Upload 9 files

README.md

@@ -0,0 +1,177 @@
+---
+language: en
+license: apache-2.0
+library_name: sentence-transformers
+tags:
+- sentence-transformers
+- feature-extraction
+- sentence-similarity
+- transformers
+datasets:
+- s2orc
+- flax-sentence-embeddings/stackexchange_xml
+- ms_marco
+- gooaq
+- yahoo_answers_topics
+- code_search_net
+- search_qa
+- eli5
+- snli
+- multi_nli
+- wikihow
+- natural_questions
+- trivia_qa
+- embedding-data/sentence-compression
+- embedding-data/flickr30k-captions
+- embedding-data/altlex
+- embedding-data/simple-wiki
+- embedding-data/QQP
+- embedding-data/SPECTER
+- embedding-data/PAQ_pairs
+- embedding-data/WikiAnswers
+pipeline_tag: sentence-similarity
+---
+
+
+# all-MiniLM-L6-v2
+This is a [sentence-transformers](https://www.SBERT.net) model: It maps sentences & paragraphs to a 384 dimensional dense vector space and can be used for tasks like clustering or semantic search.
+
+## Usage (Sentence-Transformers)
+Using this model becomes easy when you have [sentence-transformers](https://www.SBERT.net) installed:
+
+```
+pip install -U sentence-transformers
+```
+
+Then you can use the model like this:
+```python
+from sentence_transformers import SentenceTransformer
+sentences = ["This is an example sentence", "Each sentence is converted"]
+
+model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
+embeddings = model.encode(sentences)
+print(embeddings)
+```
+
+## Usage (HuggingFace Transformers)
+Without [sentence-transformers](https://www.SBERT.net), you can use the model like this: First, you pass your input through the transformer model, then you have to apply the right pooling-operation on-top of the contextualized word embeddings.
+
+```python
+from transformers import AutoTokenizer, AutoModel
+import torch
+import torch.nn.functional as F
+
+#Mean Pooling - Take attention mask into account for correct averaging
+def mean_pooling(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)
+
+
+# Sentences we want sentence embeddings for
+sentences = ['This is an example sentence', 'Each sentence is converted']
+
+# Load model from HuggingFace Hub
+tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
+model = AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
+
+# Tokenize sentences
+encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
+
+# Compute token embeddings
+with torch.no_grad():
+    model_output = model(**encoded_input)
+
+# Perform pooling
+sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
+
+# Normalize embeddings
+sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
+
+print("Sentence embeddings:")
+print(sentence_embeddings)
+```
+
+## Evaluation Results
+
+For an automated evaluation of this model, see the *Sentence Embeddings Benchmark*: [https://seb.sbert.net](https://seb.sbert.net?model_name=sentence-transformers/all-MiniLM-L6-v2)
+
+------
+
+## Background
+
+The project aims to train sentence embedding models on very large sentence level datasets using a self-supervised 
+contrastive learning objective. We used the pretrained [`nreimers/MiniLM-L6-H384-uncased`](https://huggingface.co/nreimers/MiniLM-L6-H384-uncased) model and fine-tuned in on a 
+1B sentence pairs dataset. We use a contrastive learning objective: given a sentence from the pair, the model should predict which out of a set of randomly sampled other sentences, was actually paired with it in our dataset.
+
+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 intervention from Googles Flax, JAX, and Cloud team member about efficient deep learning frameworks.
+
+## Intended uses
+
+Our model is intended to be used as a sentence and short paragraph encoder. Given an input text, it outputs a vector which captures 
+the semantic information. The sentence vector may be used for information retrieval, clustering or sentence similarity tasks.
+
+By default, input text longer than 256 word pieces is truncated.
+
+
+## Training procedure
+
+### Pre-training 
+
+We use the pretrained [`nreimers/MiniLM-L6-H384-uncased`](https://huggingface.co/nreimers/MiniLM-L6-H384-uncased) model. 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 our model on a TPU v3-8. We train the model during 100k 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: `train_script.py`.
+
+#### Training data
+
+We use the concatenation from multiple datasets to fine-tune our model. The total number of sentence pairs is above 1 billion sentences.
+We sampled each dataset given a weighted probability which configuration is detailed in the `data_config.json` file.
+
+
+| Dataset                                                  | Paper                                    | Number of training tuples  |
