初始化项目，由ModelHub XC社区提供模型

Model: cross-encoder/mmarco-mMiniLMv2-L12-H384-v1
Source: Original Platform

.gitattributes

@@ -0,0 +1,28 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+model.safetensors filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,63 @@
+---
+license: apache-2.0
+language:
+- en
+- ar
+- zh
+- nl
+- fr
+- de
+- hi
+- in
+- it
+- ja
+- pt
+- ru
+- es
+- vi
+- multilingual
+datasets:
+- unicamp-dl/mmarco
+base_model:
+- nreimers/mMiniLMv2-L12-H384-distilled-from-XLMR-Large
+pipeline_tag: text-ranking
+library_name: sentence-transformers
+tags:
+- transformers
+---
+# Cross-Encoder for multilingual MS Marco
+
+This model was trained on the [MMARCO](https://hf.co/unicamp-dl/mmarco) dataset. It is a machine translated version of MS MARCO using Google Translate. It was translated to 14 languages. In our experiments, we observed that it performs also well for other languages.
+
+As a base model, we used the [multilingual MiniLMv2](https://huggingface.co/nreimers/mMiniLMv2-L12-H384-distilled-from-XLMR-Large) model.
+
+The model can be used for Information Retrieval: Given a query, encode the query will all possible passages (e.g. retrieved with ElasticSearch). Then sort the passages in a decreasing order. See [SBERT.net Retrieve & Re-rank](https://www.sbert.net/examples/applications/retrieve_rerank/README.html) for more details. The training code is available here: [SBERT.net Training MS Marco](https://github.com/UKPLab/sentence-transformers/tree/master/examples/training/ms_marco)
+
+## Usage with SentenceTransformers
+
+The usage becomes easy when you have [SentenceTransformers](https://www.sbert.net/) installed. Then, you can use the pre-trained models like this:
+```python
+from sentence_transformers import CrossEncoder
+model = CrossEncoder('model_name')
+scores = model.predict([('Query', 'Paragraph1'), ('Query', 'Paragraph2') , ('Query', 'Paragraph3')])
+```
+
+
+
+
+## Usage with Transformers
+
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+
+model = AutoModelForSequenceClassification.from_pretrained('model_name')
+tokenizer = AutoTokenizer.from_pretrained('model_name')
+
+features = tokenizer(['How many people live in Berlin?', 'How many people live in Berlin?'], ['Berlin has a population of 3,520,031 registered inhabitants in an area of 891.82 square kilometers.', 'New York City is famous for the Metropolitan Museum of Art.'],  padding=True, truncation=True, return_tensors="pt")
+
+model.eval()
+with torch.no_grad():
+    scores = model(**features).logits
+    print(scores)
+```

config.json

@@ -0,0 +1,34 @@
+{
+  "_name_or_path": "nreimers/mMiniLMv2-L12-H384-distilled-from-XLMR-Large",
+  "architectures": [
+    "XLMRobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 384,
+  "id2label": {
+    "0": "LABEL_0"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 1536,
+  "label2id": {
+    "LABEL_0": 0
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "xlm-roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "torch_dtype": "float32",
+  "transformers_version": "4.18.0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 250002,
+  "sbert_ce_default_activation_function": "torch.nn.modules.linear.Identity"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5daeca2481a76b5976a2bdc32f0a78532b6716da4f8cd3ff59460ef8d2f359b4
+size 470592698

onnx/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e9a03ed1e966f7c5288dd4230e3d6a9bf5e3a170a06f1f4241c5bca12c6487c
+size 470883696

onnx/model_O1.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6230c9f55c7126a852c27655bdaf63df6f039fdb4e070ad3a73bab924dfc61ef
+size 470782077

onnx/model_O2.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04aca53ef8e0c1a553b6958c8e4480f96667a68fad9b6c24f42ba4144052b342
+size 470605122

onnx/model_O3.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3a4da3a7b0c57aa8abe5b0ce2c1afb30dbd74181d8a0943df3272e5deb39643
+size 470604977

onnx/model_O4.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:613b96ad598e707de925349280296d3fe0f792a03bb17ff4661b11781380a0c3
+size 235460006

onnx/model_qint8_arm64.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1825907d6c1a9001ff78124780bbde20a614a8c3df3b63409cf3c72c6fe5c8b4
+size 118620017

onnx/model_qint8_avx512.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1825907d6c1a9001ff78124780bbde20a614a8c3df3b63409cf3c72c6fe5c8b4
+size 118620017

onnx/model_qint8_avx512_vnni.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1825907d6c1a9001ff78124780bbde20a614a8c3df3b63409cf3c72c6fe5c8b4
+size 118620017

onnx/model_quint8_avx2.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c2513767fb63d008a4377bef7a7a3555433d9436342bb53e35a3a72ffc52d4b
+size 118620016

openvino/openvino_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b00b3db7a1a0d53c64b459e81900c1dfa7fd34c253556e4cc9b862fcd356c10
+size 470568628

openvino/openvino_model_qint8_quantized.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2911002d538ad8ea038cf554cb391b6f96fa65ba0ea524214ccf884d6ae28120
+size 119130716

