初始化项目，由ModelHub XC社区提供模型

Model: flax-sentence-embeddings/stackoverflow_mpnet-base
Source: Original Platform

.gitattributes

@@ -0,0 +1,17 @@
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tar.gz filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack 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
+*tfevents* filter=lfs diff=lfs merge=lfs -text

1_Pooling/config.json

@@ -0,0 +1,7 @@
+{
+  "word_embedding_dimension": 768,
+  "pooling_mode_cls_token": false,
+  "pooling_mode_mean_tokens": true,
+  "pooling_mode_max_tokens": false,
+  "pooling_mode_mean_sqrt_len_tokens": false
+}

README.md

@@ -0,0 +1,66 @@
+---
+pipeline_tag: sentence-similarity
+tags:
+- sentence-transformers
+- feature-extraction
+- sentence-similarity
+---
+
+# stackoverflow_mpnet-base
+
+This is a microsoft/mpnet-base model trained on 18,562,443 (title, body) pairs from StackOverflow. 
+
+SentenceTransformers is a set of models and frameworks that enable training and generating sentence embeddings from given data. The generated sentence embeddings can be utilized for Clustering, Semantic Search and other tasks. We used a pretrained [microsoft/mpnet-base](https://huggingface.co/microsoft/mpnet-base) model and trained it using Siamese Network setup and contrastive learning objective. 18,562,443 (title, body) pairs from StackOverflow was used as training data. For this model, mean pooling of hidden states were used as sentence embeddings. See data_config.json and train_script.py in this respository how the model was trained and which datasets have been used.
+
+We developed this model during the 
+[Community week using JAX/Flax for NLP & CV](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104), 
+organized by Hugging Face. We developed this model as part of the project:
+[Train the Best Sentence Embedding Model Ever with 1B Training Pairs](https://discuss.huggingface.co/t/train-the-best-sentence-embedding-model-ever-with-1b-training-pairs/7354). We benefited from efficient hardware infrastructure to run the project: 7 TPUs v3-8, as well
+as assistance from Google’s Flax, JAX, and Cloud team members about efficient deep learning frameworks.
+
+## Intended uses
+
+Our model is intended to be used as a sentence encoder for a search engine. Given an input sentence, it outputs a vector which captures 
+the sentence semantic information. The sentence vector may be used for semantic-search, clustering or sentence similarity tasks.
+
+## How to use
+
+Here is how to use this model to get the features of a given text using [SentenceTransformers](https://github.com/UKPLab/sentence-transformers) library:
+
+```python
+from sentence_transformers import SentenceTransformer
+
+model = SentenceTransformer('flax-sentence-embeddings/stackoverflow_mpnet-base')
+text = "Replace me by any question / answer you'd like."
+text_embbedding = model.encode(text)
+# array([-0.01559514,  0.04046123,  0.1317083 ,  0.00085931,  0.04585106,
+#        -0.05607086,  0.0138078 ,  0.03569756,  0.01420381,  0.04266302 ...],
+#        dtype=float32)
+```
+
+# Training procedure
+
+## Pre-training 
+
+We use the pretrained [microsoft/mpnet-base](https://huggingface.co/microsoft/mpnet-base). Please refer to the model
+card for more detailed information about the pre-training procedure.
+
+## Fine-tuning 
+
+We fine-tune the model using a contrastive objective. Formally, we compute the cosine similarity from each possible sentence pairs from the batch.
+We then apply the cross entropy loss by comparing with true pairs.
+
+### Hyper parameters
+
+We trained on model on a TPU v3-8. We train the model during 80k steps using a batch size of 1024 (128 per TPU core).
+We use a learning rate warm up of 500. The sequence length was limited to 128 tokens. We used the AdamW optimizer with
+a 2e-5 learning rate. The full training script is accessible in this current repository.
+
+### Training data
+
+We used 18,562,443 (title, body) pairs from StackOverflow as training data.
+
+| Dataset                                                  | Paper                                    | Number of training tuples  |
+|:--------------------------------------------------------:|:----------------------------------------:|:--------------------------:|
+| StackOverflow title body pairs | - | 18,562,443 |
+

config.json

@@ -0,0 +1,23 @@
+{
+  "_name_or_path": "microsoft/mpnet-base",
+  "architectures": [
+    "MPNetForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "mpnet",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "relative_attention_num_buckets": 32,
+  "transformers_version": "4.8.2",
+  "vocab_size": 30527
+}

config_sentence_transformers.json

@@ -0,0 +1,7 @@
+{
+  "__version__": {
+    "sentence_transformers": "2.0.0",
+    "transformers": "4.6.1",
+    "pytorch": "1.8.1"
+  }
+}