+|--------------------------------------------------------|:----------------------------------------:|:--------------------------:|
+| [Reddit comments (2015-2018)](https://github.com/PolyAI-LDN/conversational-datasets/tree/master/reddit) | [paper](https://arxiv.org/abs/1904.06472) | 726,484,430 |
+| [S2ORC](https://github.com/allenai/s2orc) Citation pairs (Abstracts) | [paper](https://aclanthology.org/2020.acl-main.447/) | 116,288,806 |
+| [WikiAnswers](https://github.com/afader/oqa#wikianswers-corpus) Duplicate question pairs | [paper](https://doi.org/10.1145/2623330.2623677) | 77,427,422 |
+| [PAQ](https://github.com/facebookresearch/PAQ) (Question, Answer) pairs | [paper](https://arxiv.org/abs/2102.07033) | 64,371,441 |
+| [S2ORC](https://github.com/allenai/s2orc) Citation pairs (Titles) | [paper](https://aclanthology.org/2020.acl-main.447/) | 52,603,982 |
+| [S2ORC](https://github.com/allenai/s2orc) (Title, Abstract) | [paper](https://aclanthology.org/2020.acl-main.447/) | 41,769,185 |
+| [Stack Exchange](https://huggingface.co/datasets/flax-sentence-embeddings/stackexchange_xml) (Title, Body) pairs  | - | 25,316,456 |
+| [Stack Exchange](https://huggingface.co/datasets/flax-sentence-embeddings/stackexchange_xml) (Title+Body, Answer) pairs  | - | 21,396,559 |
+| [Stack Exchange](https://huggingface.co/datasets/flax-sentence-embeddings/stackexchange_xml) (Title, Answer) pairs  | - | 21,396,559 |
+| [MS MARCO](https://microsoft.github.io/msmarco/) triplets | [paper](https://doi.org/10.1145/3404835.3462804) | 9,144,553 |
+| [GOOAQ: Open Question Answering with Diverse Answer Types](https://github.com/allenai/gooaq) | [paper](https://arxiv.org/pdf/2104.08727.pdf) | 3,012,496 |
+| [Yahoo Answers](https://www.kaggle.com/soumikrakshit/yahoo-answers-dataset) (Title, Answer) | [paper](https://proceedings.neurips.cc/paper/2015/hash/250cf8b51c773f3f8dc8b4be867a9a02-Abstract.html) | 1,198,260 |
+| [Code Search](https://huggingface.co/datasets/code_search_net) | - | 1,151,414 |
+| [COCO](https://cocodataset.org/#home) Image captions | [paper](https://link.springer.com/chapter/10.1007%2F978-3-319-10602-1_48) | 828,395|
+| [SPECTER](https://github.com/allenai/specter) citation triplets | [paper](https://doi.org/10.18653/v1/2020.acl-main.207) | 684,100 |
+| [Yahoo Answers](https://www.kaggle.com/soumikrakshit/yahoo-answers-dataset) (Question, Answer) | [paper](https://proceedings.neurips.cc/paper/2015/hash/250cf8b51c773f3f8dc8b4be867a9a02-Abstract.html) | 681,164 |
+| [Yahoo Answers](https://www.kaggle.com/soumikrakshit/yahoo-answers-dataset) (Title, Question) | [paper](https://proceedings.neurips.cc/paper/2015/hash/250cf8b51c773f3f8dc8b4be867a9a02-Abstract.html) | 659,896 |
+| [SearchQA](https://huggingface.co/datasets/search_qa) | [paper](https://arxiv.org/abs/1704.05179) | 582,261 |
+| [Eli5](https://huggingface.co/datasets/eli5) | [paper](https://doi.org/10.18653/v1/p19-1346) | 325,475 |
+| [Flickr 30k](https://shannon.cs.illinois.edu/DenotationGraph/) | [paper](https://transacl.org/ojs/index.php/tacl/article/view/229/33) | 317,695 |
+| [Stack Exchange](https://huggingface.co/datasets/flax-sentence-embeddings/stackexchange_xml) Duplicate questions (titles) | | 304,525 |
+| AllNLI ([SNLI](https://nlp.stanford.edu/projects/snli/) and [MultiNLI](https://cims.nyu.edu/~sbowman/multinli/) | [paper SNLI](https://doi.org/10.18653/v1/d15-1075), [paper MultiNLI](https://doi.org/10.18653/v1/n18-1101) | 277,230 | 
+| [Stack Exchange](https://huggingface.co/datasets/flax-sentence-embeddings/stackexchange_xml) Duplicate questions (bodies) | | 250,519 |
+| [Stack Exchange](https://huggingface.co/datasets/flax-sentence-embeddings/stackexchange_xml) Duplicate questions (titles+bodies) | | 250,460 |
+| [Sentence Compression](https://github.com/google-research-datasets/sentence-compression) | [paper](https://www.aclweb.org/anthology/D13-1155/) | 180,000 |
+| [Wikihow](https://github.com/pvl/wikihow_pairs_dataset) | [paper](https://arxiv.org/abs/1810.09305) | 128,542 |
+| [Altlex](https://github.com/chridey/altlex/) | [paper](https://aclanthology.org/P16-1135.pdf) | 112,696 |
+| [Quora Question Triplets](https://quoradata.quora.com/First-Quora-Dataset-Release-Question-Pairs) | - | 103,663 |
+| [Simple Wikipedia](https://cs.pomona.edu/~dkauchak/simplification/) | [paper](https://www.aclweb.org/anthology/P11-2117/) | 102,225 |
+| [Natural Questions (NQ)](https://ai.google.com/research/NaturalQuestions) | [paper](https://transacl.org/ojs/index.php/tacl/article/view/1455) | 100,231 |
+| [SQuAD2.0](https://rajpurkar.github.io/SQuAD-explorer/) | [paper](https://aclanthology.org/P18-2124.pdf) | 87,599 |
+| [TriviaQA](https://huggingface.co/datasets/trivia_qa) | - | 73,346 |
+| **Total** | | **1,170,060,424** |