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1abc209e54d70bbcb08c1b5111a924fb99c0428f51cab1659310ccdcab69dc03
+size 470633197

sentencepiece.bpe.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfc8146abe2a0488e9e2a0c56de7952f7c11ab059eca145a0a727afce0db2865
+size 5069051

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62c24cdc13d4c9952d63718d6c9fa4c287974249e16b7ade6d5a85e7bbb75626
+size 17082660

tokenizer_config.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "sep_token": "</s>", "cls_token": "<s>", "unk_token": "<unk>", "pad_token": "<pad>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "nreimers/mMiniLMv2-L12-H384-distilled-from-XLMR-Large", "tokenizer_class": "XLMRobertaTokenizer"}

train_script.py

@@ -0,0 +1,362 @@
+from codecs import EncodedFile
+from datetime import datetime
+from typing import Optional
+
+import datasets
+import torch
+from pytorch_lightning import LightningDataModule, LightningModule, Trainer, seed_everything
+from torch.utils.data import DataLoader
+from transformers import (
+    AutoConfig,
+    AutoModelForSequenceClassification,
+    AutoTokenizer,
+    get_linear_schedule_with_warmup,
+    get_scheduler,
+)
+import torch
+import sys
+import os
+from argparse import ArgumentParser
+from datasets import load_dataset
+import tqdm
+import json
+import gzip
+import random
+from pytorch_lightning.callbacks import ModelCheckpoint
+import numpy as np
+from shutil import copyfile
+from pytorch_lightning.loggers import WandbLogger
+import transformers
+
+
+class MSMARCOData(LightningDataModule):
+    def __init__(
+        self,
+        model_name: str,
+        triplets_path: str,
+        langs,
+        max_seq_length: int = 250,
+        train_batch_size: int = 32,
+        eval_batch_size: int = 32,
+        num_negs: int = 3,
+        cross_lingual_chance: float = 0.0,
+        **kwargs,
+    ):
+        super().__init__()
+        self.model_name = model_name
+        self.triplets_path = triplets_path
+        self.max_seq_length = max_seq_length
+        self.train_batch_size = train_batch_size
+        self.eval_batch_size = eval_batch_size
+        self.langs = langs
+        self.num_negs = num_negs
+        self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
+        self.cross_lingual_chance = cross_lingual_chance  #Probability for cross-lingual batches
+
+    #def setup(self, stage: str):
+        print(f"!!!!!!!!!!!!!!!!!! SETUP {os.getpid()}  !!!!!!!!!!!!!!!")
+
+        #Get the queries
+        self.queries = {lang: {} for lang in self.langs}
+    
+        for lang in self.langs:
+            for row in tqdm.tqdm(load_dataset('unicamp-dl/mmarco', f'queries-{lang}')['train'], desc=lang):
+                self.queries[lang][row['id']] = row['text']
+
+        #Get the passages
+        self.collections = {lang: load_dataset('unicamp-dl/mmarco', f'collection-{lang}')['collection'] for lang in self.langs}
+
+        #Get the triplets
+        with gzip.open(self.triplets_path, 'rt') as fIn:
+            self.triplets = [json.loads(line) for line in tqdm.tqdm(fIn, desc="triplets", total=502938)] 
+            """
+            self.triplets = []
+            for line in tqdm.tqdm(fIn):
+                self.triplets.append(json.loads(line))
+                if len(self.triplets) >= 1000:
+                    break
+            """
+
+    def collate_fn(self, batch):
+        cross_lingual_batch = random.random() < self.cross_lingual_chance 
+
+        #Create data for list-rank-loss
+        query_doc_pairs = [[] for _ in  range(1+self.num_negs)]
+
+        for row in batch:
+            qid = row['qid']
+            pos_id = random.choice(row['pos'])
+
+            query_lang = random.choice(self.langs)
+            query_text = self.queries[query_lang][qid]
+            
+            doc_lang = random.choice(self.langs) if cross_lingual_batch else query_lang 
+            query_doc_pairs[0].append((query_text, self.collections[doc_lang][pos_id]['text']))
+
+            dense_bm25_neg = list(set(row['dense_neg'] + row['bm25_neg']))
+            neg_ids = random.sample(dense_bm25_neg, self.num_negs)
+
+            for num_neg, neg_id in enumerate(neg_ids):
+                doc_lang = random.choice(self.langs) if cross_lingual_batch else query_lang
+                query_doc_pairs[1+num_neg].append((query_text, self.collections[doc_lang][neg_id]['text']))
+
+        #Now tokenize the data