data_config.json

@@ -0,0 +1,7 @@
+[
+{
+        "name": "stackexchange_title_body/stackoverflow.com-Posts.jsonl.gz",
+        "lines": 18562443,
+        "weight": 100
+}
+]

modules.json

@@ -0,0 +1,20 @@
+[
+  {
+    "idx": 0,
+    "name": "0",
+    "path": "",
+    "type": "sentence_transformers.models.Transformer"
+  },
+  {
+    "idx": 1,
+    "name": "1",
+    "path": "1_Pooling",
+    "type": "sentence_transformers.models.Pooling"
+  },
+  {
+    "idx": 2,
+    "name": "2",
+    "path": "2_Normalize",
+    "type": "sentence_transformers.models.Normalize"
+  }
+]

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7dcdca5dfbca9becac708f520fa11b36e7c615bec56e033f1566eb8745e2d9d3
+size 438011953

sentence_bert_config.json

@@ -0,0 +1,4 @@
+{
+  "max_seq_length": 512,
+  "do_lower_case": false
+}

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_config.json

@@ -0,0 +1 @@
+{"do_lower_case": true, "bos_token": "<s>", "eos_token": "</s>", "sep_token": "</s>", "cls_token": "<s>", "unk_token": "[UNK]", "pad_token": "<pad>", "mask_token": "<mask>", "tokenize_chinese_chars": true, "strip_accents": null, "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "microsoft/mpnet-base", "tokenizer_class": "MPNetTokenizer"}