config.json

@@ -0,0 +1,24 @@
+{
+  "_name_or_path": "nreimers/MiniLM-L6-H384-uncased",
+  "architectures": [
+    "BertModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 384,
+  "initializer_range": 0.02,
+  "intermediate_size": 1536,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 6,
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "transformers_version": "4.8.2",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 30522
+}

data_config.json

@@ -0,0 +1,344 @@
+"""
+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, queue):
+    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
+        batch = queue.get()
+        #print(index, "batch {}x{}".format(len(batch), ",".join([str(len(b)) for b in batch])))
+        
+
+        if len(batch[0]) == 2: #(anchor, positive)
+            text1 = tokenizer([b[0] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+            text2 = tokenizer([b[1] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+
+            ### 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 512 x 512
+            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]
+            
+            ## Symmetric loss as in CLIP
+            loss = (cross_entropy_loss(scores, labels) + cross_entropy_loss(scores.transpose(0, 1), labels)) / 2
+
+        else:   #(anchor, positive, negative)
+            text1 = tokenizer([b[0] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+            text2 = tokenizer([b[1] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+            text3 = tokenizer([b[2] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+
+            embeddings_a  = model(**text1.to(device))
+            embeddings_b1 = model(**text2.to(device))
+            embeddings_b2 = model(**text3.to(device))
+
+            embeddings_a  = torch_xla.core.functions.all_gather(embeddings_a)
+            embeddings_b1 = torch_xla.core.functions.all_gather(embeddings_b1)
+            embeddings_b2 = torch_xla.core.functions.all_gather(embeddings_b2)
+
+            embeddings_b = torch.cat([embeddings_b1, embeddings_b2])
+
+            ### Compute similarity scores 512 x 1024
+            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)
+
+        
+        # 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, "final")
+    xm.master_print("save model final: "+ output_path)
+    model.save_pretrained(output_path)
+
+
+def produce_data(args, queue, filepaths, dataset_indices):
+    global_batch_size = args.batch_size*args.nprocs    #Global batch size
+    size_per_dataset = int(global_batch_size / args.datasets_per_batch)    #How many datasets per batch
+    num_same_dataset = int(size_per_dataset / args.batch_size)
+    print("producer", "global_batch_size", global_batch_size)
+    print("producer", "size_per_dataset", size_per_dataset)
+    print("producer", "num_same_dataset", num_same_dataset)
+    
+    datasets = []
+    for filepath in filepaths:
+        if "reddit_" in filepath:       #Special dataset class for Reddit files
+            data_obj = RedditDataset(filepath)
+        else:
+            data_obj = Dataset(filepath)
+        datasets.append(iter(data_obj)) 
+    
+    # Store if dataset is in a 2 col or 3 col format
+    num_cols = {idx: len(next(dataset)) for idx, dataset in enumerate(datasets)}
+
+    while True:
+        texts_in_batch = set()
+        batch_format = None     #2 vs 3 col format for this batch
+        
+        #Add data from several sub datasets
+        for _ in range(args.datasets_per_batch):
+            valid_dataset = False   #Check that datasets have the same 2/3 col format
+            while not valid_dataset:
+                data_idx = random.choice(dataset_indices)
+                if batch_format is None:
+                    batch_format = num_cols[data_idx]
+                    valid_dataset = True
+                else:   #Check that this dataset has the same format
+                    valid_dataset = (batch_format == num_cols[data_idx])
+            
+            #Get data from this dataset
+            dataset = datasets[data_idx]
+            for _ in range(num_same_dataset):
+                for _ in range(args.nprocs):