+        features = [self.tokenizer(qd_pair, max_length=self.max_seq_length, padding=True, truncation='only_second', return_tensors="pt") for qd_pair in query_doc_pairs]
+    
+        return features
+
+    def train_dataloader(self):
+        return DataLoader(self.triplets, shuffle=True, batch_size=self.train_batch_size, num_workers=1, pin_memory=True, collate_fn=self.collate_fn)
+
+    
+
+
+
+class ListRankLoss(LightningModule):
+    def __init__(
+        self,
+        model_name: str,
+        learning_rate: float = 2e-5,
+        warmup_steps: int = 1000,
+        weight_decay: float = 0.01,
+        train_batch_size: int = 32,
+        eval_batch_size: int = 32,
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.save_hyperparameters()
+        print(self.hparams)
+
+        self.config = AutoConfig.from_pretrained(model_name, num_labels=1)
+        self.model = AutoModelForSequenceClassification.from_pretrained(model_name, config=self.config)
+        self.loss_fct = torch.nn.CrossEntropyLoss()
+        self.global_train_step = 0
+        
+
+    def forward(self, **inputs):
+        return self.model(**inputs)
+
+    def training_step(self, batch, batch_idx):
+        pred_scores = []
+        scores = torch.tensor([0] * len(batch[0]['input_ids']), device=self.model.device)
+   
+        for feature in batch:
+            pred_scores.append(self(**feature).logits.squeeze())
+
+        pred_scores = torch.stack(pred_scores, 1)
+        loss_value = self.loss_fct(pred_scores, scores)
+        self.global_train_step += 1
+        self.log('global_train_step', self.global_train_step)
+        self.log("train/loss", loss_value)
+
+        return loss_value
+     
+
+    def setup(self, stage=None) -> None:
+        if stage != "fit":
+            return
+        # Get dataloader by calling it - train_dataloader() is called after setup() by default
+        train_loader = self.trainer.datamodule.train_dataloader()
+
+        # Calculate total steps
+        tb_size = self.hparams.train_batch_size * max(1, self.trainer.gpus)
+        ab_size = self.trainer.accumulate_grad_batches
+        self.total_steps = (len(train_loader) // ab_size) * self.trainer.max_epochs
+
+        print(f"{tb_size=}")
+        print(f"{ab_size=}")
+        print(f"{len(train_loader)=}")
+        print(f"{self.total_steps=}")
+        
+
+    def configure_optimizers(self):
+        """Prepare optimizer and schedule (linear warmup and decay)"""
+        model = self.model
+        no_decay = ["bias", "LayerNorm.weight"]
+        optimizer_grouped_parameters = [
+            {
+                "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
+                "weight_decay": self.hparams.weight_decay,
+            },
+            {
+                "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
+                "weight_decay": 0.0,
+            },
+        ]
+        optimizer =  torch.optim.AdamW(optimizer_grouped_parameters, lr=self.hparams.learning_rate)
+
+        lr_scheduler = get_scheduler(
+            name="linear",
+            optimizer=optimizer,
+            num_warmup_steps=self.hparams.warmup_steps,
+            num_training_steps=self.total_steps,
+        )
+
+        scheduler = {"scheduler": lr_scheduler, "interval": "step", "frequency": 1}
+        return [optimizer], [scheduler]
+
+   
+
+def main(args):
+    dm = MSMARCOData(
+        model_name=args.model,
+        langs=args.langs,
+        triplets_path='data/msmarco-hard-triplets.jsonl.gz',
+        train_batch_size=args.batch_size,
+        cross_lingual_chance=args.cross_lingual_chance,
+        num_negs=args.num_negs
+    )
+    output_dir = f"output/{args.model.replace('/', '-')}-{datetime.now().strftime('%Y-%m-%d_%H-%M-%S')}"
+    print("Output_dir:", output_dir)
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    wandb_logger = WandbLogger(project="multilingual-cross-encoder", name=output_dir.split("/")[-1])
+
+    train_script_path = os.path.join(output_dir, 'train_script.py')
+    copyfile(__file__, train_script_path)
+    with open(train_script_path, 'a') as fOut:
+        fOut.write("\n\n# Script was called via:\n#python " + " ".join(sys.argv))
+
+    
+    # saves top-K checkpoints based on "val_loss" metric
+    checkpoint_callback = ModelCheckpoint(
+        every_n_train_steps=25000,
+        save_top_k=5,
+        monitor="global_train_step",
