normalize(embeddings, p=2, dim=-1)
                    logger.info(f"Encoded shape: {embeddings.shape}, Norm of first embedding: {torch.norm(embeddings[0]).item()}")
                    all_embeddings.append(embeddings.cpu().numpy())

        self.model = self.model.cpu()
        return np.concatenate(all_embeddings, axis=0)

def load_corpus_embeddings(dataset_name):
    corpus_path = f"{dataset_name}/corpus_emb.*.pkl"
    index_files = glob.glob(corpus_path)
    index_files.sort(key=lambda x: int(x.split('.')[-2]))
    
    all_embeddings = []
    corpus_lookups = []
    
    for file in index_files:
        with open(file, 'rb') as f:
            embeddings, p_lookup = pickle.load(f)
        all_embeddings.append(embeddings)
        corpus_lookups.extend(p_lookup)
    
    all_embeddings = np.concatenate(all_embeddings, axis=0)
    logger.info(f"Loaded corpus embeddings for {dataset_name}. Shape: {all_embeddings.shape}")
    
    return all_embeddings, corpus_lookups

def create_faiss_index(embeddings):
    dimension = embeddings.shape[1]
    index = faiss.IndexFlatIP(dimension)
    index.add(embeddings)
    logger.info(f"Created FAISS index with {index.ntotal} vectors of dimension {dimension}")
    return index

def load_or_create_faiss_index(dataset_name):
    embeddings, corpus_lookups = load_corpus_embeddings(dataset_name)
    index = create_faiss_index(embeddings)
    return index, corpus_lookups

def initialize_faiss_and_corpus(dataset_name):
    global corpus_lookups
    index, corpus_lookups[dataset_name] = load_or_create_faiss_index(dataset_name)
    logger.info(f"Initialized FAISS index and corpus lookups for {dataset_name}")
    return index

def search_queries(dataset_name, q_reps, depth=100):
    global faiss_index    
    logger.info(f"Searching queries. Shape of q_reps: {q_reps.shape}")
    
    # Perform the search
    all_scores, all_indices = faiss_index.search(q_reps, depth)
    
    logger.info(f"Search completed. Shape of all_scores: {all_scores.shape}, all_indices: {all_indices.shape}")
    logger.info(f"Sample scores: {all_scores[0][:5]}, Sample indices: {all_indices[0][:5]}")
    
    psg_indices = [[str(corpus_lookups[dataset_name][x]) for x in q_dd] for q_dd in all_indices]
    
    return all_scores, np.array(psg_indices)

def load_queries(dataset_name):
    global queries, q_lookups, qrels, query2qid
    dataset = ir_datasets.load(f"beir/{dataset_name.lower()}" + ("/test" if dataset_name == "scifact" else ""))
    
    queries[dataset_name] = []
    query2qid[dataset_name] = defaultdict(dict)
    q_lookups[dataset_name] = {}
    qrels[dataset_name] = {}
    for query in dataset.queries_iter():
        queries[dataset_name].append(query.text)
        q_lookups[dataset_name][query.query_id] = query.text
        query2qid[dataset_name][query.text] = query.query_id
    
    for qrel in dataset.qrels_iter():
        if qrel.query_id not in qrels[dataset_name]:
            qrels[dataset_name][qrel.query_id] = {}
        qrels[dataset_name][qrel.query_id][qrel.doc_id] = qrel.relevance

    logger.info(f"Loaded queries for {dataset_name}. Total queries: {len(queries[dataset_name])}")
    logger.info(f"Loaded qrels for {dataset_name}. Total query IDs: {len(qrels[dataset_name])}")

def evaluate(qrels, results, k_values):
    qrels = {str(k): {str(k2): v2 for k2, v2 in v.items()} for k, v in qrels.items()}
    results = {str(k): {str(k2): v2 for k2, v2 in v.items()} for k, v in results.items()}
    evaluator = pytrec_eval.RelevanceEvaluator(
        qrels, {f"ndcg_cut.{k}" for k in k_values} | {f"recall.{k}" for k in k_values}
    )
    scores = evaluator.evaluate(results)