+                    batch_device = []   #A batch for one device
+                    while len(batch_device) < args.batch_size:
+                        sample = next(dataset)
+                        in_batch = False
+                        for text in sample:
+                            if text in texts_in_batch:
+                                in_batch = True
+                                break
+                        
+                        if not in_batch:
+                            for text in sample:
+                                texts_in_batch.add(text)
+                            batch_device.append(sample)
+
+                    queue.put(batch_device)
+                      
+
+class RedditDataset:
+    """
+    A class that handles the reddit data files
+    """
+    def __init__(self, filepath):
+        self.filepath = filepath
+
+    def __iter__(self):
+        while True:
+            with gzip.open(self.filepath, "rt") as fIn:
+                    for line in fIn:
+                        data = json.loads(line)
+
+                        if "response" in data and "context" in data:
+                            yield [data["response"], data["context"]]
+
+class Dataset:
+    """
+    A class that handles one dataset
+    """
+    def __init__(self, filepath):
+        self.filepath = filepath
+
+    def __iter__(self):
+        max_dataset_size = 10*1000*1000    #Cache small datasets in memory
+        dataset = []
+        data_format = None
+
+        while dataset is None or len(dataset) == 0:
+            with gzip.open(self.filepath, "rt") as fIn:
+                for line in fIn:
+                    data = json.loads(line)
+                    if isinstance(data, dict):
+                        data = data['texts']
+
+                    if data_format is None:
+                        data_format = len(data)
+                    
+                    #Ensure that all entries are of the same 2/3 col format
+                    assert len(data) == data_format
+
+                    if dataset is not None:
+                        dataset.append(data)
+                        if len(dataset) >= max_dataset_size:
+                            dataset = None
+
+                    yield data
+                
+        # Data loaded. Now stream to the queue
+        # Shuffle for each epoch
+        while True:
+            random.shuffle(dataset)
+            for data in dataset:
+                yield 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=64)
+    parser.add_argument('--max_length', type=int, default=128)
+    parser.add_argument('--nprocs', type=int, default=8)
+    parser.add_argument('--datasets_per_batch', type=int, default=2, help="Number of datasets per batch")
+    parser.add_argument('--scale', type=float, default=20, help="Use 20 for cossim, and 1 when you work with unnormalized embeddings with dot product")
+    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()
+
+    # Ensure global batch size is divisble by data_sample_size
+    assert (args.batch_size*args.nprocs) % args.datasets_per_batch == 0
+
+    logging.info("Output: "+args.output)
+    if os.path.exists(args.output):
+        print("Output folder already exists.")
+        input("Continue?")
+
+    # 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)
+
+    queue = mp.Queue(maxsize=100*args.nprocs)
+    
+    filepaths = []
+    dataset_indices = []
+    for idx, data in enumerate(data_config):
+        filepaths.append(os.path.join(os.path.expanduser(args.data_folder), data['name']))
+        dataset_indices.extend([idx]*data['weight'])
+
+    # Start producer
+    p = mp.Process(target=produce_data, args=(args, queue, filepaths, dataset_indices))
+    p.start()
+
+    # Run training
+    print("Start processes:", args.nprocs)
+    xmp.spawn(train_function, args=(args, queue), 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")
+    p.kill()
+    exit()
+
+
+
+# Script was called via:
+#python train_many_data_files_v2.py --steps 1000000 --batch_size 128 --model nreimers/MiniLM-L6-H384-uncased train_data_configs/all_datasets_v4.json output/all_datasets_v4_MiniLM-L6-H384-uncased-batch128