+        mode="max",
+        dirpath=output_dir,
+        filename="ckpt-{global_train_step}",
+    )
+
+
+    model = ListRankLoss(model_name=args.model)
+
+    trainer = Trainer(max_epochs=args.epochs, 
+                      accelerator="gpu", 
+                      devices=args.num_gpus, 
+                      precision=args.precision, 
+                      strategy=args.strategy,    
+                      default_root_dir=output_dir,
+                      callbacks=[checkpoint_callback],
+                      logger=wandb_logger
+                      )
+
+    trainer.fit(model, datamodule=dm)
+
+    #Save final HF model 
+    final_path = os.path.join(output_dir, "final")
+    dm.tokenizer.save_pretrained(final_path)
+    model.model.save_pretrained(final_path)
+
+  
+def eval(args):
+    import ir_datasets
+    
+ 
+    model = ListRankLoss.load_from_checkpoint(args.ckpt)
+    hf_model = model.model.cuda()
+    tokenizer = AutoTokenizer.from_pretrained(model.hparams.model_name)
+
+    dev_qids = set()
+
+    dev_queries = {}
+    dev_rel_docs = {}
+    needed_pids = set()
+    needed_qids = set()
+
+    corpus = {}
+    retrieved_docs = {}
+
+    dataset = ir_datasets.load("msmarco-passage/dev/small")
+    for query in dataset.queries_iter():
+        dev_qids.add(query.query_id)
+
+    
+    with open('data/qrels.dev.tsv') as fIn:
+        for line in fIn:
+            qid, _, pid, _ = line.strip().split('\t')
+
+            if qid not in dev_qids:
+                continue
+
+            if qid not in dev_rel_docs:
+                dev_rel_docs[qid] = set()
+            dev_rel_docs[qid].add(pid)
+
+            retrieved_docs[qid] = set()
+            needed_qids.add(qid)
+            needed_pids.add(pid)
+
+    for query in dataset.queries_iter():
+        qid = query.query_id
+        if qid in needed_qids:
+            dev_queries[qid] = query.text
+
+    with open('data/top1000.dev', 'rt') as fIn:
+        for line in fIn:
+            qid, pid, query, passage = line.strip().split("\t")
+            corpus[pid] = passage
+            retrieved_docs[qid].add(pid)
+
+
+    ## Run evaluator
+    print("Queries: {}".format(len(dev_queries)))
+
+    mrr_scores = []
+    hf_model.eval()
+
+    with torch.no_grad():
+        for qid in tqdm.tqdm(dev_queries, total=len(dev_queries)):
+            query = dev_queries[qid]
+            top_pids = list(retrieved_docs[qid])
+            cross_inp = [[query, corpus[pid]] for pid in top_pids]
+
+            encoded = tokenizer(cross_inp, padding=True, truncation=True, return_tensors="pt").to('cuda')
+            output = model(**encoded)
+            bert_score = output.logits.detach().cpu().numpy()
+            bert_score = np.squeeze(bert_score)
+        
+            argsort = np.argsort(-bert_score)
+
+            rank_score = 0
+            for rank, idx in enumerate(argsort[0:10]):
+                pid = top_pids[idx]
+                if pid in dev_rel_docs[qid]:
+                    rank_score = 1/(rank+1)
+                    break
+
+            mrr_scores.append(rank_score)
+        
+            if len(mrr_scores) % 10 == 0:
+                print("{} MRR@10: {:.2f}".format(len(mrr_scores), 100*np.mean(mrr_scores)))
+
+    print("MRR@10: {:.2f}".format(np.mean(mrr_scores)*100))
+
+
+if __name__ == '__main__':
+    parser = ArgumentParser()
+    parser.add_argument("--num_gpus", type=int, default=1)
+    parser.add_argument("--batch_size", type=int, default=32)
+    parser.add_argument("--epochs", type=int, default=10)
+    parser.add_argument("--strategy", default=None)
+    parser.add_argument("--model", default='microsoft/mdeberta-v3-base')
+    parser.add_argument("--eval", action="store_true")
+    parser.add_argument("--ckpt")
+    parser.add_argument("--cross_lingual_chance", type=float, default=0.33)
+    parser.add_argument("--precision", type=int, default=16)
+    parser.add_argument("--num_negs", type=int, default=3)
+    parser.add_argument("--langs", nargs="+", default=['english', 'chinese', 'french', 'german', 'indonesian', 'italian', 'portuguese', 'russian', 'spanish', 'arabic', 'dutch', 'hindi', 'japanese', 'vietnamese'])
+    
+    
+    args = parser.parse_args()
+
+    if args.eval:
+        eval(args)
+    else:
+        main(args)
+    
+
+# Script was called via:
+#python cross_mutlilingual.py --model nreimers/mMiniLMv2-L12-H384-distilled-from-XLMR-Large