    metrics = {}
    for k in k_values:
        ndcg_scores = [query_scores[f"ndcg_cut_{k}"] for query_scores in scores.values()]
        recall_scores = [query_scores[f"recall_{k}"] for query_scores in scores.values()]
        metrics[f"NDCG@{k}"] = round(np.mean(ndcg_scores), 3)
        metrics[f"Recall@{k}"] = round(np.mean(recall_scores), 3)
        logger.info(f"NDCG@{k}: mean={metrics[f'NDCG@{k}']}, min={min(ndcg_scores)}, max={max(ndcg_scores)}")
        logger.info(f"Recall@{k}: mean={metrics[f'Recall@{k}']}, min={min(recall_scores)}, max={max(recall_scores)}")

    # delete nDCG@100 and Recall@10
    del metrics["NDCG@100"]
    del metrics["Recall@100"]
    return metrics

@spaces.GPU
def run_evaluation(dataset, postfix):
    global current_dataset, queries, model, query2qid
    current_dataset = dataset

    input_texts = [f"query: {query.strip()} {postfix}".strip() for query in queries[current_dataset]]
    logger.info(f"Number of input texts: {len(input_texts)}")
    logger.info(f"Sample input text: {input_texts[0]}")

    q_reps = model.encode(input_texts)
    logger.info(f"Encoded query first five: {q_reps[0][:5]}")
    logger.info(f"Encoded query representations shape: {q_reps.shape}")

    all_scores, psg_indices = search_queries(dataset, q_reps)
    
    results = {}
    logging.info(f"Number of queries in q_lookups: {len(q_lookups[dataset])}")
    logging.info("Size of all_scores: " + str(len(all_scores)))
    logging.info("Size of psg_indices: " + str(len(psg_indices)))
    for query, scores, doc_ids in zip(queries[current_dataset], all_scores, psg_indices):
        qid = query2qid[dataset][query]
        qid_str = str(qid)
        results[qid_str] = {}
        for doc_id, score in zip(doc_ids, scores):
            doc_id_str = str(doc_id)
            results[qid_str][doc_id_str] = float(score)

        if not results[qid_str]:  # If no results for this query
            logger.warning(f"No results for query {qid_str}")

    logger.info(f"Number of queries in results: {len(results)}")
    logger.info(f"Sample result: {next(iter(results.items()))}")

    qrels[dataset] = {str(qid): {str(doc_id): rel for doc_id, rel in rels.items()}
                  for qid, rels in qrels[dataset].items()}

    logger.info(f"Number of results: {len(results)}")
    logger.info(f"Sample result: {list(results.items())[0]}")
    
    logger.info(f"Number of queries in qrels: {len(qrels[dataset])}")
    logger.info(f"Sample qrel: {list(qrels[dataset].items())[0]}")
    logger.info(f"Number of queries in results: {len(results)}")
    logger.info(f"Sample result: {list(results.items())[0]}")
    
    # Check for mismatches
    qrels_keys = set(qrels[dataset].keys())
    results_keys = set(results.keys())
    logger.info(f"Queries in qrels but not in results: {qrels_keys - results_keys}")
    logger.info(f"Queries in results but not in qrels: {results_keys - qrels_keys}")
    
    metrics = evaluate(qrels[dataset], results, k_values=[10, 100])
    
    return metrics

@spaces.GPU
def gradio_interface(dataset, postfix):
    return run_evaluation(dataset, postfix)

if model is None:
    model = RepLlamaModel(model_name_or_path=CUR_MODEL)
    load_queries(current_dataset)   
    faiss_index = initialize_faiss_and_corpus(current_dataset)

# Create Gradio interface
iface = gr.Interface(
    fn=gradio_interface,
    inputs=[
        gr.Dropdown(choices=datasets, label="Dataset", value="scifact"),
        gr.Textbox(label="Prompt")
    ],
    outputs=gr.JSON(label="Evaluation Results"),
    title="Promptriever Demo",
    description="Enter a prompt to evaluate the model's performance on SciFact. Note: it takes about **30 seconds** to evaluate.",
    examples=[
        ["scifact", ""],
        ["scifact", "Think carefully about these conditions when determining relevance"]
    ],
    cache_examples=False,
)

# Launch the interface
iface.launch(share=False)