sentence_bert_config.json

@@ -0,0 +1,4 @@
+{
+  "max_seq_length": 256,
+  "do_lower_case": false
+}

special_tokens_map.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}

tokenizer_config.json

@@ -0,0 +1 @@
+{"do_lower_case": true, "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]", "tokenize_chinese_chars": true, "strip_accents": null, "name_or_path": "nreimers/MiniLM-L6-H384-uncased", "do_basic_tokenize": true, "never_split": null, "tokenizer_class": "BertTokenizer", "model_max_length": 512}

train_script.py

@@ -0,0 +1,344 @@
+"""
+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, queue):
+    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
+        batch = queue.get()
+        #print(index, "batch {}x{}".format(len(batch), ",".join([str(len(b)) for b in batch])))
+        
+
+        if len(batch[0]) == 2: #(anchor, positive)
+            text1 = tokenizer([b[0] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+            text2 = tokenizer([b[1] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+
+            ### 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 512 x 512
+            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]
+            
+            ## Symmetric loss as in CLIP
+            loss = (cross_entropy_loss(scores, labels) + cross_entropy_loss(scores.transpose(0, 1), labels)) / 2
+
+        else:   #(anchor, positive, negative)
+            text1 = tokenizer([b[0] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+            text2 = tokenizer([b[1] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+            text3 = tokenizer([b[2] for b in batch], return_tensors="pt", max_length=args.max_length, truncation=True, padding="max_length")
+
+            embeddings_a  = model(**text1.to(device))
+            embeddings_b1 = model(**text2.to(device))
+            embeddings_b2 = model(**text3.to(device))
+
+            embeddings_a  = torch_xla.core.functions.all_gather(embeddings_a)
+            embeddings_b1 = torch_xla.core.functions.all_gather(embeddings_b1)
+            embeddings_b2 = torch_xla.core.functions.all_gather(embeddings_b2)
+
+            embeddings_b = torch.cat([embeddings_b1, embeddings_b2])
+
+            ### Compute similarity scores 512 x 1024
+            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)
+
+        
+        # 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, "final")
+    xm.master_print("save model final: "+ output_path)
+    model.save_pretrained(output_path)
+
+
+def produce_data(args, queue, filepaths, dataset_indices):
+    global_batch_size = args.batch_size*args.nprocs    #Global batch size
+    size_per_dataset = int(global_batch_size / args.datasets_per_batch)    #How many datasets per batch
+    num_same_dataset = int(size_per_dataset / args.batch_size)
+    print("producer", "global_batch_size", global_batch_size)
+    print("producer", "size_per_dataset", size_per_dataset)
+    print("producer", "num_same_dataset", num_same_dataset)
+    
+    datasets = []
+    for filepath in filepaths:
+        if "reddit_" in filepath:       #Special dataset class for Reddit files
+            data_obj = RedditDataset(filepath)
+        else:
+            data_obj = Dataset(filepath)
+        datasets.append(iter(data_obj)) 
+    
+    # Store if dataset is in a 2 col or 3 col format
+    num_cols = {idx: len(next(dataset)) for idx, dataset in enumerate(datasets)}
+
+    while True:
+        texts_in_batch = set()
+        batch_format = None     #2 vs 3 col format for this batch
+        
+        #Add data from several sub datasets
+        for _ in range(args.datasets_per_batch):
+            valid_dataset = False   #Check that datasets have the same 2/3 col format
+            while not valid_dataset:
+                data_idx = random.choice(dataset_indices)
+                if batch_format is None:
+                    batch_format = num_cols[data_idx]
+                    valid_dataset = True
+                else:   #Check that this dataset has the same format
+                    valid_dataset = (batch_format == num_cols[data_idx])
+            
+            #Get data from this dataset
+            dataset = datasets[data_idx]
+            for _ in range(num_same_dataset):