train_script.py

@@ -0,0 +1,235 @@
+"""
+Train script for a single file
+
+Need to set the TPU address first:
+export XRT_TPU_CONFIG="localservice;0;localhost:51011"
+"""
+
+import torch.multiprocessing as mp
+import threading
+import time
+import random
+import sys
+import argparse
+import gzip
+import json
+import logging
+import tqdm
+import torch
+from torch import nn
+from torch.utils.data import DataLoader
+import torch
+import torch_xla
+import torch_xla.core
+import torch_xla.core.functions
+import torch_xla.core.xla_model as xm
+import torch_xla.distributed.xla_multiprocessing as xmp
+import torch_xla.distributed.parallel_loader as pl
+import os
+from shutil import copyfile
+
+
+from transformers import (
+    AdamW,
+    AutoModel,
+    AutoTokenizer,
+    get_linear_schedule_with_warmup,
+    set_seed,
+)
+
+class AutoModelForSentenceEmbedding(nn.Module):
+    def __init__(self, model_name, tokenizer, normalize=True):
+        super(AutoModelForSentenceEmbedding, self).__init__()
+
+        self.model = AutoModel.from_pretrained(model_name)
+        self.normalize = normalize
+        self.tokenizer = tokenizer
+
+    def forward(self, **kwargs):
+        model_output = self.model(**kwargs)
+        embeddings = self.mean_pooling(model_output, kwargs['attention_mask'])
+        if self.normalize:
+            embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
+
+        return embeddings
+
+    def mean_pooling(self, model_output, attention_mask):
+        token_embeddings = model_output[0]  # First element of model_output contains all token embeddings
+        input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
+
+    def save_pretrained(self, output_path):
+        if xm.is_master_ordinal():
+            self.tokenizer.save_pretrained(output_path)
+            self.model.config.save_pretrained(output_path)
+
+        xm.save(self.model.state_dict(), os.path.join(output_path, "pytorch_model.bin"))
+       
+
+
+
+def train_function(index, args, queues, dataset_indices):
+    dataset_rnd = random.Random(index % args.data_word_size)  #Defines which dataset to use in every step
+    tokenizer = AutoTokenizer.from_pretrained(args.model)
+    model = AutoModelForSentenceEmbedding(args.model, tokenizer)
+    
+  
+    ### Train Loop
+    device = xm.xla_device()
+    model = model.to(device)
+
+    # Instantiate optimizer
+    optimizer = AdamW(params=model.parameters(), lr=2e-5, correct_bias=True)
+
+    lr_scheduler = get_linear_schedule_with_warmup(
+        optimizer=optimizer,
+        num_warmup_steps=500,
+        num_training_steps=args.steps,
+    )
+    
+    # Now we train the model
+    cross_entropy_loss = nn.CrossEntropyLoss()
+    max_grad_norm = 1
+
+    model.train()
+   
+    for global_step in tqdm.trange(args.steps, disable=not xm.is_master_ordinal()):
+        
+        #### Get the batch data
+        dataset_idx = dataset_rnd.choice(dataset_indices)
+        text1 = []
+        text2 = []
+        for _ in range(args.batch_size):
+            example = queues[dataset_idx].get()
+            text1.append(example[0])
+            text2.append(example[1])
+
+        #print(index, f"dataset {dataset_idx}", text1[0:3])
+
+        text1 = tokenizer(text1, return_tensors="pt", max_length=128, truncation=True, padding="max_length")
+        text2 = tokenizer(text2, return_tensors="pt", max_length=128, truncation=True, padding="max_length")
+
+        ### Compute embeddings
+        #print(index, "compute embeddings")
+        embeddings_a = model(**text1.to(device))
+        embeddings_b = model(**text2.to(device))
+        
+
+        ### Gather all embedings 
+        embeddings_a = torch_xla.core.functions.all_gather(embeddings_a)
+        embeddings_b = torch_xla.core.functions.all_gather(embeddings_b)
+
+        ### Compute similarity scores
+        scores = torch.mm(embeddings_a, embeddings_b.transpose(0, 1)) * args.scale
+     
+        ### Compute cross-entropy loss
+        labels = torch.tensor(range(len(scores)), dtype=torch.long, device=embeddings_a.device)  # Example a[i] should match with b[i]
+        
+        ## One-way loss
+        #loss = cross_entropy_loss(scores, labels)
+      
+        ## Symmetric loss as in CLIP
+        loss = (cross_entropy_loss(scores, labels) + cross_entropy_loss(scores.transpose(0, 1), labels)) / 2
+        
+        # Backward pass
+        optimizer.zero_grad()
+        loss.backward()
+        torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)
+        
+        xm.optimizer_step(optimizer, barrier=True)
+        lr_scheduler.step()
+
+        #Save model
+        if (global_step+1) % args.save_steps == 0:
+            output_path = os.path.join(args.output, str(global_step+1))
+            xm.master_print("save model: "+output_path)
+            model.save_pretrained(output_path)
+          
+            
+    output_path = os.path.join(args.output)
+    xm.master_print("save model final: "+ output_path)
+    model.save_pretrained(output_path)
+
+
+
+def load_data(path, queue):
+    dataset = []
+
+    with gzip.open(path, "rt") as fIn:
+        for line in fIn:
+            data = json.loads(line)
+            if isinstance(data, dict):
+                data = data['texts']
+            
+            #Only use two columns
+            dataset.append(data[0:2])
+            queue.put(data[0:2])
+
+    # Data loaded. Now stream to the queue
+    # Shuffle for each epoch
+    while True:
+        random.shuffle(dataset)
+        for data in dataset:
+            queue.put(data)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model', default='nreimers/MiniLM-L6-H384-uncased')
+    parser.add_argument('--steps', type=int, default=2000)
+    parser.add_argument('--save_steps', type=int, default=10000)
+    parser.add_argument('--batch_size', type=int, default=32)
+    parser.add_argument('--nprocs', type=int, default=8)
+    parser.add_argument('--data_word_size', type=int, default=2, help="How many different dataset should be included in every train step. Cannot be larger than nprocs")
+    parser.add_argument('--scale', type=float, default=20)
+    parser.add_argument('--data_folder', default="/data", help="Folder with your dataset files")
+    parser.add_argument('data_config', help="A data_config.json file")
+    parser.add_argument('output')
+    args = parser.parse_args()
+
+    logging.info("Output: "+args.output)
+    if os.path.exists(args.output):
+        print("Output folder already exists. Exit!")
+        exit()
+
+    # Write train script to output path
+    os.makedirs(args.output, exist_ok=True)
+
+    data_config_path = os.path.join(args.output, 'data_config.json')
+    copyfile(args.data_config, data_config_path)
+
+    train_script_path = os.path.join(args.output, '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))
+
+
+
+    #Load data config
+    with open(args.data_config) as fIn:
+        data_config = json.load(fIn)
+
+    threads = []
+    queues = []
+    dataset_indices = []
+    for data in data_config:
+        data_idx = len(queues)
+        queue = mp.Queue(maxsize=args.nprocs*args.batch_size)
+        th = threading.Thread(target=load_data, daemon=True, args=(os.path.join(os.path.expanduser(args.data_folder), data['name']), queue))
+        th.start()
+        threads.append(th)
+        queues.append(queue)
+        dataset_indices.extend([data_idx]*data['weight'])
+
+
+    print("Start processes:", args.nprocs)
+    xmp.spawn(train_function, args=(args, queues, dataset_indices), nprocs=args.nprocs, start_method='fork')
+    print("Training done")
+    print("It might be that not all processes exit automatically. In that case you must manually kill this process.")
+    print("With 'pkill python' you can kill all remaining python processes")
+    exit()
+
+
+
+# Script was called via:
+#python train_many_data_files.py --steps 100000 --batch_size 64 --model microsoft/mpnet-base train_data_configs/stackoverflow.json output/stackoverflow_mpnet-base