+                for _ in range(args.nprocs):
+                    batch_device = []   #A batch for one device
+                    while len(batch_device) < args.batch_size:
+                        sample = next(dataset)
+                        in_batch = False
+                        for text in sample:
+                            if text in texts_in_batch:
+                                in_batch = True
+                                break
+                        
+                        if not in_batch:
+                            for text in sample:
+                                texts_in_batch.add(text)
+                            batch_device.append(sample)
+
+                    queue.put(batch_device)
+                      
+
+class RedditDataset:
+    """
+    A class that handles the reddit data files
+    """
+    def __init__(self, filepath):
+        self.filepath = filepath
+
+    def __iter__(self):
+        while True:
+            with gzip.open(self.filepath, "rt") as fIn:
+                    for line in fIn:
+                        data = json.loads(line)
+
+                        if "response" in data and "context" in data:
+                            yield [data["response"], data["context"]]
+
+class Dataset:
+    """
+    A class that handles one dataset
+    """
+    def __init__(self, filepath):
+        self.filepath = filepath
+
+    def __iter__(self):
+        max_dataset_size = 10*1000*1000    #Cache small datasets in memory
+        dataset = []
+        data_format = None
+
+        while dataset is None or len(dataset) == 0:
+            with gzip.open(self.filepath, "rt") as fIn:
+                for line in fIn:
+                    data = json.loads(line)
+                    if isinstance(data, dict):
+                        data = data['texts']
+
+                    if data_format is None:
+                        data_format = len(data)
+                    
+                    #Ensure that all entries are of the same 2/3 col format
+                    assert len(data) == data_format
+
+                    if dataset is not None:
+                        dataset.append(data)
+                        if len(dataset) >= max_dataset_size:
+                            dataset = None
+
+                    yield data
+                
+        # Data loaded. Now stream to the queue
+        # Shuffle for each epoch
+        while True:
+            random.shuffle(dataset)
+            for data in dataset:
+                yield 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=64)
+    parser.add_argument('--max_length', type=int, default=128)
+    parser.add_argument('--nprocs', type=int, default=8)
+    parser.add_argument('--datasets_per_batch', type=int, default=2, help="Number of datasets per batch")
+    parser.add_argument('--scale', type=float, default=20, help="Use 20 for cossim, and 1 when you work with unnormalized embeddings with dot product")
+    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()
+
+    # Ensure global batch size is divisble by data_sample_size
+    assert (args.batch_size*args.nprocs) % args.datasets_per_batch == 0
+
+    logging.info("Output: "+args.output)
+    if os.path.exists(args.output):
+        print("Output folder already exists.")
+        input("Continue?")
+
+    # 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)
+
+    queue = mp.Queue(maxsize=100*args.nprocs)
+    
+    filepaths = []
+    dataset_indices = []
+    for idx, data in enumerate(data_config):
+        filepaths.append(os.path.join(os.path.expanduser(args.data_folder), data['name']))
+        dataset_indices.extend([idx]*data['weight'])
+
+    # Start producer
+    p = mp.Process(target=produce_data, args=(args, queue, filepaths, dataset_indices))
+    p.start()
+
+    # Run training
+    print("Start processes:", args.nprocs)
+    xmp.spawn(train_function, args=(args, queue), 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")
+    p.kill()
+    exit()
+
+
+
+# Script was called via:
+#python train_many_data_files_v2.py --steps 1000000 --batch_size 128 --model nreimers/MiniLM-L6-H384-uncased train_data_configs/all_datasets_v4.json output/all_datasets_v4_MiniLM-L6-H384-uncased-batch128